Skills, hooks, repos, and MCP servers we use daily. Each one comes with a walkthrough so you can plug it into your own setup immediately. Updated regularly.
Part 1: What I'd Set Up First For Claude Code
Three things I'd do on day one, and this is the first. Skip it and you'll spend month one wondering why Claude keeps forgetting your rules.
This is part one of three. There's a ton of overwhelm around Claude Code and where to actually begin, so I stripped it back to the three moves I'd make on day one if I were starting over.
Part 1 is this page. Parts 2 and 3 land over the next few days and will be linked right here when they do.
Set up a memory layer before you begin any serious work. Not after. Before.
Vanilla Claude Code does its best to remember things. It has no system for streamlining specific rules to specific projects. So it guesses, and it guesses fresh every session.
I use BASE. Install it, activate the knowledge graph, and Claude registers all your projects, tasks, rules, and your operator profile in one place it can actually query.
Why CLAUDE.md alone isn't enough
Your CLAUDE.md holds static rules. Rules change over time, and you have to hand-update the file every time they do. It does have a baked-in memory system, and that part is fine.
Here's the thing. Every time you open Claude you're working off stale data and stale rules. Between sessions Claude loses track of your active projects, so at session start it has no idea what's going on.
That's not a prompting problem. It's a state problem, and you fix it once at the infrastructure layer.
How the fix actually works
A session start hook fires the moment you open Claude Code.
The hook activates the BASE MCP, which pulls your knowledge graph.
The graph hands over your state: active work, tasks, operator profile, and your always-on global rules.
You decide what to work on next with Claude already holding full awareness of the board.
Then, as you use it, Claude creates new keyword-triggered knowledge domains that are self-learned and carry their own rules. Say "carousel" and only your carousel rules load. Nothing else fires, so nothing else eats context.
Close every session with *end
At the end of every session, type *end.
It conglomerates everything from that session, the context, the projects, the tasks, the decisions, and updates the knowledge graph across the board.
That's the compounding part. Your Claude Code gets measurably smarter every session instead of resetting to zero.
Go deeper on BASE
The full breakdown lives in this same toolkit: BASE: A Knowledge Graph For Your Rules. That one covers domains and triggers, what to move out of your CLAUDE.md and what to leave in, star commands, and the mode pack.
Worth pairing with it once you're set up:
Claude Resources To Read is the nine Anthropic articles that explain how Claude Code actually works under the hood.
Top 10 Tips Claude Code covers worktrees, subagents, and the claude.md habits worth keeping.
Graphify is the other half of the memory story. BASE graphs how you work, Graphify graphs what you know.
What's coming in parts 2 and 3
Parts 2 and 3 publish over the next few days and cover the other two things I'd set up before doing real work. When they go live they'll be linked from this page, and each part links back to the others so you can run the whole setup in order.
Every one of these does the job of a tool you pay for monthly. All ten links up top, then the exact prompt to hand your AI agent so it does the setup for you.
All 10 repos
Every link goes straight to the source. All ten are free. Star counts checked 2 September 2026.
Credit where it is due: Eric Michaud put most of these on my radar first.
Let your agent do the setup
Here is the part most lists skip. You do not have to install any of this by hand.
Most of these run in Docker. That word scares people off, and it should not. Docker is a way to run an app in its own sealed box on your computer so it cannot break anything else. Install Docker Desktop once, then your AI agent handles the rest.
The pattern is the same every time. Open Claude Code in a fresh folder and paste a prompt like this:
Read the install docs at <repo URL>.
Set it up on this machine using Docker.
Explain each step before you run it, and stop
if anything needs a decision from me.
When it is running, tell me the URL to open.
Four things make that prompt work:
You point it at the real docs. The agent reads the current instructions instead of guessing from memory, so it does not use a command that stopped working two versions ago.
You ask it to explain before running. You see what is about to happen. That is how you learn this instead of just watching it happen.
You tell it to stop on decisions. Ports, passwords, where your files get stored. Those are yours to choose, not the agent's.
You ask for the URL at the end. Otherwise you are hunting through terminal output for it.
Run that prompt once per tool. Start with one, get it working, then do the next.
Copy and paste prompts, one per repo
Same shape as above, filled in for each. Paste one at a time.
Immich: Read the install docs at github.com/immich-app/immich and set up Immich with Docker Compose on this machine. Ask me where I want my photo library stored before you start. Explain each step, then give me the URL.
Stirling PDF: Set up Stirling PDF from github.com/Stirling-Tools/Stirling-PDF using Docker on port 8080. Confirm nothing else is already using that port first, then give me the URL.
Vaultwarden: Read the wiki at github.com/dani-garcia/vaultwarden and set up Vaultwarden with Docker. This one holds passwords, so walk me through the security settings and do not pick them for me.
Meetily: Go to github.com/Zackriya-Solutions/meetily, find the right download for my operating system, and walk me through installing it and pointing it at a local transcription model.
Actual Budget: Read the docs at actualbudget.org/docs/install and set up Actual Budget with Docker on this machine. Ask me where the data file should live before you start.
Dub: Read github.com/dubinc/dub and tell me plainly whether self-hosting or their free hosted plan makes more sense for someone doing a few thousand clicks a month. Then set up whichever you recommend.
Activepieces: Set up Activepieces from github.com/activepieces/activepieces with Docker. Once it runs, show me how to build one automation that watches a folder and posts to Slack.
Documenso: Read github.com/documenso/documenso and set it up with Docker. I need to send contracts for signature, so tell me what is needed to make signatures legally valid where I live.
Kdenlive: Install Kdenlive from kdenlive.org for my operating system, then give me a five minute orientation on the timeline, since I am coming from CapCut.
FreeFlow: Download and install FreeFlow from github.com/zachlatta/freeflow, then help me pick a keyboard shortcut that does not clash with anything I already use.
What to expect, honestly
Two of these are just downloads. Kdenlive and FreeFlow are normal apps. Double click, done, five minutes each.
The rest need Docker and a bit of patience. Budget half an hour for your first one and about ten minutes for each one after, once you know the rhythm.
Free means free software, not free hosting. Running these on your own computer costs nothing. Running them on a server so they are online all the time costs a few dollars a month. That is still far less than the subscriptions they replace.
Start with the one that annoys you most. Not the one with the most stars. The tool you resent paying for is the one you will actually finish setting up.
Want help setting this up
Get stuck on any of the ten, bring it to the community. This stack comes up on the weekly calls constantly.
Join CC Strategic AI on Skool and post what you tried and what broke. You will get a real answer instead of a link to docs you already read.
The 5 Claude Code Plugins for UI/UX Design (September 2026)
Claude has one default look and you have seen it. These five free repos take it off autopilot. All five links, what each one fixes, and the order to install them.
All 5 repos
Every link goes straight to the source. All five are free and open source. Star counts checked September 2026.
4. Playwright CLI · github.com/microsoft/playwright-cli · 12,900 stars, Apache 2.0. Microsoft's tool. Ten invisible browsers that test the thing you just built.
Ask Claude to build you a page and you get the same page every time. Big centered headline. Purple gradient. Three rounded cards in a row. You have seen it on a hundred sites because everyone is asking the same tool the same way.
That is not Claude being bad at design. It is Claude having no idea what good looks like to you. Nobody told it. So it falls back on the average of everything it has ever read, and the average of everything is boring.
These five repos fix that from different angles. One shows Claude what good looks like. One writes your taste down as a rule. One blocks the lazy default. One checks the result actually works. The last one does something none of the others can do.
None of them cost money. Setting up all five takes about half an hour.
1. Awesome Design: real websites, taken apart
The problem it fixes: telling an AI "make it look like Notion" does nothing. It has no idea what that means in numbers.
What it does: this repo holds 73 files called DESIGN.md. Each one is a real website pulled apart into the parts that matter. The exact colors. The font sizes and how they step up. How much space sits between things. What the buttons and cards look like.
DESIGN.md is an idea from Google Stitch. It is a plain text file that tells a design agent how your project should look, the same way AGENTS.md tells a coding agent how to build it. No Figma file, no export, no setup. It is markdown, which is the format these models read best.
The part people miss: you are not copying somebody's website. You are taking the building blocks they used and putting your own thing on top. Same lumber, different house.
Install: there is nothing to install. Pick a site, copy its DESIGN.md into the root folder of your project, and tell Claude to build against it. Browse all 73 at getdesign.md.
2. Impeccable: your taste, written down once
The problem it fixes: you say "make it cleaner" and get a different answer every session, because you never wrote down what clean means.
What it does: Impeccable calls itself a design language. In plain terms, it puts a rules file inside your project that your AI tool reads before it builds anything. So the rule survives the session instead of dying with it.
It also ships a live editor in your browser. You change how something looks by looking at it, not by reading CSS in a terminal window and hoping.
It works across Claude Code, Cursor, Codex, Gemini CLI, GitHub Copilot and Grok Build, so the rules follow you if you switch tools.
Install: from your project folder, run this, then run /impeccable init inside Claude Code.
npx impeccable install
If you would rather add it as a Claude Code plugin, use the marketplace instead:
3. Taste Skill: kill the default before it gets built
The problem it fixes: the generic layout shows up before you get a chance to say anything about it.
What it does: Taste Skill is a skill, so it loads on its own when Claude starts working on design. It pushes Claude off the three or four layouts it hands everybody and toward choices that look deliberate.
Where Impeccable writes down your rules, this one stops the bad reflex whether or not you have written anything down yet. That is why both are on the list. They cover different halves of the same problem.
Install: installs through Vercel's skills CLI. This adds every skill in the repo:
The problem it fixes: everyone treats this as an automation tool. It is a design tool, and that is the part people are sleeping on.
What it does: Playwright CLI is Microsoft's tool for driving real browsers with no window on screen. Point Claude at it and it opens ten of them at once, fills in your form ten different ways, and reports back what broke.
Ask yourself when you last tested your own contact form by hand. With a bad email. With an empty field. On a phone screen. With the browser back button halfway through. Most people never do, and that is a design failure, not a code failure. A page that looks perfect and eats a lead is still a broken page.
It also records what you do in a browser and writes the test for you, and it grabs screenshots on command, which is handy when you want Claude to actually see what it built.
Install:
npm install -g @playwright/cli@latest
No global install? Run it through npx instead: npx playwright cli.
5. img2threejs: one photo in, 3D out
The problem it fixes: putting a 3D object on a website normally means modeling software you do not know and a file too heavy to load.
What it does: you hand it a photo of an object. It rebuilds that object as a Three.js model written in code. No modeling software. No mesh file to host. It comes out ready to spin and it checks its own work before handing it back.
Because it is code and not a downloaded model, it is small, it loads fast, and you can edit it after the fact by editing text.
Install: clone it straight into your Claude skills folder.
This is the one to save for last. The other four fix pages you are shipping this week. This one is the thing you add when the page already works and you want it to be memorable.
The order to install them
They stack. Each one is worth more once the one before it is in place, so this order saves you from redoing steps:
Awesome Design first. Nothing to install and it changes the output immediately. Copy one file, see the difference, then decide how deep to go.
Impeccable second. Now that you know what you like, write it down so it survives past this session.
Taste Skill third. Your rules cover the things you thought about. This covers the things you did not.
Playwright CLI fourth. The page looks right. Now find out whether it works.
img2threejs last. This is the flourish. Add it once the foundation is solid, not before.
Half an hour start to finish if you go in this order. Longer if you jump around, because you will end up re-running the same setup twice.
Want help setting this up
Get stuck on any of the five, bring it to the community. We run weekly calls and this stack comes up constantly.
Join CC Strategic AI on Skool and post what you tried and what broke. You will get a real answer instead of a link to documentation you already read.
EZ Thumb: Thumbnails Built From Your Competitors, Not Generic Rules
Every thumbnail guide hands you the same twelve rules. They were measured on somebody else's audience. This free skill onboards once against your actual competitors, reads every one of their thumbnails, and packages your videos against that instead.
npx ez-thumb install --with-commands · also on npm
The rules you were given are a photograph of a moving thing
Rule of thirds. Bright colors. Big face. Keep the text short.
None of that is wrong exactly. It's just not about you. Those rules got measured on somebody else's audience, in somebody else's niche, at some point in the past.
Here's how fast it moves. In one real fifteen day window in a single niche, a framing that was running at 19.55x median breakout dropped to 1.79x. Nothing changed except that everybody copied it and the audience stopped seeing it.
Fifteen days. Any static rule you were handed is already a snapshot of something that has since moved.
And half the thumbnails you're studying never earned a click
This is the part nobody mentions. Bought views are everywhere in AI, business, finance and crypto.
One channel in the corpus this skill was built against averaged 111,777 views on 56 likes. That's a 0.05 percent like rate while its neighbours in the same window ran 2 to 4 percent. Forty to eighty times below every peer, and flat across five straight uploads on a channel where organic variance normally swings 5x.
Their thumbnails looked incredible. Professionally produced sponsor assets. They proved absolutely nothing, because the clicks didn't happen the way they appeared to.
Copy that thumbnail and you have copied an advertisement.
So it builds you a corpus of your own
Once. Then it designs against it forever.
/ez-thumb-onboard ONE TIME
1 profiles your channel, assets, what has already worked
2 finds your competitors vidIQ MCP, or you name them
3 pulls their best thumbnails Apify MCP, 90 days, top 5 each
4 DOWNLOADS every image saved into the skill itself
5 reads every single one vision, one CSV row per thumbnail
6 flags the bought views like rate under 0.5 percent
7 writes all of it back into itself
/ez-thumb EVERY VIDEO AFTER
reads what it already learned, packages your video against it
The corpus is a one time build. It doesn't re-pull, it doesn't re-interview, and it never asks you to set the same thing up twice.
The competitor pull is what makes the output yours
This is the part worth understanding before you install it.
Step 4 isn't a cache. The thumbnails get downloaded and kept, filed by channel and sorted by performance:
skill/reference/2026-08-31-your-niche/
README.md the findings, with a generated: date stamp
records.csv one row per thumbnail, 15 columns
thumbnails/
competitor-a/000212554-2026-07-10-how-i-built-a.jpg
competitor-a/000103342-2026-07-09-gpt-made-this.jpg
competitor-b/000285219-2026-07-23-turn-this-into.jpg
That nine digit prefix is the zero-padded view count, so every channel folder sorts worst-to-best on sight.
Those saved images are then used as the working reference every time you run /ez-thumb. Not as background flavour. As a hard requirement:
Before it writes anything, it reads the corpus and names the current cliche in your niche, so your concepts either avoid it or break it on purpose
It ranks your niche's concept carriers by breakout, so it knows whether faces, screens, comparisons or objects actually win where you compete
Every concept it hands you cites a real thumbnail by filename. A concept with no citation is a guess, and it won't ship one
It opens the top outliers and tells you what they do that the median frame doesn't
Anything flagged as bought traffic gets labelled composition reference only out loud whenever it's cited
So the thumbnails you get out are shaped by the actual visual language of your competitors, in your niche, from the last 90 days. Not by the average of everybody's YouTube advice.
The part that actually makes it different: it rewrites itself
Most skills are a form you fill out every session. This one remembers.
After onboarding, the installed skill folder holds your channel profile, your chosen data source, your competitor corpus, and a dated findings section appended to its own frameworks. Two things follow from that:
The whole configured skill is one folder. Copy it to another machine and it's still onboarded. No hidden state file to lose.
npx ez-thumb update protects it. A version bump never costs you your onboarding.
And when new evidence contradicts something already written in a framework, it conditions the old rule instead of deleting it. The old finding keeps its date and its scope, the new one sits beside it. Six months later you can still tell whether the guidance changed because something was measured or because somebody had an opinion.
Two things it does that almost no thumbnail research does
1. It ranks on breakout, never raw views.
Channel A 30,000 subs 20,000 views -> 0.67x the thumbnail worked
Channel B 900,000 subs 200,000 views -> 0.22x it underperformed
Sort those by raw views and B looks like the winner, so you copy the wrong thumbnail. Nearly every thumbnail teardown on the internet makes exactly this mistake.
2. It flags bought views before it learns anything from them.
Like rate under 0.5 percent, in a set where the peers run 2 to 4, gets marked as paid inventory and pulled out of every aggregate. Still usable as a composition reference. Never cited as proof that a thumbnail earned a click.
What a run actually gives you
Three genuinely different concepts. Not three rewordings of one. Each carries a title, the concept object, one named curiosity lever, an ascii composition sketch with the bottom right kept clear for YouTube's timestamp, and either the corpus evidence backing it or the niche cliche it deliberately breaks.
Pick one and it writes five image prompts as five different visual bets: angle, moment, contrast, distance, emotional beat.
Then it scores everything on the Four Cs and takes the lowest, not the average. A frame that's a 10 on composition and a 4 on curiosity is a 4, because nobody clicks it. Pass bar is 8. Anything under comes back with one specific fix, never "make it pop."
What you need to run it
Required: the Apify MCP. Pulls the competitor videos, stats and thumbnails. A free token gets you going.
Optional: the vidIQ MCP. Auto-discovers who your competitors are. Skip it and you name them yourself, which is honestly often the better list.
Optional: any image generation MCP. Generates the variations.
Without an image generator you still get everything that matters: the concepts, the compositions, the corpus evidence, the scoring, and five copy-paste prompts for whatever generator you already use. Generation is the last mile, not the product.
npx ez-thumb doctor check what you have
npx ez-thumb status install state, onboarding state, corpus age
On other people's thumbnails
The thumbnails EZ Thumb pulls belong to the creators who made them. They land on your machine for your own analysis, the same way you'd open a competitor's channel and study it.
The package gitignores the corpus by default and ships with none of its own. Don't commit competitor thumbnails to a public repo, and don't redistribute a corpus. If you fork it, leave that rule alone.
It also writes its flags as data, not as accusations. "0.05 percent like rate, composition reference only" is a measurement. Anything stronger than that is a claim you don't need to make.
Where the knowledge came from
None of this is opinion. Every framework traces to a named source, and where the sources disagree it says so instead of quietly picking a side.
The forensic teardown. Five of the highest performing public thumbnail teachings on YouTube got run through a full forensic pass: downloaded, sampled at one frame per second, transcribed locally with Whisper, and read frame by frame with vision. All five transcripts end to end. 108 contact sheets. Nothing skimmed, nothing sampled.
That pass produced fifteen rules that survived the cross-check, a merged twelve format library, and seven places where these creators flatly contradict each other, each with the resolution the skill actually encodes. A few:
Research your niche, or refuse to. One says copy what's working. Another calls that the worst thing you can do, because it makes you the eleventh instance of a frame the audience already saw at 100K. Resolved as: research in-niche to find the cliche, take visual inspiration out-of-niche. Research is for diagnosis, not imitation.
Colour theory, or a fixed palette. The wheel maximizes contrast inside one frame. A locked five slot palette maximizes recognition across a channel page. Different problems, so use both.
Consistency, or formula fatigue. Consistency wins for about five videos, then starts costing clicks. Hold the font and palette skeleton, rotate the accent and the format.
The rules only one person teaches got kept too, because those are often the most valuable: the dark mode contrast trap, before goes left and after goes right, the 45 percent of viewing that happens on a TV, the recognizable interface workaround for creators with no famous face, and more elements equals less focus, fewer words equals larger text.
And the MasterClass it's built on
The backbone comes from the Thumbnail MasterClass by david, dill and ant. Go buy it, it's worth it. What EZ Thumb encodes from it:
The Four Cs — Composition, Colour, Clean Assets, Curiosity
The four curiosity levers, and good clickbait as exaggerating interest, never reality
The five step process — Needs, Sketch, Prompt, Mockup, Design
Packaging concepts before design, and the Safe / Emotion / Contrast / Leverage spread
Audience calibration by age, and the credibility versus energy split
Watch time above CTR, and testing three distinct ideas as paired title-plus-thumbnail units
The bottom right belongs to YouTube's duration stamp. Not one other source in the whole corpus mentions this, and every one of them shows example thumbnails with a timestamp sitting on top of the design
No slides, no deck text and no course material are redistributed. These are the ideas, restated, credited, and cross-checked against five independent sources.
Everything measured — the breakout ranking, the bought view filter, the pattern decay numbers — came out of production runs on my own channel.
Paste a video link into Claude and it reads the transcript. It never sees the cuts, the captions, or a single word on screen. This free skill fixes that, and it runs entirely on your own machine.
If you are still pasting video links into Claude and expecting it to watch the whole thing, it is not watching. It is reading.
Claude gets the transcript, so it knows what was said. It has no idea what was shown. Every caption, every cut, every sticker, every number burned onto the screen, gone. Then it summarizes the video back to you like it sat through the whole thing.
For a podcast that barely matters. For a competitor reel where the entire hook is a text card, you just threw away the part you were trying to study.
What it does instead
watchvideo downloads the video, takes actual frames off the screen, transcribes the audio on your own laptop, and hands Claude both halves at once. Then it writes the whole thing to a folder you can read later.
Four free tools, no API key, no account, no per-minute billing:
yt-dlp pulls the file down
ffmpeg extracts the audio and the visual frames
Whisper transcribes locally, so nothing gets uploaded anywhere
Claude reads the frames with its own eyes, inside the session you already have open
Nothing here talks to a paid service. A run costs you CPU time and that is it.
The storyboard trick (this is the whole thing)
Here is the part that makes it usable instead of theoretical.
The obvious way to give Claude a video is to rip a frame every second and hand over the pile. A ten minute video is 600 separate images. That does not fit in a context window. It does not fit in anything.
So watchvideo tiles them. Thirty frames get stitched into one storyboard image, laid out in a grid, with each cell still wide enough that you can read the burned-in captions. Those same 600 frames become 20 images instead of 600.
That one decision is the difference between analyzing a single clip and analyzing an entire competitor roster in one sitting. It saves your context and it saves your tokens.
Three modes, because not every video needs the same depth
You pick the mode at the start of every run. It is the first thing the skill asks.
Hook mode. First 20 seconds only, one frame per second, and the audio gets trimmed to match. One storyboard per video, flat, no matter how long the video is. This is the competitor sweep mode: point it at fifteen reels and you get fifteen images and a scoreboard comparing every hook side by side.
Condensed mode. The whole video, one frame every three seconds. This is the default, and it is what you want for one reel or for getting the shape of a longform.
Forensic mode. The whole video, one frame every single second. Use it when you want to go deep on one particular video and see the edit rhythm cut by cut.
And it will not quietly eat your window. Before it cuts a single frame it checks the duration and does the math. If a run would go over 15 storyboards it stops, shows you the real numbers, and offers you the cheaper mode.
What you get back
Run /watchvideo, drop in the link, pick your mode, and tell it what you want out of the run. What lands is a dated folder:
transcript.md keeps the source link and the full transcript in the same file, with real timestamps, so a transcript is never orphaned from the video it came from.
analysis.md is the breakdown: the spoken hook quoted word for word, the visual hook, whether those two actually agree, the format, the topic, a beat sheet, and web research on whether the person is even right.
The third file is yours. Every run it asks what you want out of this one. Ask for hook rewrites, a list of every claim they made, a shot list, a comparison against your own video, and that becomes its own file in the same folder.
The hook report breaks the opening down second by second: what was spoken, what text was on screen, any stickers, and what the person's body language was doing the whole way through.
Install it in 60 seconds
One command:
npx watchvideo install
That puts the skill in ~/.claude/skills/watchvideo/. Restart Claude Code, type /watchvideo, and you are in.
A few others if you need them:
npx watchvideo install --with-commands also drops a /watchvideo command stub.
npx watchvideo install --project installs into just the current project.
npx watchvideo update pulls the newest version over an older copy.
npx watchvideo doctor checks your machine and tells you exactly what is missing.
You need three free things installed once. On a Mac:
No API key, no account, no subscription. Run npx watchvideo doctor and it will tell you if anything is missing and how to fix it.
What it will not do
I would rather tell you the limits now than have you find them later.
It samples, it does not see every frame. Even forensic mode is one frame per second on 30fps footage, so it sees one frame in thirty. Every run stamps the real percentage in the file instead of claiming it watched the whole thing.
Hook mode only knows the first 20 seconds. Not the middle, not the CTA. So it drops the beat sheet and the topic research rather than making them up.
It cannot tell you why a video performed. Retention and traffic source are not in the file, and it says unverified instead of guessing.
Public video only. No login walls, no age-gate workarounds. Analyze, do not repost.
YouTube blocks some downloads. Keep yt-dlp current. When one fails anyway, the skill says so instead of pretending.
Where to take this next
The obvious use is competitor research. Point hook mode at every creator in your niche, go to bed, and read the scoreboard in the morning. That is what I built it for.
If you want help wiring this into an actual content system instead of running it one-off, join CC Strategic AI on Skool. Post what you tried and what broke and you will get a real answer.
I am Charles J Dove, and I run Charlie Automates. At my agency CC Strategic we build these systems for clients, and everything I test lands free on charlieautomates.com first.
Not sure what to fix first? The free AI Bottleneck Finder tells you the one constraint capping your revenue right now, plus two builds you could ship this week. Five minutes, no call required. Take the free AI Bottleneck Finder →
The 5 Claude Code Plugins Worth Installing in 2026
Most of you are paying $200 a month for limits these five plugins remove. Here are all five repos, what each one actually does, and the order to install them.
All 5 repos
Every link below goes straight to the source. Nothing here costs money.
Claude Code out of the box is one assistant, in one window, that forgets your project every time you close it. That is the version most people are paying for.
These five plugins change that. You get a team of agents instead of one. The team remembers your project. It keeps working after you hit your usage limit. And it gets a little better every week without you doing anything.
None of them cost money. The only thing they cost is the twenty minutes it takes to set them up once.
1. herdr: run a whole team of agents at once
The problem it fixes: you can only babysit one Claude window at a time. Close your laptop and whatever it was doing dies with it.
What it does: herdr is a background server that your terminals live inside. Each agent gets its own terminal, and they all keep running on the server even if you close the lid, lose wifi, or restart the machine. When you come back, they are still there, still working.
It also watches each agent and figures out on its own whether that agent is working, blocked and waiting on you, done, or idle. When one finishes or gets stuck, it plays a sound so you know to look. You are not tabbing between windows guessing who needs you.
Agents can also talk to each other through it. One agent can start another one, send it a message, and wait until it is actually blocked before stepping in.
Install:
curl -fsSL https://herdr.dev/install.sh | sh
It runs Claude Code, Codex, Cursor, OpenCode and others, so you are not locked to one tool. Written in Rust, which is why it is fast and does not eat your battery.
2. Graphify: a memory that does not reset
The problem it fixes: Claude reads your files fresh every session. You pay for that reading every single time, and it still does not understand how your project fits together.
What it does: Graphify reads a folder of stuff and builds a map of it. Not a list of files, an actual map: which pieces exist, and which pieces are connected to which. Then your agent asks the map instead of reading everything again.
It works on more than code. Point it at PDFs, markdown notes, images, audio, or video and it maps those too. I stack mine with Obsidian so I can see the map with my own eyes instead of trusting that it is in there.
The part I like most: every connection it makes is labeled extracted, inferred, or ambiguous. So you always know the difference between something it actually found and something it guessed.
To be clear, this is not my tool. It is safishamsi/graphify. I use it on every project and I am recommending someone else's work.
3. OmniRoute: keep working after you hit your limit
The problem it fixes: you hit your Claude usage limit in the middle of something and just sit there.
What it does: OmniRoute is a small program that runs on your own computer, in between Claude Code and the internet. Normally Claude Code talks straight to Anthropic. With OmniRoute running, it can talk to hundreds of other AI companies instead, and a lot of those have free plans.
So when your Claude limit runs dry, OmniRoute hands the work to a free model and you keep going. When that one runs out, it hops to the next. You do not do anything, it just keeps switching.
Last count it reaches 330 or so providers and 1,200 or so models, including Kimi, GPT, Gemini, GLM and DeepSeek. Its own dashboard tracks roughly 1.6 billion free tokens a month across those free tiers. It is MIT licensed and built by 320 plus contributors.
One honest note: "free" here means other companies' free tiers, not infinite. Free tiers change and sometimes disappear. The reason this still works is that OmniRoute keeps moving you to whichever one is up.
4. Claude Code Setup: Anthropic tells you what to install
The problem it fixes: there are thousands of hooks, skills, subagents and MCP servers out there. You have no idea which ones matter for the project in front of you.
What it does: this one is from Anthropic themselves. You run it on a project and it reads your codebase, then hands you a list of the hooks, skills, subagents and MCP servers that fit that project.
Think of it like a contractor walking your house before you go buy tools. Same advice everyone gives, except it looked at your actual house first.
Run it once on every new project, take the recommendations, and move on. It is not a daily thing.
One correction to what you may have heard elsewhere: it recommends. It does not delete anything or clean anything out for you. You still decide what to install.
5. Task Observer: your skills get better on their own
The problem it fixes: you correct Claude the same way over and over, and that lesson dies when the session ends.
What it does: Task Observer sits in the background while you work and takes notes. Every time you correct it, or make a judgement call, or something works unusually well, it writes that down in a log file.
Later, on a review cycle, it goes back through those notes, compares them against every skill you own, and writes up the improvements. Then it hands you the changes to look at.
That last part matters: nothing goes live by itself. It gives you a diff and you decide whether to install it. So it is not silently rewriting your setup behind your back, which is exactly what you should want from a tool with this much reach.
It is called "One Skill To Rule Them All" on GitHub, and Task Observer is the skill inside it that does the watching.
The order to install them
They stack. Each one removes the ceiling the one before it runs into, so this order saves you time:
OmniRoute first. No point setting anything else up if you are going to run out of usage an hour in.
Graphify second. Now you can work as long as you want, so make it stop forgetting what you are working on.
herdr third. One agent that remembers everything is still one agent. Now go wide.
Claude Code Setup fourth. More agents means more ways to set it up wrong. Let Anthropic's plugin tell you what belongs.
Task Observer last. It improves the other four, so it should be watching a setup that is already worth improving.
Total time: about twenty minutes if you go in this order. Longer if you skip around, because you will end up redoing steps.
Want help setting this up
If you get stuck on any of the five, bring it to the community. We run weekly calls and this exact stack comes up constantly.
Join CC Strategic AI on Skool and ask in the feed. Post what you tried and what broke, and you will get a real answer instead of a link to documentation you already read.
Mission Control: The Claude Code Stack I Actually Run in 2026
Most of you are paying $200 a month for limits these five tools remove. Here are all five repos, what each one actually does, and the order to install them. BASE is number one and it is not close.
All 5 repos
Every link below goes straight to the source. Nothing here costs money to start.
1. BASE · github.com/ChristopherKahler/base · gives Claude Code a memory of your whole workspace and lets several agents work the same repo without colliding. Free, with a paid tier if you want it.
Claude Code out of the box is one assistant, in one window, that forgets your project every time you close it. That is the version most people are paying for.
These five change that. You get a team of agents instead of one. The team remembers your workspace. It keeps working after you hit your usage limit. And it gets a little better every week without you doing anything.
The order below is not the order I found them in. It is the order that costs you the least rework.
1. BASE: the one that matters most
Want the deep dive instead of the summary?BASE: A Knowledge Graph For Your Rules is the full walkthrough in this same toolkit. What follows here is the short version and why it sits at number one.
The problem it fixes: every new session, Claude re-reads your repo like it has never seen it. You re-explain the same architecture, the same conventions, the same decisions you already made three weeks ago. Then you open a second window and the two agents overwrite each other.
What it does: BASE is a Rust binary that sits underneath Claude Code and holds the state Claude does not. It parses your codebase into a knowledge graph with tree-sitter across 35+ languages, so an agent can ask base ast query -c "auth" and get every matching function, or --calls "validate" and get every caller, instead of grepping blind.
On top of the graph it keeps things Claude has no place to put: decisions and the reasoning behind them, domain rules, project and task state, and recallable notes. It injects the relevant pieces automatically at four points, session start, on your prompt, before a tool runs, and after results come back. You do not ask for context. It arrives.
Better than Graphify for memory and session context. Graphify maps your code and gives Claude a graph to read. BASE does that and then keeps the layer Graphify does not: decisions, domain rules, project state, and the session-to-session continuity that makes the next conversation start where the last one ended. Graphify is a map. BASE is a map plus a memory plus a set of rules that fire on their own.
The part nobody talks about: the relay. Multiple named agent sessions work the same repo at the same time and coordinate through it.
Here is a real exchange out of my own relay log, two agents named puffin and marlin working this site:
"I did NOT re-add a local useCheckout. I hit that exact conflict on rebase, took origin/main's enrollment.tsx, and put trackGA('begin_checkout') in use-checkout.ts next to your ported trackMeta('InitiateCheckout')."
That is one agent telling another how it resolved a conflict on a file they were both touching. In the same message it flagged a bug in the other agent's work that neither of them was looking for:
"your schedule.ts ASCII normalisation left my threeDayLabels emitting en dashes, so a 3-Day row read differently from the 2-Day row beside it in the same dropdown. Fixed in 491f377."
Before touching shared files it announces it: "Shout NOW if you or Sal have anything in flight on those two files." It holds claims on directories and releases them when asked. And when something is a judgment call rather than a code problem, it stops and kicks it up: "Escalating to Charles, this is his call not ours."
That is the difference between running four Claude windows and running a team. The other four tools on this list make one agent better. BASE is the only one that makes several agents work together without stepping on each other.
Install:
cargo build --release
./target/release/base install
cd ~/my-workspace
base scaffold
The installer drops the binary at ~/.local/bin/base, creates global config in ~/.base-gbl/, and wires the hooks into your ~/.claude/settings.json. Windows users can take the prebuilt zip from Releases instead of building. Current version is 0.13.2.
The repo is free. There is a paid tier on top of it if you want more, but everything above works without paying anything.
2. Herdr: run a whole team of agents at once
The problem it fixes: you can only babysit one Claude window at a time. Close your laptop and whatever it was doing dies with it.
What it does: Herdr is a background server that your terminals live inside. Each agent gets its own terminal, and they all keep running on the server even if you close the lid, lose wifi, or restart the machine. When you come back, they are still there, still working.
It also watches each agent and works out on its own whether that agent is working, blocked and waiting on you, done, or idle. When one finishes or gets stuck, it plays a sound so you know to look. You are not tabbing between windows guessing who needs you.
Pairs directly with BASE. BASE decides what every agent knows and stops them colliding. Herdr decides where they run and tells you which one needs a human.
Install:
curl -fsSL https://herdr.dev/install.sh | sh
It runs Claude Code, Codex, Cursor, OpenCode and others, so you are not locked to one tool. Written in Rust, which is why it is fast and does not eat your battery.
3. OmniRoute: keep working past your limit
The problem it fixes: you hit your Claude usage limit mid-task and the day stops.
What it does: OmniRoute sits between you and the model and routes your requests to whatever you still have capacity on. You keep working instead of watching a reset timer.
Free, MIT license, so there is nothing to think about on the legal side.
Install it first if you are only going to install one thing today. Everything else on this list is wasted if you run out of usage an hour in.
4. Claude Code Setup: Anthropic's own plugin
The problem it fixes: more agents and more tools means more ways to configure it wrong, and most people are guessing.
What it does: it looks at the project you are in and tells you what belongs there. Anthropic ships it themselves, which means it tracks what Claude Code actually supports rather than what a blog post said six months ago.
Free. Low effort, and it stops you carrying a bad setup from one project into the next.
5. Task Observer: gets better while you work
The problem it fixes: your skills stay exactly as good as the day you wrote them.
What it does: it watches how your other skills perform in real work and improves them off what it sees. You do not run a training step. It just gets sharper the more you use everything else.
Free, CC BY 4.0.
Install it last on purpose. It improves whatever it is watching, so it should be watching a setup that is already worth improving.
Install in this order
OmniRoute first. No point setting anything else up if you are going to run out of usage an hour in.
BASE second. Now you can work as long as you want, so make it stop forgetting your workspace. Do this before you go wide, because BASE is what keeps multiple agents from overwriting each other.
Herdr third. One agent that remembers everything is still one agent. Now go wide, with BASE already underneath keeping them coordinated.
Claude Code Setup fourth. More agents means more ways to set it up wrong. Let Anthropic's plugin tell you what belongs.
Task Observer last. It improves the other four, so it should be watching a setup that is already worth improving.
Total time: about half an hour if you go in this order. Longer if you skip around, because you will end up redoing steps.
Want help setting this up
If you get stuck on any of the five, bring it to the community. We run weekly calls and this exact stack comes up constantly.
Join CC Strategic AI on Skool and ask in the feed. Post what you tried and what broke, and you will get a real answer instead of a link to documentation you already read.
Automate Your Facebook Ads With Claude Code (Meta Ads MCP)
Meta has an official MCP for ads. Connect it to Claude Code and you can build campaigns, ad sets, creatives and copy by typing what you want.
Building is not the problem anymore
Fable 5.1 dropped and somebody rebuilt Cities: Skylines with it. He called it Fable Cities. That took a day.
So here is the honest read. Everybody can build incredible things now. Almost nobody can get them in front of a human being.
The real problem is not the build. It is distribution. A great product does not hand you users and it does not fill your bank account.
Which means you point Claude Code at the other half of the job. Marketing, paid traffic, and the follow-up that turns a lead into a customer. This page is where that starts.
What this is
Meta runs an official MCP server for ads at https://mcp.facebook.com/ads. An MCP is a link that gives Claude real controls inside another tool. Hand Claude this one and it can work inside your ad account instead of telling you where to click.
Once it is connected you stop building campaigns by hand. You describe the campaign, and Claude builds it. Campaign, ad sets, targeting, budget, creatives, copy.
Setup takes about ten minutes. Most of it is Meta paperwork, not code. Follow the steps in order and you only do this once.
What you need before you start
A Facebook account you use for business.
A Meta Business Manager (also called a business portfolio) with your ad account and Page inside it.
Claude Code installed on your computer.
A privacy policy URL you can paste. Your website's privacy page is fine.
Give the app a name you will recognize later, like "Charlie Ads". Nobody else sees it.
On the Use cases screen, check these two boxes:
Create & manage ads with Marketing API
Create & manage app ads with Meta Ads Manager
On the next screen, pick your business portfolio. If you do not have one yet, create it here. Then finish creating the app.
Step 2: Publish the app
A brand new app is in development mode, and Meta will not let it touch a live ad account until you publish it.
In the left menu of your app, open Publish. Meta asks for a Privacy Policy URL. Paste your privacy page and fill in whatever else it asks for. Then click publish.
That is the step most people skip, and then nothing works later. Do it now.
In the left menu go to Accounts, then Apps. Click Add and paste the App ID of the app you just made. You will find the App ID on your app's dashboard on developers.facebook.com.
Your app is now owned by your business, which is what lets it reach your ad account.
Step 4: Create a system user
A system user is a robot account. It is how you give Claude access without handing over your personal login.
Still in Business Manager settings, go to Users, then System users. Click Add, name it anything you like, and set the role to Admin.
Step 5: Assign the assets
Open the system user you just made and click Assign assets. The robot can only touch what you hand it, so give it everything the ads need:
Your ad account, with full control.
Your Page, with full control.
The app you created in step 1. This one is easy to forget and nothing works without it.
Your Instagram account and catalog, if you run ads on either.
Step 6: Generate the token
On the same system user, click Generate new token. Pick the app you created, then tick these permissions:
ads_management
ads_read
business_management
pages_show_list
instagram_basic
catalog_management
ads_mcp_management
Copy the token the moment it appears. Meta shows it once and never again. If you lose it, generate a new one.
Step 7: Give Claude the link and the token
Open Claude Code in your project and say this:
Set up the Meta Ads MCP for me.
Server URL: https://mcp.facebook.com/ads
Token: [paste your token]
Claude adds the server, connects with your token, and confirms it can see your ad account. If it cannot, the usual cause is a missing asset from step 5 or an app that never got published in step 2.
Step 8: Ask for the campaign
Now you just describe what you want in plain English:
Build a lead generation campaign for my agency.
Budget $40 a day, United States, ages 28 to 55,
business owner interests. Write three ad variations
with different hooks and set them all live paused
so I can review them first.
Claude builds the campaign, the ad sets, the targeting, the creatives and the copy. Then you open Ads Manager and it is all sitting there.
Ask it to leave new campaigns paused until you have seen them. That way you review before a dollar moves.
Things you can ask for after setup
Fix what is losing money. "Show me every ad set over a $40 cost per lead this week and pause them."
Scale what is working. "Find my best ad by cost per lead and raise its budget 20%."
Test more copy. "Clone my top ad and write five new hooks for it."
Report without logging in. "Give me last week's spend, leads and cost per lead by campaign."
This is the part clients pay agencies a monthly retainer for. It is now a sentence.
Set up your Meta pixel before you spend a dollar
The pixel is a small piece of code that sits on your website. Its whole job is to tell Meta what people did after they clicked your ad. Landed on the page. Gave you an email. Booked a call. Paid.
Here is why that decides whether the ads work. When you run a leads campaign, Meta is not looking for people who click. It is looking for people who convert, and the only way it knows what a conversion looks like is the pixel telling it. No pixel means no examples. No examples means Meta shows your ad to whoever is cheapest to reach, and you pay for traffic that was never going to buy.
I learned this the expensive way, twice, on my own account.
A pixel that was never installed. I created the pixel for Your Claude Academy in August and left it. Weeks later Events Manager still said the last event was received in 1970, which is what it shows when the answer is never. The pixel existed. The code was never on the site. Meta had learned nothing the entire time.
A pixel that was installed and still reported nothing. Later, on a live leads campaign, a real lead came through into GoHighLevel while the pixel recorded zero Lead events for the whole campaign, against real spend. Ad blockers had stopped the pixel in the browser, and a blocked pixel does not throw an error. It just quietly reports nothing while the campaign keeps optimizing toward a signal it has never once received.
So the pixel is two jobs, not one. The code in the browser, and a server-side copy called the Conversions API that the visitor's browser cannot block. Both send the same event with the same ID, and Meta collapses the pair instead of counting your one lead twice.
Pick whichever of these two you are. The instructions are the same either way.
Give this to Claude Code (copy the whole block)
Paste this into Claude Code inside your website project. It builds the browser pixel and the server-side half in one pass, and it names the traps that cost me money.
Set up the Meta pixel on this site properly. Properly means all of this.
1. Base pixel. Put Meta's base snippet in the root layout so every page
inherits PageView. Do not paste it page by page, because a pixel you
have to remember to add is a pixel that goes missing invisibly.
2. If this is a single page app, Meta's copy-paste snippet only fires
PageView on a hard reload, so I would get one PageView per session
instead of one per page. Fire PageView on every route change too, and
guard the very first render so the base snippet and the route change
do not both count the landing hit. Double counting reads as an
artificially low cost per result.
3. Wrap fbq in one typed helper. Nothing else in the codebase touches
window directly. The helper no-ops silently when fbq is missing, so
an ad blocker can never take a checkout button down with it.
4. Use standard event names only: PageView, ViewContent, Lead,
SubmitApplication, InitiateCheckout, Schedule, Purchase. A custom
event name cannot be optimized toward until someone hand-builds a
matching Custom Conversion in Events Manager, so do not invent one.
5. Add the Conversions API on the server for every event that must
actually arrive, which is Lead, SubmitApplication and Purchase.
Hash email, phone and name with SHA-256 after trimming and
lowercasing. Phone is digits only, including country code, with no
plus sign. Never send an empty string as a match key, leave the key
out instead, because an empty value drags Event Match Quality down.
Send client IP and user agent raw, never hashed. Build fbc as
fb.1.TIMESTAMP_IN_MS.FBCLID and read the _fbp cookie server side.
6. Mint one event ID in the browser and send it to both the pixel and
the Conversions API so Meta deduplicates the pair.
7. Every Conversions API call swallows and logs its own errors. A
conversion event is telemetry. Losing one is a reporting gap. Failing
the request that carries it loses an actual lead, so telemetry must
never be able to do that.
8. Add the noscript tracking image fallback.
Then tell me which environment variable holds the access token, and how
to confirm each event inside Events Manager Test Events.
If your checkout happens on someone else's domain, tell Claude that too. The last event you can fire from your own site is InitiateCheckout, and Purchase has to fire on whatever page they get sent back to.
Do it yourself (the click path)
Make the pixel. Go to Events Manager, click Connect data source, choose Web, then Meta Pixel. Name it something you will recognize. Meta hands you a long number. That number is your pixel, and you will paste it everywhere.
Put the code on your site. Meta shows you a block of code to copy. It goes in your site's site-wide header, not on one page. Wix, Squarespace, Shopify and WordPress all have a custom code or header scripts box in settings. That box is where it goes. Some of them have a Meta pixel field where you just paste the number instead.
Check that it is actually there. Install the free Meta Pixel Helper extension in Chrome, then open your own site. The icon should show your pixel firing PageView. Nothing showing means the code is not on the page, and no amount of waiting will change that.
Watch yourself convert. In Events Manager open your pixel and click Test Events. Paste your site address, then go walk your own funnel. Opt in with your own email. Every step should pop up in that list within seconds. This is the only check that proves the thing works, so do not skip it.
Fire an event on the action that matters. PageView alone is not enough to optimize on. You want Lead on your thank-you step, Schedule when someone books a call, and Purchase with a dollar value when someone pays.
Look at Last Received. Back on the Events Manager overview, your pixel should show a recent time. If it shows 1970 or blank, it has never fired once. That was my exact situation for weeks.
Do not point a campaign at a rare event. Wait until the event happens about fifty times a week before you optimize toward it. Below that the ad set never finishes learning and your delivery gets worse, not better.
Then add the Conversions API. The browser pixel gets blocked and tells you nothing when it does. Have Claude build the server-side copy, or turn on the Conversions API connection your CRM offers if it has one. This is the step almost nobody does, and it is the one that keeps the campaign learning.
Do this before you turn on spend, not after. Everything Claude built you in the steps above is only as smart as the events you feed back to Meta.
Ads bring the leads in. Here is what closes them.
Say the ads work and a hundred people a day start asking about the thing you built. Now what?
Speed to lead decides that. Reach someone in five minutes and you are in the conversation. Reach them tomorrow and you are a name in an inbox.
Two pieces do that job for me, and Claude Code builds both.
Chloe AI calls the lead. Chloe is the AI sales teammate built into Close CRM. A lead hits the pipeline, Chloe calls it, qualifies it, handles the questions, and books the call straight into your calendar. No pickup and it leaves a voicemail. Every call gets transcribed and logged. Mine calls the leads that came in overnight, the ones I was never going to get to, and books them. Usage based pricing on any Close plan.
GoHighLevel runs the follow-up. The emails and texts that fire after someone opts in, and again after they book. That is the part nobody wants to build by hand, and Claude Code will build it for you. This is my affiliate link, and there is a free trial on it.
These two are not an either/or, they run in tandem. Chloe is the caller and GoHighLevel is everything around the call. The lead opts in, GoHighLevel tags it and starts the text and email sequence, Chloe dials it inside a few minutes and books the appointment, and the GoHighLevel workflows keep working the ones who did not pick up. One handles the conversation, the other handles the memory and the follow-through.
Once the CRM is doing the follow-up, wire it into the terminal too. See GoHighLevel MCP: Run Your CRM From Claude Code, which is how I triage the whole pipeline every morning from one command.
Meta Ads MCP for the traffic, Chloe for the call, GoHighLevel for the follow-up. That is the stack sitting behind about $5,000 a week for me right now, and every piece of it is a sentence typed into Claude Code.
Keep your token safe
That token can spend your money. Treat it like a credit card, not a password reset link.
Never paste it into a chat, a document, a screenshot or a public repo.
Never give it to a client or a contractor. Make them their own system user instead.
If you think it leaked, delete the system user in Business Manager. That kills the token instantly.
Test on a small budget first. Claude follows what you asked for, so a typo in a budget is still a typo in a budget.
Get on the weekly calls
Join CC Strategic AI on Skool
Connecting the MCP is the easy part. Running paid traffic that actually closes is the part that pays. We work through real campaigns together on the weekly calls, live.
Every member gets the community, the courses, and the weekly call replays.
Premium members get the weekly live strategy calls, monthly live workshops, and the exclusive Claude Code: From Zero To Dangerous course.
You installed the GoHighLevel MCP and found out it cannot build workflows. That is not a bug. GoHighLevel's public API has no endpoint for writing a workflow at all.
Check the official SDK and you find two workflow methods: read one, and remove a contact from one. There is no create. So the MCP cannot build automations, and it never will on the public API.
Which leaves you paying for their AI builder to do the one thing you wanted your own AI to do.
How it gets around that
GoHighLevel's own web app builds workflows every day. It does that by calling an internal API that is not in the public docs.
This repo documents those calls. Give that documentation to Claude Code, and Claude can make the same calls the GoHighLevel interface makes. That is the whole trick.
Read this before you install
This is the honest part, and it matters more than the setup.
The key you copy is your account password. Not a scoped API key. Anyone holding it can mint fresh access to your GoHighLevel account until you revoke it. Never commit it, never post it, never send it to a client or a contractor.
Use a throwaway sub-account first. Point this at a test location until you have seen it work. Never test on a client's sub-account.
It can break without warning. This is an internal API. HighLevel does not support it. Any front-end update on their side can change the call shapes and stop it working mid-build.
HighLevel has not blessed this. Depending on how you read their terms, using an undocumented internal API may sit outside them. Your call, your account.
No license on the repo. That means no permission is formally granted for use or modification. Fine for personal testing, worth thinking about before you build a product on it.
None of that means do not use it. It means use it with your eyes open, on your own account, not a client's.
Step 1 — Install it
Copy the repo link and paste it into Claude Code:
Install this for me: https://github.com/drleadflow/ghl-automation-builder
Claude clones it and reads the schemas. Let it finish while you do step 2.
Step 2 — Grab the key from your browser
Open GoHighLevel in Chrome, then:
Right-click anywhere on the page and choose Inspect.
In the panel that opens, click the arrows at the top and choose Network.
Set the filter to Fetch/XHR.
Click your Automation tab. The list fills with requests.
Right-click any row with the orange icon, hover Copy, then click Copy as cURL.
You now have a full request on your clipboard, and your login key is inside it.
Step 3 — Hand it to Claude
Back in Claude Code:
I want to use the automation builder. Here is my request:
[paste]
Claude pulls the key and the location ID out of that request and is ready to build.
Step 4 — Ask for the workflow
Now describe what you want in plain English. Because Claude Code already knows the pages and code in your project, you can point at them:
Build a lead intake workflow in GoHighLevel that connects to the
opt-in page we just built. Tag the contact, send a welcome text,
wait 10 minutes, then email me if they have not booked.
Claude builds the step graph, attaches the trigger, and publishes it. Open GoHighLevel and it is sitting there on the canvas.
What to build with it
The value is not one workflow. It is that you can now build the complicated ones without clicking through a builder for an hour.
Page and workflow in one sitting. Build the landing page with Claude Code, then have Claude build the intake automation that matches the exact fields on that page. No mismatched field names.
Client onboarding you can clone. Build one onboarding automation properly, then have Claude clone it per client with the names and offers swapped.
Speed-to-lead. New lead comes in, text goes out in under a minute, sales rep gets a task, no-response leads roll into a follow-up branch after an hour.
Recovery flows. Booked but no-showed, started checkout but did not pay, opened three emails but never replied. These are the branch-heavy ones nobody wants to build by hand.
Reporting nudges. A weekly workflow that fires a webhook at your own dashboard so you get a summary instead of logging in.
Wiring it to your own tools with webhooks
This is where it stops being a CRM and starts being infrastructure. Workflows can talk to anything you build.
Going out. Add a Custom Webhook action and the workflow sends the contact's details to any URL you choose. Ask Claude for it directly:
Add a custom webhook step that POSTs the contact name, email,
phone and the form answers to https://myapp.com/api/new-lead
Your own app receives that and can score the lead, write to your database, or ping Slack.
Coming in. GoHighLevel workflows can also start from an inbound webhook. So your app, your site, or another automation can kick off a GoHighLevel sequence whenever you want.
Both directions. Claude Code builds the endpoint on your site and the workflow that calls it, in the same session, using the same field names on both ends. That is the part that usually breaks when a developer and a CRM person build the two halves separately.
Where this fits with the GHL MCP
Keep both. They do different jobs and neither replaces the other.
This repo: building, cloning and publishing workflows. That is all it does.
Ask for an automation, use this. Ask for anything else, use the MCP.
Want to go deeper?
Building one workflow is easy. Running an agency where every client's automations, pages and follow-up are built and maintained by Claude Code is a different skill, and it is the one that actually pays.
That is what we work on inside the community: real builds, real client systems, and the Claude Code workflows behind them.
Book to Skill: Turn Any Book or PDF Into a Claude Code Skill
Feed Claude a book PDF and it makes things up. This turns the book into a real skill instead, so Claude pulls the right chapter and answers from the actual pages.
You point it at a book, a PDF, or a folder of documents. It reads the whole thing once and builds a Claude Code skill out of it.
From then on, the book is a command you can call in any session. You ask a question, Claude opens the right chapter, and answers from the real text.
Why you need it
You have probably already tried dropping a PDF into Claude and asking about it. Two things go wrong.
It makes things up. The model skims, loses the thread, and fills the gaps with something that sounds right.
It burns tokens. Every question makes it hunt through the file again. The repo measures this at 24x to 51x more tokens than asking a proper skill.
Book to Skill fixes both by doing the hard reading once, up front, and saving the structure.
Use the book as a framework, not a reference
This is the part most people miss. You are not building a search box for a PDF. You are giving your agent a working method it can apply while it does real work.
Turn a book on offers into a skill, then tell Claude to write your sales page using that skill. It runs the author's framework on your business instead of quoting the book back at you.
Same move works for:
Your own docs. Runbooks, onboarding guides, architecture notes. Fold a whole docs/ folder into one skill and ask it while you build.
Brand and voice guides. Turn a 60 page brand book into a skill your agent checks before it writes a single line.
Specs and standards. API contracts, compliance rules, RFCs. The stuff you look up constantly and never memorize.
Research. A stack of papers plus your notes, merged into one skill you can add to later.
If you keep reopening a document, it belongs in a skill.
What it builds
One conversion gives you five files in your skills folder:
SKILL.md, the core ideas plus a chapter index. Roughly 4,000 tokens.
chapters/ch01-*.md, one file per chapter. These only load when you ask about that topic, so they cost nothing until you need them.
glossary.md, every key term in alphabetical order with chapter references.
patterns.md, the techniques and methods pulled out on their own.
cheatsheet.md, decision tables and quick rules.
It pulls out frameworks, principles, techniques, and anti patterns. It is not a book summary. It is a toolkit.
Formats it reads
PDF, EPUB, DOCX, HTML, Markdown, plain text, and RTF out of the box. MOBI and AZW work too if you have Calibre installed.
It asks you a few questions about how you want the book used, then builds the skill. Give it a folder or a wildcard instead of one file and it merges everything into a single skill.
After that, call the book by name in any session:
/your-book-slug pricing
Claude opens the chapter that covers pricing and answers from it.
Four ways to run it
Full conversion. The default. Give it a path, get a finished skill.
Analyze only. Say "analyze" and it shows you the frameworks it found before it writes anything.
Generate from analysis. Hand it notes you already have and skip straight to building.
Update and fold in. Point it at new sources plus an existing skill and it merges them.
One thing to be clear about
Use books you actually bought. This reads a file you already own and turns it into a private skill on your own machine. It is not a way to get around buying the book, and it does not put anything on the internet.
Works with more than Claude
It follows the open Agent Skills standard, so the same install works in GitHub Copilot CLI and Amp as well as Claude Code. There is also a pip install book-to-skill option, but that only gives you the text extractor, not the slash command. Use the git clone above.
Obsidian Web Clipper: Save Any Webpage as Clean Markdown for Claude Code
Stop copy-pasting webpages into Claude. This free Chrome extension saves any page as clean markdown, drops it straight into your Obsidian vault, and now Claude can just read the file.
This Chrome extension helps you save any webpage as clean markdown for Claude Code to read. And the coolest part? You can add it directly into your Obsidian vault.
Here's why that matters. Claude works best when it can read a file sitting on your machine. A webpage is the opposite of that: it's locked in a browser, wrapped in ads, menus, and popups. The old way was to copy-paste the text into your chat or hand Claude a link and hope it grabs the right stuff. Now you don't have to do either. Clip the page once, it lands in your vault as a clean markdown file, and Claude can read it any time. Just reference the file and have a conversation about it.
What it does
One click, clean markdown. It strips out the ads, nav bars, and junk, and keeps the actual content.
Saves straight into Obsidian. Your vault is just a folder of markdown files on your computer, which is exactly the format Claude Code loves.
Clip the whole page or just a piece. Grab the full article, a selection, or highlights.
Free. The extension is free, and Obsidian is free for personal use.
Setup in 2 minutes
Download Obsidian (free) at obsidian.md/download and create a vault. A vault is just a folder where your notes live. Name it something like research.
Point it at your vault. Open the extension's settings and set your vault name and the folder where clips should land (I use a clippings folder).
Clip something. Go to any article, hit the extension icon, and save. Open Obsidian and the page is sitting there as a markdown file.
The playbook: feed your clips to Claude Code
This is where it gets good. Open Claude Code in (or pointed at) your vault folder and the clips become context Claude can actually use.
Research a topic. Clip 5 articles on a subject, then ask Claude: "Read everything in my clippings folder and give me the 5 big takeaways with sources."
Steal what works. Clip a competitor's landing page, then: "Read this clip and break down the offer, the hooks, and what I should test on my own page."
Save docs that keep moving. Clip the docs page of a tool you use, then reference the file whenever you ask Claude to build with that tool. No more hoping Claude's memory of the docs is current.
Turn reading into content. Clip an article that fired you up, then: "Read this file and draft my take on it as a LinkedIn post in my voice."
Build a swipe file. Every good email, ad, or hook you see, clip it. Ask Claude to pull patterns from the whole folder when you write something new.
The pattern is always the same: clip it once, then just reference the file and talk about it. No copy-paste, no link-grabbing.
Go further: turn the vault into a brain
Once your vault has real research in it, level it up. Graphify turns the whole vault into a knowledge graph Claude can query instead of re-reading every file, and it's the same brain-building move I use in my own agentic OS. Pair it with NotebookLM for deep research and Understand Anything when the thing you're mapping is a codebase instead of articles.
FreeFlow: Free Mac Dictation That Replaces Wispr Flow and Superwhisper
Stop typing prompts. FreeFlow listens while you talk, cleans up the transcript with AI, and pastes it wherever your cursor is. Same job as the $15-a-month dictation apps, open source, no subscription.
Install it first, read after:github.com/zachlatta/freeflow. Grab the FreeFlow.dmg from that repo and you are running in under three minutes. Apple Silicon and Intel both work.
The mouse I use:Logitech MX Master 4. Asked about it constantly, so it is up here with the download instead of buried at the bottom.
FreeFlow is a free, open source Mac dictation app built by Zach Latta and maintained by Marc Bodea. It does what Wispr Flow, Superwhisper, and Monologue do: you hold a key, you talk, and clean text lands in whatever field your cursor is sitting in. The difference is you are not paying a monthly bill for it.
I talk my prompts into Claude Code now instead of typing them. A prompt I would have typed in ninety seconds takes fifteen seconds to say, and the long ones actually get written instead of getting cut short because my hands got tired.
Install it in 3 minutes
Download the app. Pull FreeFlow.dmg from the repo and drag it into Applications.
Get a free Groq API key at groq.com and paste it into FreeFlow's settings. Groq is what runs the transcription, and the free tier covers normal daily use.
Give it accessibility permission when macOS asks. That is how it pastes into other apps.
Hold Fn and talk. Let go and the text appears. Tap Command-Fn instead if you want hands-free mode that stays on until you tap again.
What it does
Context-aware cleanup. It reads what is on screen in the app you are dictating into, so names, product terms, and file paths get spelled right instead of turned into mush.
Custom vocabulary. Feed it your client names, your repo names, your jargon. It stops mangling them.
Edit Mode. Highlight text you already wrote, hit your shortcut, and say "make this shorter" or "turn this into bullets." It rewrites the selection in place.
Your own shortcuts. Hold-to-talk and toggle are both remappable. If the toggle shortcut extends the hold shortcut, you can start holding and add the extra modifier to latch it without breaking the recording.
Any OpenAI-compatible provider. Groq by default, or point it at your own model and API URL. Ollama and LM Studio work, which means you can run the whole thing local.
No server in the middle. There is no FreeFlow backend. The only thing leaving your Mac is the API call to the provider you configured.
Why this matters if you live in Claude Code
The bottleneck in agentic coding is not Claude. It is how much context you are willing to type before you get bored and send a lazy prompt. Short prompts get short thinking.
Dictation kills that. You describe the whole problem out loud, the constraints, the files involved, what you already tried, and FreeFlow hands Claude a paragraph instead of a sentence. Three things I use it for every day:
Long prompts in the terminal. Talk the full spec into Claude Code instead of typing four lines and hoping it guesses the rest.
Voice notes into CLAUDE.md. Say the rule out loud, paste it into the file, done. Your project instructions actually get maintained.
Rewriting on the fly. Edit Mode on a paragraph beats deleting it and starting over.
Pair it with the token-saving playbook so those longer prompts do not cost you more than they are worth.
The mouse I use
People ask about my desk setup every time I record. The mouse is the Logitech MX Master 4. Ultra-fast scrolling for ripping through long files and terminal output, side buttons I map to workspace switching, and it is comfortable enough that a six-hour build session does not wreck my wrist. Not required for FreeFlow, just the thing on my desk.
OpenSEO: The Free Semrush Replacement That Plugs Straight Into Claude Code
Claude can build you a beautiful site. It cannot guess which words people actually type into Google. OpenSEO hands it the real keyword data through one MCP and seven skills, so the site gets built around words that already have traffic.
If you cannot seem to rank on Google no matter what you do, this is the piece you are missing.
Claude Code is great at building. It is not great at knowing what people search for. It has never seen your search volume, your competition, or the exact phrases that bring buyers to a site like yours. So it writes nice pages about the wrong words.
OpenSEO fixes that. It is an open-source alternative to Ahrefs and Semrush, and it connects directly to your AI agents. Claude asks it for the top keywords, gets real numbers back, and then builds the site around those words instead of guessing.
It also costs a fraction of what Semrush costs. Usage-based, no big monthly bill, and you can start free with no credit card.
Step 1: Sign up for free
Go to openseo.so and create your account. No credit card needed to start.
Take two minutes to click around the dashboard first. Keyword research, domain overview, backlinks, rank tracking, site audit, AI visibility. Everything the MCP can do, you can also see in the web app. Knowing what the buttons do makes it much easier to tell Claude what you want.
Step 2: Open the AI and MCP page, then pick your agent
Once you are logged in, find the AI and MCP section. That page gives you the exact setup command for whichever agent you use.
Pick your agent from the list. Claude Code, Claude Desktop, Cursor, Codex, whatever you run. The page hands you a command. You copy it. That is the whole idea.
If you are on the CLI like I am, you paste the command into your terminal and hit enter. That is it.
Step 3: Install the MCP
For Claude Code, the command is one line in your terminal:
claude mcp add --transport http --scope user openseo https://app.openseo.so/mcp
The --scope user part matters. It installs OpenSEO for every project on your machine instead of just the folder you happen to be standing in.
On other tools the destination is the same, only the wrapper changes:
Hit enter. That installs all seven OpenSEO skills at once.
Step 5: Make sure it is actually connected
Restart Claude Code, then type /mcp.
You should see openseo in the list with a connected status. If it is sitting there unauthenticated, select it and approve the login. It uses OAuth, so it opens your browser, you click approve, and you are done.
Quick sanity check: ask Claude to run whoami on OpenSEO. If it returns your account, the pipe is open.
What OpenSEO can actually do once it is connected
Worth knowing, because Claude will only use what you know to ask for. Everything below is a real tool it can call for you:
Find keywords: research keywords from a seed topic, pull volume, difficulty and CPC on any list, and get keyword suggestions for a whole domain.
Read the competition: pull the live Google results page for any keyword, find who really competes with you across a keyword set, see a competitor's organic footprint, and pull the full list of keywords a competitor already ranks for.
Local: local search results by location, local business lookups near a set of coordinates, and the questions people ask on Google Business profiles.
Track: save keywords into a working set, list what you saved, and read rank tracker positions over time.
Fix the site: run a full technical site audit, check its status, then pull the issues and the affected pages.
Links: backlink overview and full referring-domain profile for you or anyone else.
Truth from Google itself: Google Search Console performance data, plus URL inspection for index status, crawl, and canonicals, up to ten URLs per call.
The seven skills wrap those tools into jobs you can run by name: /seo-project-setup, /seo-coach, /keyword-research, /keyword-clustering, /competitive-landscape, /competitor-analysis, and /link-prospecting.
If you are not sure which one to run, run /seo-coach. That is literally what it is for. If you are starting something new, run /seo-project-setup first so everything after it has a home.
The order I run it in for a brand new site
This is the part most people get backwards. They build the site, then try to bolt SEO on at the end. Do it the other way around. Words first, then build.
Here is my exact order, using tools that are already in this free resource library:
1. SEED, to shape the idea. Run /seed and talk through what the site is, who it serves, and what it needs to do. You end up with a real plan instead of a vibe. Grab it here: SEED in my free resources.
2. OpenSEO, to find the words. Before a single page gets written, run /keyword-research on your topic, then /competitive-landscape to see who you are up against, then /keyword-clustering to group the winners into page-sized buckets. One cluster equals one page. Save the set so it sticks around.
3. Feed the words back into the plan. Tell Claude to update the SEED plan so every page maps to a keyword cluster, with the main keyword in the page title, the URL, and the first paragraph.
4. PAUL, to build it. Now let PAUL run the build in plan, apply, unify loops with checkpoints, so you approve each phase instead of watching one giant prompt go off the rails. Grab it here: PAUL in my free resources.
5. The anti-slop stack, so it does not look like every other AI site. Keywords get you found. Design gets you trusted. Run the four free design tools alongside PAUL: the anti-slop design stack.
6. OpenSEO again, to check the build. After it ships, run a site audit, pull the issues, and hand them straight back to Claude to fix.
SEED decides what. OpenSEO decides which words. PAUL builds it. The anti-slop stack makes it look like a real company made it.
The prompt that makes this work
The single most useful instruction to give Claude before a build:
Use the OpenSEO MCP to pull the top keywords for this niche, then interview me on the results before you build anything.
That one sentence does two things. It forces Claude to go get real data instead of inventing keywords. And it stops it from sprinting into code before you have agreed on the target.
Answer its questions. Kill the keywords that do not fit your buyer even if the volume looks great. Then let it build.
How to steal a competitor's keyword list
This is the fastest win in the whole tool, and it takes about five minutes.
Give Claude a competitor's domain and ask it to pull their ranked keywords and their domain overview.
Ask it to filter that list down to keywords with real volume and difficulty low enough that a new site can compete.
Ask it to check which of those you already rank for using your Search Console data, and flag anything sitting on page two. Those are the cheapest wins you will ever get, because you are already close.
Ask it to cluster what is left into page ideas, and save the set.
You now have a content plan built from words that are already proven to send traffic to somebody in your market.
Getting picked up by ChatGPT and AI search, not just Google
This is the part people are sleeping on. More and more buyers ask ChatGPT, Claude, Perplexity, and Google's AI answers instead of scrolling ten blue links. OpenSEO tracks AI visibility and has a prompt explorer so you can compare what different models say about your topic.
Ranking in AI answers is a slightly different game. What works:
Write the question, then answer it in the next sentence. AI engines quote clean, short answers. Bury it in paragraph six and you do not get quoted.
Use question keywords. Ask OpenSEO for the question-shaped phrases in your niche, and the questions people ask on local business profiles. Those map perfectly to how people talk to a chatbot.
Name real things. Say the actual tool names, product names, city names, and your own brand name. AI engines connect entities, not vibes.
Get indexed. Most AI search still leans on a search index underneath. Use OpenSEO's URL inspection to confirm Google actually has your pages before you wonder why no model mentions you.
Get mentioned elsewhere. Run /link-prospecting. Links and mentions on other sites are still how these systems decide you are real.
Check your work. Use the AI visibility tracking to see whether you show up when someone asks a model about your category. If you do not, you now know what to fix instead of guessing.
A simple weekly rhythm once the site is live
You do not need an SEO routine that eats your Monday. This takes about fifteen minutes.
Weekly: ask Claude to pull rank tracker positions and Search Console performance, and tell you only what moved.
Weekly: ask for anything sitting in positions 8 through 20. Improve one of those pages instead of writing a brand new one. It is far less work for far more traffic.
Monthly: run a site audit, pull the issues, and let Claude fix them in one pass.
Monthly: re-run the competitive landscape to catch new keywords your competitors just started ranking for.
That is the whole loop. Find words, build around them, check what moved, fix the closest thing to a win.
Where people actually learn to run this
The tools are the easy part. Knowing which keywords are worth a page, when to ignore a big volume number, and how to turn a cluster into a page that converts, that is the part nobody puts in a docs site.
That is what we work through inside CC Strategic AI on Skool. Real builds, real critiques, and the exact setups I run on my own sites and client work.
ADHD: The Skill That Stops Claude From Giving You The Boring Answer
Claude locks onto the first idea it thinks of and defends it for the rest of the chat. This skill splits it into 5 separate brains that never talk to each other, then brings in a critic to sort the good from the traps.
What it is
ADHD is a free, MIT-licensed Claude Code skill built by Udit Akhouri. You install it with one command and then type /adhd plus your problem.
The name is not a joke and it is not a medical tool. It is named after the way an ADHD brain actually thinks: 30 directions at once, no filter, weird connections nobody asked for. That part is a superpower. The part that hurts is the second half, sorting the 30 ideas down to the 2 worth doing.
This skill gives Claude the first half on purpose, then adds the second half as a separate step. You get the scatter AND the sort.
The problem it actually fixes
There is a name for what goes wrong when you ask Claude an open question. It is called premature convergence.
Claude writes one word at a time, and every word is shaped by the words before it. So the very first sentence it commits to becomes the frame for everything after. Ask it "how should I handle retries in my app" and it locks onto exponential backoff in the first line, then spends 800 words defending that lane. The answer is correct. It is also the answer any senior engineer gives in 30 seconds, and you did not need an AI for that.
Asking Claude to "think of more options" does not fix it, because those extra options are written in the same chat, downstream of the first idea. They are variations, not alternatives.
ADHD treats this as a plumbing problem instead of a prompting problem. If the branches can see each other, they copy each other. So it makes sure they cannot see each other.
How it works, in one picture
Two phases, with a hard wall in the middle. Phase 1 is not allowed to judge. Phase 2 is not allowed to generate.
YOUR PROBLEM
|
v
PHASE 1: DIVERGE (no critic allowed)
|
+--> Agent 1 [frame: biology] --> 6 ideas
+--> Agent 2 [frame: speedrunner] --> 6 ideas
+--> Agent 3 [frame: 3am on-call] --> 6 ideas
+--> Agent 4 [frame: inversion] --> 6 ideas
+--> Agent 5 [frame: $0, 1 hour] --> 6 ideas
(parallel + isolated, ~30 raw ideas)
|
=====WALL===== branches never see each other =====
|
v
PHASE 2: FOCUS (critic on)
|
+--> SCORE novelty / viability / fit, 0-10 each
+--> TRAP FLAG "looks great, here is why it breaks"
+--> CLUSTER group by underlying angle, not keywords
+--> DEEPEN top 3 survivors get a full sketch
|
v
BRIEF -> WIDE SET -> SHORTLIST (with a starred pick) -> 3 DEEP DIVES -> PROVOCATION
The important detail is that the generator and the critic are different Claude calls with opposite instructions. It is not one prompt promising to be creative and fair at the same time. The generator is literally forbidden to evaluate. That is why the weird ideas survive long enough to be considered.
The frames, and why they are the secret sauce
A frame is a costume Claude has to wear while it thinks. Each of the 5 parallel agents gets a different one, so they physically cannot arrive at the same answer.
The skill ships 15 of them:
hardware engineer, thinks in latency, memory, and physical limits
regulator, asks what has to be provable, traceable, or refusable
10-year-old, has never seen software, ignores every convention
competitor trying to break it, attacks the obvious solution, then flips the attacks into ideas
biology, steals a mechanism from immune systems, evolution, or cell signaling
game design, looks for loops, rewards, friction, save states, speedrun tricks
markets, turns it into buyers, sellers, auctions, and clearing houses
inversion, asks how to guarantee the opposite, then negates each answer back
extreme: $0 budget, 1 hour, the crudest thing that still works
extreme: infinite budget, 10 years, the maximalist version
remove the load-bearing assumption, deletes the thing everyone treats as fixed
speedrunner, hunts glitches, skips, and the abusive-but-legal path
ant colony, no central planner, dumb agents and local rules only
3am on-call, designs so nobody ever gets paged
For code problems it picks 4 practical frames plus 1 wild one, so you always get range without the whole run going silly. It also varies the picks between runs, which means running /adhd twice on the same question gives you two different idea pools.
Install it in one command
Open a terminal and run:
npx skills add UditAkhourii/adhd
That auto-detects your agent. It works with Claude Code, Cursor, Codex, Cline, Windsurf, Gemini CLI, Antigravity, and around 50 more. Restart Claude Code and the skill is live.
If you are on Codex and the universal command does not register it, force the target:
npx skills add UditAkhourii/adhd -a codex -g
There is also a standalone command line version and a Node library, if you want the same loop outside of Claude Code or in a batch job:
Nothing is required for the skill itself, though. No API key, no config file, no dependencies. Claude runs the loop by itself.
How to actually use it
Type the slash command with your problem in quotes:
/adhd "design a retry and timeout strategy for a CLI that sometimes hangs for 90 seconds"
Or from the CLI version, with knobs:
adhd "design a rate limiter that survives a leader election"
adhd "name this function" --frames 3 --ideas 8 --top 2
Three things worth knowing before your first run:
Typing /adhd skips the gate. The skill has a pre-flight check that normally aborts on questions with one right answer. Explicit invocation means you opted in, so it goes straight to work without arguing with you.
Certain words turn it off. If you say "quick", "standard", "canonical", "textbook", "just", or "one-line" in your prompt, the skill reads that as "give me the normal answer" and steps aside. Leave those words out when you want the wide version.
Give it the real constraints. Every branch only sees your problem plus its frame. Budget, deadline, stack, and team size belong in the prompt, or the branches will invent their own.
What you get back
The output has a fixed shape, and the shape is half the value:
Brief. One or two lines confirming the problem it heard.
Wide set. All ~30 ideas, grouped into 3 to 6 clusters by underlying angle, not by keyword. Each idea carries a score chip like [N7 V8 F9] for novelty, viability, and fit.
Converge. A 2 to 4 idea shortlist with reasons, and one idea marked with a star as the non-obvious-but-viable pick.
Traps. Listed separately, each with a one-line reason it is a trap. This is the part nobody else gives you.
Focus. The top 3 get a real sketch: how it works, the load-bearing risk, the first concrete step, and 3 to 5 spin-off ideas.
Provocation. One wildcard question, in case nothing landed.
Here is the difference in practice. On the hanging-CLI problem, a normal single answer walks through progressive timeouts and exponential backoff and lands on a sensible hybrid. ADHD surfaced 30-plus ideas and starred this one instead: a rage-quit button that glows hotter the longer you wait, and one click cancels and re-sends the prompt to a cheaper, faster model. It also flagged 20 traps. The obvious answer never questions the assumption that you should wait for the slow model. Maybe the slow model was just the wrong model for that prompt.
Does it actually work
The author ran 6 open-ended engineering problems, same model both ways, judged by a separate AI told to act like a skeptical staff engineer, with the A and B order shuffled so the judge could not tell which was which. Scores out of 10:
Breadth: 9.00 vs 4.83
Novelty: 7.83 vs 2.67
Trap detection: 9.50 vs 1.83
Actionability: 9.50 vs 6.50
Builder usefulness: 7.67 vs 6.83
ADHD won 5 of the 6 problems. The biggest gap by far is trap detection, because a normal answer almost never names the idea that looks great and quietly breaks.
Read the numbers with your eyes open. The judge is an AI, the sample is 6 problems, and it is the author's own benchmark. An independent blind test by another developer reproduced the direction of the result on 2 problems, at roughly 2.3x the time and 1.9x the output. The method also picked up a feature in The New Stack and has been adopted by other open source projects.
When to use it, and when to leave it alone
This is not a skill you leave running. It is a skill you reach for at a fork in the road.
Use it for: architecture and design decisions, API or CLI surface design, naming a product or a function, migration and refactor plans, bugs where you have no idea what the root cause is, code review when you want to know what else could go wrong, and any pricing, positioning, or strategy question.
Skip it for: lookups, syntax, bugs where you already know the cause, anything one search away, and anything inside a tight loop where you need an answer in two seconds.
The one-sentence test from the docs is good: if a junior would Google it and find the answer, the normal answer wins. If a senior would say "hm, let me think about this differently for a minute", that is the moment ADHD replaces.
The honest cost
A default run is about 10 Claude calls: 5 to diverge, 1 to score, 1 to cluster, 3 to deepen. Roughly 30 to 90 seconds of waiting.
Call count is not what you pay for, though. Every branch is a fresh, isolated context, and that isolation is the entire point. Which means each branch re-loads your whole base context, your CLAUDE.md, your project state files, your tool definitions, before it produces a single new idea.
cost = N x (base_context + branch_output) <- diverge
+ critic_context <- score + cluster
+ K x deepen_context <- focus
If your session carries a 26,000 token base, 5 branches burn about 130,000 tokens of setup before the interesting part starts. So the standalone CLI is much cheaper than the in-Claude skill, because the CLI branch only carries the problem and the frame.
Practical framing: a few cents to a few dollars to widen a decision that would cost you a week if you got it wrong. Do not run it on every keystroke. Run it at decision points.
How this helps if you have ADHD brain
Plenty of us building with Claude Code think in exactly this shape. Twelve tabs open, six half-built ideas, and a real gift for spotting the connection nobody else saw. The generating part was never the problem. The problem is that everything feels equally interesting, so nothing gets picked, and nothing ships.
What this skill quietly does is split those two jobs apart and put a wall between them.
You stop paying the tab tax. The 30-directions-at-once phase runs in parallel, in Claude's context, not in your head at 1am. You read the map instead of building it.
Something else does the converging. The critic pass scores, clusters, and picks. That is the exact executive-function step that costs the most and gets skipped the most.
The shiny trap gets named out loud. Twenty traps flagged with reasons is a guardrail against the idea that feels amazing at midnight and eats two weeks.
It ends with one concrete first step. Not 30 options. Each deepened idea comes with the literal next action, which is the difference between a great idea and a shipped one.
It refuses to sit on the fence. "Here are 20 ideas, you decide" is listed in the skill as an anti-pattern. It has to take a position and star a pick.
Read it as external executive function. Your brain does divergence for free. The skill rents you the convergence.
I Need Cruise: Find Absurdly Cheap Cruises Inside Claude Code
Cruise sites show you a price that is not the real price. This skill adds back the taxes and the tips, prices your flight, and tells you what the trip actually costs per night.
What it is
I Need Cruise is a free Claude Code skill. You tell it roughly what you want, like "somewhere warm and cheap this winter," and it comes back with real sailings and the real total.
You do not need to pick a destination. Not picking one is the whole point. The cheapest trips are the ones you would never have searched for.
Why cruise prices lie
A cruise advertised at "from $389" is not $389. There are two costs stacked on top, and most sites show you neither.
Taxes and port fees. These are in the data, but buried. On a real sailing we tested, they added $92 to a $389 fare.
Gratuities. These are the daily tips the ship charges you automatically. They run $16 to $25 per person, per day. They show up in no price you will find online. On a 7-night trip for two people, that is $252 nobody told you about.
Add it up and the real number looks like this:
Fare $389 + Taxes $92 + Gratuities $252 = $733
That is 88% more than the price on the ad. The skill does that math every time and shows you every line of it, so you can see exactly where the money went.
The flight is usually half the trip
Here is the part that surprised me most. Cruise sites do not sell flights, and flight sites do not know about ships. So nobody adds them together.
We measured a real example. A round-trip flight from Newark to Miami for two people came to $478. The cruise it paired with was $414. The flight cost more than the boat.
So the skill prices the flight too, and ranks on the two of them combined. A cheap cruise from a far-away port is not a cheap trip.
It also checks that the flight actually works. Ships sail in the early evening and do not wait. On one real search, three of the five flights offered were useless at any price because they landed after the ship left. The skill filters those out first, then picks the cheapest one that is left. If it skips a cheaper flight, it tells you why.
The trick: repositioning cruises
Twice a year, cruise lines have to move their ships. In spring they sail from the Caribbean over to Europe. In autumn they come back. Those crossings are called repositioning cruises.
Almost nobody books them, so the lines dump them at $40 to $70 a night. A two-week crossing of the Atlantic can cost less per night than a hotel room.
The reason they are cheap is simple: it is a one-way trip, so you end up in the wrong country. That is exactly why pricing the flight home matters, and it is the step the skill was built around.
What you need: an Apify account
This is the one thing you have to set up, and I want to be straight with you: it is not free. There is no free cruise data anywhere. Every official cruise API is locked behind a travel agency license. That lock is the reason no free tool does this.
Instead, the skill uses Apify, a service that runs small scraper programs for you.
Step 1: make an account.Sign up at Apify here. The free plan gives you $5 in credit every month and does not ask for a card. That is enough to run this.
Use code CHARLIE20 for a discount if you decide to upgrade later.
Step 2: connect Apify to Claude Code. This is the part I show in the video, and it is how I run it myself. You are installing something called an MCP, which is just a way to plug an outside service into Claude Code so Claude can use it directly.
Open your terminal and run this one line:
claude mcp add --transport http apify https://mcp.apify.com
Then open Claude Code, type /mcp, and log in when it asks. That is it. Once Apify is connected, the skill finds it on its own and there is no key or token for you to manage.
Not sure if it worked? Type /ineedcruise setup. It checks your setup, runs a small live test, and tells you plainly what is working and what is not.
What a search costs
Every search spends a little of your Apify credit. Here is the honest breakdown:
Searching for cruises: about $0.50 to $0.80
Pricing the flights: about $0.03 to $0.05
Total per search: roughly $0.55 to $0.85
On the free $5 plan that is about 6 to 9 searches a month.
The skill never spends your money without asking. It shows you the plan and the exact cost first, and waits for you to say go. When it finishes, it tells you what the search actually cost.
Install it in 60 seconds
Open a terminal and run one command:
npx ineedcruise install --with-commands
That puts the skill in ~/.claude/skills/ineedcruise/ and creates the /ineedcruise command. Restart Claude Code, type /ineedcruise, and you are in.
A few other commands if you need them:
npx ineedcruise install --project installs it into just the current project instead of everywhere.
npx ineedcruise update replaces an older copy with the newest version.
npx ineedcruise uninstall --with-commands removes it cleanly.
How to actually use it
Type /ineedcruise, then say what you want in plain English. Any of these work:
"Cheapest cruise out of Florida this winter" "Repositioning crossing to Europe, two weeks, under $1500" "Somewhere warm, 7 nights or more, as cheap as possible" "I want to get to Europe cheap"
It will ask which airport you fly out of. Answer that one, because flights are part of the ranking. Leave it out and the answer is not just incomplete, it is wrong.
What comes back is a short brief: the best trip with every cost broken out, two runners-up, a table of everything else that qualified, and a booking link on every single one.
What it will not do
I would rather tell you the limits up front than have you find them later.
It does not book anything. It finds the trip and hands you the link. You book it yourself.
It does not pretend to see everything. The cruise data source returns only 5 results per query, no matter what you ask. The skill works around that by running several narrow searches, and it always tells you how many it ran.
It does not guess a fee. If a cruise line is not in its tip table, it says so instead of making up a number.
It does not invent flight prices. Flights are only published about 11 months out. Past that you get a clearly labeled estimate, never a fake price.
It will not do the shady stuff. No hidden-city ticketing, no Skiplagged tricks, no fuel dumps. It will explain them if you ask. It will not build them, because airlines catch it, close accounts, and wipe out your miles.
One more honest note: the tip rates are estimates, last checked on July 25, 2026. Cruise lines change them about once a year. Every result prints that date and reminds you to confirm when you book.
ZenMillionaire: A Council of Money Advisors Inside Claude Code
Ten advisors built from twelve real wealth books. Ask a money question and they debate it in front of you, and every answer names the book it came from.
What it is
ZenMillionaire is a free Claude Code skill that counsels you on money like you have a board of billionaire advisors sitting at your table.
It runs on something called persona counseling. Instead of one AI voice giving you one flat answer, the skill splits into separate advisors. Each advisor is built from one real book, and each one is only allowed to argue using ideas from that book. Then they talk to each other about your actual situation.
Two rules make it different from asking a chatbot about money:
Every claim gets a citation. If an advisor tells you something, it names the source, like (Rich Dad Poor Dad, p.42). If an idea is not in the library, the advisor has to say "that's not in the library, here's my own read." No made-up financial doctrine dressed up as fact.
It is built for augmentation, not automation. It does not move your money or run your budget. It helps you make a better decision, then hands the decision back to you.
The 10 advisors and the 12 books behind them
Each advisor pulls from its own source, so they naturally disagree. That disagreement is the whole point.
The Zen Master — Ken Honda, Happy Money
The Definite Thinker — Napoleon Hill, Think and Grow Rich
The Realist — Morgan Housel, The Psychology of Money
The Life-Energy Guide — Robin & Dominguez, Your Money or Your Life
The Ancient Sage — George Clason, The Richest Man in Babylon
Rich Dad — Robert Kiyosaki, Rich Dad Poor Dad
The Quiet Millionaire — Stanley & Danko, The Millionaire Next Door
The Operator — Alex Hormozi, $100M Offers and $100M Leads
The Angel Philosopher — Naval Ravikant, The Almanack of Naval Ravikant
The Persuader — Robert Cialdini, Influence
MJ DeMarco's The Millionaire Fastlane is in the library too, so the blended advisor can pull from it.
The advisors sit on three poles: mindset-zen (peace, meaning, "how much is enough"), mechanics-iq (assets, saving, investing), and modern-leverage (offers, selling, earning more). The skill picks advisors from different poles on purpose so you hear the other side.
The six modes
You type /zenmillionaire, it shows the menu, asks you a few quick questions, then suggests the right mode. It always confirms before it locks anything in.
1. Talk to one advisor — pick a single voice and have a normal back-and-forth in that style.
2. Council / roundtable — the best one. Claude spins up subagents, one per advisor, and they debate your question. You get each opinion plus a synthesis at the end.
3. Blended advisor — one voice that pulls from all twelve sources at once.
4. Diagnose me — a money-blueprint survey that saves a profile, so later advice is personalized to how you actually think about money.
5. Create a persona — build your own advisor.
6. Add a source — add a book PDF or a YouTube video to the library.
Install it in 60 seconds
Open a terminal and run one command:
npx zenmillionaire install --with-commands
That drops the skill into ~/.claude/skills/zenmillionaire/ and writes the /zenmillionaire slash command. Restart Claude Code, type /zenmillionaire, and you're in.
Other commands if you need them:
npx zenmillionaire install --project installs into the current project's ./.claude/ instead of globally.
npx zenmillionaire update overwrites an existing install with the newest version.
npx zenmillionaire uninstall --with-commands removes it and the slash command.
Three ways I actually use it
1. Pressure-test a decision. Type /zenmillionaire, pick council, and ask something real:
"I'm thinking about dropping a $4,000/mo retainer client to free up time for my own product. Challenge me on this."
Rich Dad argues about building an asset instead of trading hours. The Realist argues about how long your runway really is. The Operator argues about whether the product offer is even good enough yet. You end up seeing the parts of the decision you were skipping.
2. Have the council review a document. Paste in notes from a coaching call, a business plan, or a budget, and say:
"Have the council review this consulting conversation and tell me where the advice is weak."
The subagents each read the document through their own lens, then come back with feedback from ten angles instead of one.
3. Work on a money belief, not a money number. Run mode 4, "Diagnose me," and let it survey your money blueprint. It stores a profile, so the next time you ask whether to spend or save, the advice already knows how you're wired.
Make it your own
Nothing is locked in code. Every advisor, framework, and citation is plain Markdown you can open and edit.
Want a new advisor? Run mode 5, hand it a book you already own, and ask it to build a persona from that book. Want to grow the library? Run mode 6 and feed it a PDF or a YouTube link. It distills the ideas, tags them with page numbers or timestamps, and adds them to the citation index so every advisor can quote it.
One honest note on copyright: the public repo ships the distilled, cited frameworks only. The actual book PDFs are not included, because owning a book does not give anyone the right to redistribute it. You bring your own copies.
Understand Anything: Turn Any Codebase Into a Brain You Can Explore
Most of us build apps with Claude Code and have no real idea how they're wired. This plugin maps every file, function, and dependency into an interactive knowledge graph, then hands you a guided tour of your own project.
What it does
Understand Anything (Egonex-AI/Understand-Anything, the #1 repo of the day with 200,000+ users) is a Claude Code plugin that turns a codebase into an interactive knowledge base anyone can actually understand. One command runs a multi-agent pipeline over your project:
/understand scans the whole project, extracts every file, function, class, and dependency, and builds a knowledge graph saved to .understand-anything/knowledge-graph.json.
/understand-dashboard opens a local web dashboard where the codebase is drawn as a clickable graph, color-coded by architectural layer (API, service, data, UI, utility). Click any node to see its code, its relationships, and a plain-English explanation of what it's for.
It auto-generates guided tours ordered by dependency, so you learn the codebase in the right order instead of grepping blind.
It runs incrementally after the first pass, only re-analyzing files that changed, so day-two runs are cheap.
Why it's needed
Here's the uncomfortable truth for the Claude Code era: most of us are shipping apps we didn't hand-write and can't fully explain. The AI built it, it works, and the mental model lives nowhere. The day something breaks, or a client asks how it works, or you want to hand it off, you're reading your own project like a stranger.
Understand Anything fixes the exact gap. It's the difference between "the AI made this" and "I understand what the AI made." It splits the work the smart way: Tree-sitter deterministically parses the real structure (imports, call sites, inheritance) so the edges are reproducible, and an LLM layer reads that structure to write the summaries, layer assignments, and tours a parser never could. Same code always yields the same graph, but you still get intent, not just syntax.
Legit use cases
Where this actually earns its keep, especially if you run an agency or ship AI-built software:
Understand your own vibe-coded app. You built it with Claude Code across ten sessions. Now graph it and finally see how the pieces fit before you scale it, sell it, or debug it under pressure.
Inherit a client's or legacy codebase. Walk into 200,000 lines you've never seen, run /understand, and get an onboarding tour instead of a week of blind exploration. Use /understand-onboard to generate a new-hire guide.
Scope a client project before you quote it. Graph the existing system first so your proposal is based on the real architecture, not a guess. Fewer surprise hours, tighter estimates.
Review changes before you commit./understand-diff shows which parts of the system your current changes ripple into, so you catch blast radius before it ships.
Answer "how does X work?" instantly./understand-chat How does the payment flow work? queries the graph in plain English. /understand-explain src/auth/login.ts deep-dives a single file.
Onboard a team or VA without a meeting. The graph is just JSON. Commit it once and teammates skip the whole pipeline, they open the dashboard and explore. Docs-as-code that never goes stale (wire up /understand --auto-update for a post-commit refresh).
Map a business's logic, not just its files./understand-domain extracts domains, flows, and process steps, so a non-engineer stakeholder can see how the software maps to the actual business.
How to install
Two commands, straight into Claude Code. First add the marketplace, then install the plugin:
Restart Claude Code, then point it at a project and run /understand. When it finishes, run /understand-dashboard to open the visual graph. Scope a huge monorepo with /understand src/frontend, or generate non-English output with /understand --language zh (supports en, zh, zh-TW, ja, ko, ru).
Honest framing: the first run costs tokens
The initial /understand reads your entire codebase, so on a large project it can burn a meaningful number of tokens. Run it on a subscription / token plan, or point it at a local model (Ollama) for the first pass if you're cost-sensitive. Every run after that is incremental by default, only changed files get re-analyzed, so ongoing use is cheap. Worth knowing before you aim it at a 200k-line repo.
How it works under the hood
The /understand command orchestrates five specialized agents: a project-scanner (detects files, languages, frameworks), a file-analyzer (extracts functions, classes, imports into graph nodes and edges, running up to 5 in parallel), an architecture-analyzer (assigns each file to a layer), a tour-builder (writes the guided walkthroughs), and a graph-reviewer (validates completeness). /understand-domain adds a sixth for business-logic mapping. It's MIT-licensed and open source, an Egonex project originally created by Lum1104.
RE Walkthrough Pro: Zillow Listings → Cinematic Video Walkthroughs
Stop paying a videographer $1,500 a listing. This skill scrapes a Zillow listing, animates every room into a cinematic camera move, and stitches one finished walkthrough you can sell to the agent.
What it does
RE Walkthrough Pro is a Claude Code skill that turns a Zillow listing into a cinematic, room-by-room walkthrough video. One command runs the whole pipeline:
/re-walkthrough-pro takes a Zillow listing link (or finds you listings: "find 5 single-family homes in Montclair NJ").
It pulls every photo plus the address, price, beds/baths, and agent through an Apify Zillow actor (Zillow blocks naive scraping, so the actor is the reliable source).
It curates the hero rooms, then animates each one with the Higgsfield MCP, a camera move matched to the room: drone approach on the exterior, steadicam down the hall, orbit in the great room, a rising reveal on the backyard.
It stitches the clips with ffmpeg into a 16:9 master for the agent, plus an optional 9:16 cut for social.
Why it's needed
Agents pay for media, and the bottleneck is production time. A walkthrough that looks like a $1,000 to $3,000 shoot can be built from the listing photos in minutes, for sub-cents of scraping plus a handful of Higgsfield credits. Drop it into a real estate marketing service as a video line item, or batch tours across a farm area and pitch the listing agents.
The architecture is the actual unlock. AI video models work clip-by-clip (around 5 seconds of image-to-video), so a whole-house walkthrough has to be per-room clips stitched together on your side. This skill does exactly that, with a camera move per room type so the stitched result reads like walking through the home.
Honest framing: it's cinematic, not 3D
This produces a cinematic walkthrough, per-room camera-move clips stitched into a tour, not a true 3D or Matterport reconstruction. It looks premium and it is impossible without an AI video engine, but do not promise a walkable dollhouse. The skill is honest about this on purpose, so you sell it right.
How to install
One command with npm:
npx re-walkthrough-pro install
That drops the skill into ~/.claude/skills/re-walkthrough-pro/. Restart Claude Code, run /re-walkthrough-pro, and paste a Zillow link. Use --project to install into the current repo only.
What you need
Two MCP servers connected in Claude Code, plus ffmpeg:
Higgsfield MCP for the per-room animation and the social reframe (Seedance 2.0 / Kling 3.0 image-to-video).
Apify for the listing scrape (the Zillow detail scraper pulls photos plus specs in one call).
SkillSmith: The Plugin That Builds Your Claude Skills (and Agents)
An agent is just a skill: a named worker with its own tools and routing, packaged in plain Markdown. SkillSmith is the one plugin that builds them in minutes instead of wrestling with n8n or custom code.
Skills are the new agents
If you're still struggling to build an AI agent in 2026, here's the unlock: an agent is just a skill. Anthropic has publicly told builders to focus on skills over agents, and a Claude skill is nothing more than a named worker with access to its own tools and routing. It knows what it can touch and when to load what. It's packaged in plain Markdown, which is just a text file you can read yourself, not custom code.
The old way of building agents meant n8n workflows or custom scripts, both far harder to assemble than a skill. I've built 40+ skills now, and this is 10x easier than the old way.
What a skill actually is (the folder anatomy)
A skill is a folder with files inside. Open one of mine and you'll find the same five parts every time:
The entry point (skill.md). The beginning of everything: the triggers, the persona, and the commands. It lives outside the subfolders.
Context. Everything the skill needs to know to operate, like the details about your business it should load.
Frameworks. The playbooks and methods the skill follows.
Tasks. The workflows that tell Claude how to put the frameworks, context, and entry point together into one cohesive run.
Templates. How the output should look when the skill is done running.
SkillSmith formalizes this into seven file types, each with a syntax spec: entry points, tasks, templates, frameworks, context, checklists, and rules. Consistent structure is what makes a skill portable, so someone else can install it and not get lost.
One command, four workflows
There's a single command, /skillsmith, with four routes:
/skillsmith discover: a guided interview that turns an idea into a skill spec.
/skillsmith scaffold: builds the compliant skill directory from that spec.
/skillsmith distill: takes a long-form YouTube video or an SOP and distills it into framework chunks you can turn into a skill.
/skillsmith audit: reviews skills and frameworks you already have and improves them over time.
I built a skill live with it
I used SkillSmith to build a "Creative Director" skill, a simpler version of my Viral Audit. The prompt: make a slash command that pulls my YouTube and Instagram data from the last 30 days, acts as a creative director, ranks my videos best to worst in a Markdown file, then interviews me on what to do with the top performers and closes with a coaching brief.
SkillSmith generated the whole folder: a channels.md context for which handles to log into, frameworks for the performance signals to rank on, a task for the ranking workflow, and a template for how the report should look. Then it ran: scraped both platforms, ranked my reels, carousels, top 10 shorts, and top 10 long-form pieces, and wrote a full coaching brief telling me what to do better. Building it felt easy once the structure was right.
The data layer: Apify (optional)
To pull the social data, that skill used Apify, a marketplace of "actors" that take actions on your behalf (in this case an Instagram Post Scraper and a YouTube scraper). Apify recently shipped its own MCP, a bridge that lets Claude Code walk into the marketplace and pick the right actor for the workflow you described.
Important: you do not need API connectors or an MCP to start. Your skills can be simple at the core and you can layer bridges to other software on as you grow. Apify was just the data layer for this particular example.
How to install
One command in a fresh terminal:
npx @chrisai/skillsmith
That installs the skill to ~/.claude/commands/skillsmith/, available in every workspace. Restart Claude Code, type /skillsmith, and pick a workflow. Use --local to install to the current project only.
Coming next: making skills run without you
Right now these skills run on a manual slash command, which is totally fine to start. Part two covers making them autonomous, on a schedule or triggered from Telegram or Slack. Three paths I like: Hermes (self-hosted on a VPS, my pick over OpenClaude), Claude managed agents (the easiest), and wiring an n8n MCP to run the workflow on its own.
SEED: Turn a Raw Idea Into a PAUL-Ready Build Plan
You have the idea. You skip the planning because it is tedious, then build the wrong thing. SEED interviews you by project type, writes the plan, and hands it straight to PAUL to build.
What it does
SEED (ChristopherKahler/seed, 270+ stars, by Chris AI Systems) is an AI project incubator for Claude Code. It takes a raw idea through a type-aware guided conversation and produces a structured PLANNING.md, ready for a PAUL-managed build. The point: stop answering the same setup questions every time you start something.
/seed begins the guided ideation, shaped by what kind of thing you are building.
/seed graduate moves a finished plan into apps/ with its own git repo.
/seed launch graduates AND initializes PAUL in one command, so you go straight to building.
/seed status shows your ideation pipeline; /seed add-type creates a custom project type.
Typed, so it asks the right questions
SEED does not run one generic interview. It adapts the rigor and the sections to the project type, so a throwaway script does not get grilled like a full app:
Application (deep rigor, 10 sections): software with UI, data model, and API.
Workflow (standard, 8 sections): Claude Code commands and skills.
Client (standard, 7 sections): client websites.
Utility (tight, 6 sections): small tools, built to resist scope creep.
Campaign (creative, 7 sections): content and marketing initiatives.
Rigor levels change the behavior: tight moves fast and resists expansion, standard stays balanced, deep goes thorough, creative goes generative. You can also write your own type with /seed add-type.
Why it's needed
Most builds go sideways at the start, not the middle. You skip planning because it feels slow, then you are three hours into the wrong architecture. SEED makes planning the fast part. The interview is collaborative, it checkpoints so you can resume a half-finished idea, and it gates on quality so the plan is actually PAUL-ready before it graduates.
It is the front of the same build system the rest of my stack uses: SEED shapes the idea, PAUL runs the Plan to Apply to Unify loop on it, and SkillSmith builds any skills the plan calls for. SEED works standalone too, no PAUL required.
How to install
One global install with npm:
npm i -g @chrisai/seed
Zero runtime dependencies, pure Markdown skill install to ~/.claude/commands/seed/. Needs Claude Code and Node 16+. Restart Claude Code, type /seed, and start talking through the idea.
Resources
SEED on GitHub (the incubator, by Chris AI Systems, MIT)
Stop paying creators per video. This skill interviews you, casts a consistent AI creator, and ships a finished UGC video ad: keyframes, Seedance 2.0 clips, and an ffmpeg-stitched final cut.
What it does
UGC Factory is a Claude Code skill that turns a product, offer, and topic into a finished UGC video ad. One command runs the whole pipeline:
/ugc-factory interviews you on product, offer, topic, the on-screen creator, length, and aspect ratio.
It casts the creator as a multi-angle Higgsfield Element so the same person turns and moves believably across every clip.
It routes to one of 15 built-in Seedance genre skills (ecommerce, food, fashion, SaaS, real estate, and more) so each clip gets genre-correct prompt craft.
It writes a full script for every Seedance 2.0 clip (never blank, which is what makes AI video come out as gibberish), then stitches the clips into one ad with ffmpeg.
Why it's needed
Two things break almost every AI UGC ad: the creator's face changes every cut, and the video model hallucinates because the prompt was left thin. UGC Factory fixes both. The creator and product are built once as reusable Elements and referenced in every clip, so there is zero drift between cuts. And the skill always writes a full scene script: who is on screen, what they say, the camera, the motion. Nothing is ever left blank.
It is stateless on purpose. Every run starts blank and casts a brand-new creator. There is no saved avatar locking you into the same face forever.
How to use it
One command. Install with npm:
npx ugc-factory install
That drops the skill into ~/.claude/skills/ugc-factory/. Restart Claude Code, run /ugc-factory, and answer the interview. It returns the finished ad plus every source asset in one dated folder.
You need the Higgsfield MCP for the image, Elements, and Seedance 2.0 generation, plus ffmpeg for stitching. Grab the MCP here: install the Higgsfield MCP.
What's inside
The package is self-contained. It bundles all 15 Seedance 2.0 genre style skills so the genre router works out of the box, with no separate installs: cinematic, 3D CGI, cartoon, comic-to-video, fight scenes, motion-design ad, ecommerce ad, anime, product 360, music video, social hook, brand story, fashion lookbook, food and beverage, and real estate.
Stop writing SOPs by hand. This skill interviews you like a COO, pushes back on weak steps, flags what to automate, and writes the full SOP in minutes.
What it does
Procedure Ops is a Claude Code skill that interviews you, then writes the full SOP. Five tasks ship in the box:
/sop-build interview walks you through a guided Q&A and writes a clean SOP at the end.
/sop-build scaffold-business sets up a new business folder with every department.
/sop-build from-recording turns a Loom or voice note into a structured SOP.
/sop-build audit reviews an SOP you already have and flags gaps plus automation candidates.
/sop-build delegate-pack bundles a finished SOP for a VA or new hire (access list, training videos, first-week check-ins, escalation contact).
Why it's needed
I've written hundreds of SOPs for my clients and my own business. Thinking through every single step is a pain. Writing it up and packaging it pretty for your team eats hours. Procedure Ops fixes both.
It acts like a COO sitting next to you. It pushes back when a step is too vague ("review the form" gets challenged to "open the form, check fields X, Y, Z, confirm they match the contract"). It flags steps that could be automated and picks the right tool: Zapier, Make, n8n, or a Claude managed agent. It only recommends installing SEED and PAUL if you don't already have them.
How to use it
One command. Install with npm:
npx procedure-ops install
That drops the skill into ~/.claude/skills/sop-build/. Restart Claude Code and you're set. Now any time you need an SOP, type /sop-build interview and start talking.
My order of operations: scaffold the business folders first with /sop-build scaffold-business, then run /sop-build interview per process. The skill writes every SOP to a clean SOPs/<business>/<department>/<slug>.md path. Your team reads it in Markdown, or you pick Google Drive or Notion as the host when you scaffold the business.
What you can feed it
Voice notes. Loom videos. Old SOP drafts. The skill transcribes the recording, pulls out the steps, flags the gotchas, and writes the SOP for you. Same quality gate as a live interview, faster intake.
Compliance Ops: Build HIPAA-Safe AI in Claude Code
If you build with Claude Code in medical, legal, or finance, one leaked record gets you sued or fined. This skill makes Claude aware of the rules BEFORE it builds, so protected data never routes through an uncovered vendor.
What it does
Compliance Ops is an interview-driven compliance guardrail for Claude Code. It loads the matching framework (HIPAA, SOC 2, GDPR, PCI-DSS) into your conversation before Claude builds or automates anything, so every website, workflow, and agent is designed correct by default. Three commands ship in the box:
/compliance-ops finds out what you're building and what you're worried about, loads the right framework, and flags you the moment a build heads in a non-compliant direction.
/compliance-ops-audit reviews an existing build or data flow for gaps, anywhere protected data touches an uncovered vendor or the AI.
/compliance-ops-document generates a data-flow map plus a vendor BAA checklist you can hand straight to a compliance officer.
Why it's needed
Businesses in medical, legal, and financial spaces handle PHI, protected health information, the kind of data that gets you sued or fined if it leaks. And right now most owners don't know how to bring AI agents into the room safely. So they freeze. Stuck in digital limbo while their competitors automate and pull ahead.
The fix isn't bolted on after you build. It's baked in from line one. Compliance Ops gives the AI awareness of the rules up front, so protected data never gets routed through a non-compliant flow in the first place.
This is not legal advice
Important, and the skill says this itself: Compliance Ops is a seatbelt and a paper trail, not a compliance certificate. It does not make a Claude consumer subscription able to process protected data, and it does not replace your lawyer or a signed BAA. What it does is keep your builds inside the lines and document the entire process so you have a record of the decisions you made.
How to install
One command in a fresh terminal:
npx compliance-ops install --with-commands
That drops the skill into ~/.claude/skills/compliance-ops/ and writes the three slash-command stubs. Restart Claude Code, type /compliance-ops, and feed it what you're building and what you're worried about. Use --project to install into the current project's ./.claude/ instead of globally.
Advertising Ops: CMO-in-a-Box (Media Buyer) for Claude Code
Stop guessing at a blank ad account. This skill scrapes the ads that have been running long enough to be proven winners, tears down the video frame by frame, then acts like your CMO and generates new copy plus creative ready to launch.
What it does
/advertising-ops runs the whole loop a media buyer runs by hand, one command, five phases:
Scope. Reads your brand kit if you have one, otherwise interviews you for your business, keywords, and ICP.
Scrape. Pulls long-running winning ads from the Meta Ad Library through Apify, filtered to a minimum run time you set (2 months floor) AND still active. Started long ago plus still live equals the market already validated it.
Teardown. For video ads it downloads the MP4, extracts frames with ffmpeg, transcribes the voiceover, and breaks down the hook, structure, and CTA. An LLM cannot watch a video, so it samples it instead.
Brief. Acts as your CMO and pins down one CTA and the exact outcome you sell before generating anything.
Generate. Produces 5+ aligned copy plus image/video variations through Higgsfield, each dropped into its own creative folder.
Why it's needed
Most ad creative fails because it was made in a vacuum. No swipe file, no proof, no teardown of why the competitor's ad has been running for six months straight. You launch, you burn budget, you learn nothing.
Advertising Ops templates what already works. The core trick is the filter: it combines an old start date with an active status, so you only ever look at ads the advertiser is still paying to run. A winner, not a coin flip. Then it tags every ad back to the page that ran it, so your swipe file is fully sourced.
How it watches a video ad
The Apify run returns URLs and metadata, never the actual files. So the skill downloads what it needs to see it: image ads get pulled with curl and read directly, video ads get downloaded then sliced by ffmpeg into scene-change frames plus a dense first-2-seconds hook burst, and the audio is transcribed for the spoken script. Then Claude reads the frames and the transcript and writes the teardown.
How to install
One command with npm:
npx advertising-ops install
That drops the skill into ~/.claude/skills/advertising-ops/. Restart Claude Code, then run /advertising-ops. Tell it where to put the report and the creatives once and it saves those paths, so every future run reuses them without asking.
What you need
Two MCP servers connected in Claude Code, plus ffmpeg only if you audit video ads:
Apify MCP for the ad scrape (the brilliant_gum Facebook Ads Library actor, the only one that reports each ad's run time).
Higgsfield MCP for the image and video generation.
ffmpeg for video teardowns (brew install ffmpeg). Image-only runs never touch it.
Cold Outbound Skills: 100s of Leads from Claude Code
Stop buying a lead list and praying. This 29-skill plugin scrapes your own leads from Google Maps and Prospeo, writes the cold email, fires it through Smartlead, then scores the replies and tells you what to fix.
What it does
Cold Outbound Skills (growthenginenowoslawski/coldoutboundskills, 400+ stars) is 29 interconnected Claude Code skills that run the entire cold email engine, from zero to a campaign hitting 100s of leads. You start with one command and it chains the whole thing:
/cold-email-kickoff is the guided orchestrator: ICP, lead magnet, strategy, and plan in one flow.
/cold-email-starter-kit is the 14-step tutorial version if you want to learn the moves as you build.
From there the skills cover five tracks: strategy, infrastructure, list building, copy and send, and iteration.
List building. This is the 100s-of-leads engine: Prospeo exports, Google Maps scraping, domain-to-contact discovery via Blitz, competitor research, list-quality scoring.
Most cold outreach dies at the list. You buy a stale CSV, blast it, torch your domain, learn nothing. This flips it. The skills BUILD the list from live sources (Google Maps and Prospeo), score it for quality before you send, then run deliverability and spam checks so your inbox survives. It is the part of lead gen nobody packages, done in plain Markdown skills you can read and edit yourself.
It pairs with the rest of the leads stack: this finds and emails the leads, Advertising Ops builds your paid-ad creative from proven winners, and UGC Factory ships the AI-creator video. Three layers, zero overlap.
What it costs to run
The skills are free and open source. The tools they drive are not: budget roughly $360 for your first 2,000-lead campaign and about $130/month recurring across domains, Zapmail, Prospeo, Smartlead, and verification. You are paying the sending stack, not the brain.
How to install
Clone the repo, drop the skills into your ~/.claude/ commands directory per the README, and add your API keys (Prospeo, Smartlead, your domain provider). Restart Claude Code, then run /cold-email-kickoff and answer the interview. It scaffolds the strategy, sources the list, writes the copy, and queues the send.
Carousel Builder: Instagram Carousels End-to-End in Claude Code
Stop bouncing between an image tool, a doc, and Canva for every carousel. One command generates the images, locks the copy to your template, builds the deck, and writes the caption.
What it does
/carousel runs an interview-first, approval-gated pipeline that turns one topic into a finished, on-brand Instagram carousel:
Loads your framework. Reads a reference deck (shipped in the repo) so it mirrors your exact layout: two-tone headlines, one orange highlight word, glow images, CAPS-word body lines.
Interviews the idea. Breaks your topic into 4 to 6 pillars, each doing one job, plus a CTA keyword and deliverable.
Generates a cover (Higgsfield, 4:5). 3 to 4 options, no baked-in text. The one you pick becomes the theme anchor for every other image.
Drafts the copy with live character counts. Every line is budget-checked against the real Canva text boxes BEFORE a single image credit is spent. You approve the draft first.
Generates theme-locked pillar images (Higgsfield, 16:9). Each middle image is passed the cover as a reference so the whole deck shares one palette, lighting, and mood. No drift.
Assembles the deck in Canva. Duplicates your template, replaces text span-by-span (preserving the orange highlight), and swaps images with update_fill so the frame glow survives. Nothing commits until you approve the preview.
Writes the caption. Automatically runs the bundled /short-form-caption so the post copy and the on-slide CTA keyword match.
Why it's needed
A carousel is three tools in a trench coat: an image generator, a copy doc, and Canva. Most people tab between them for an hour per post and the deck still drifts off-brand. This command runs the whole thing from your terminal and keeps it on-template by design.
The trick is the character budgets. Canva text boxes are sized for specific lengths. Go over and the text reflows, the slide stops matching the deck, and the carousel loses its visual unity. The command enforces the budgets on the draft, so over-length copy never reaches Canva and never burns image credits on a deck that won't line up.
The bundled caption skill
/short-form-caption ships in the same package and runs as the final step of every carousel. It also works standalone on any video transcript or topic blurb. Output is a tight, copy-paste-ready IG / TikTok / Reels caption: CTA on line one, fragment rhythm, exactly 5 hashtags, 500-character hard cap, no em dashes, written in your voice from a fingerprint file you fill in once.
What you need
Two MCP servers connected in Claude Code, plus an 8-page Canva template you own (the repo ships a reference deck to recreate):
Higgsfield MCP for the cover and pillar images (uses the Nano Banana Pro model).
Canva MCP to duplicate your template and edit text plus image fills in place.
Or clone the repo and copy the two command files into ~/.claude/commands/. Then swap the placeholders (your Canva template ID, your handle, your brand voice file) per the README and you're live. The exact 8-slide reference deck ships inside the repo, so you can recreate the Canva template 1:1.
Not sure what to fix first?
Most people don't need another AI tool. They need to know which thing is actually costing them the most.
The free AI Bottleneck Finder tells you the one constraint capping your revenue right now, plus two builds you could ship this week. Five minutes, free, no call required.
Inheriting a codebase and grepping blind? Graphify maps every file, concept, and connection so you see the structure before you touch the code.
What it does
Reads your files (code, PDFs, markdown, images, audio, video) and builds an interactive knowledge graph of concepts and relationships. Exports as HTML, JSON, and an audit report. Every relationship is tagged extracted, inferred, or ambiguous, so you always know what was found versus guessed.
Why it's needed
Inheriting a codebase or a research corpus is painful. Grep only finds what you already know to look for. Graphify gives you the structure behind the files: which nodes are load-bearing, which connections are surprising, which questions to ask first.
When to use it
Walking into a new codebase. Processing a research folder full of PDFs and notes. Mapping a project's architecture before touching it. Understanding why past architectural decisions were made.
Graphify gave Claude unlimited memory for your data. BASE does the same thing for your rules, so they load only when they matter instead of on every single prompt.
Download my star commands
All 36 star commands from my live BASE setup, in one markdown file. Hand it to Claude Code and it installs every one you are missing.
Drop it at the root of your project, open Claude Code, and say:
Read base-star-commands.md. Run base commands list first to see what I already have, back up my commands.toml, then import everything I am missing and tell me what was added and what was skipped.
Fifteen of them are workflow modes like *dev and *plan. The other twenty-one are the mode pack, and those are the half most people never set up. Install BASE first, below.
Built by Chris Kahler. Install it first, then work through the steps below.
What it does
BASE turns your rules into a graph instead of a document. You group rules into domains like DEV, CONTENT, or BRAND. Each domain has trigger words and folder paths attached to it.
When you type a prompt, BASE checks it. If you say the word "carousel", only your carousel rules load. If you open a file inside your website folder, only your website rules load. Everything else stays quiet.
BASE vs Graphify: what is actually different
They solve two different halves of the same problem, and you can run both at once.
Graphify graphs what you know. Your code, your PDFs, your notes. It answers "what is in this project and how does it connect?"
BASE graphs how you work. Your rules, your standards, your projects. It answers "how should Claude behave right now?"
Graphify is something you query. You ask it a question and it gives you an answer.
BASE is something that fires on its own. A hook reads your prompt before Claude sees it and injects the rules that match.
Graphify does not stop your context from filling up. BASE does, because unmatched rules never get sent.
Short version: Graphify is memory for your data. BASE is memory for your behavior.
How it works with your CLAUDE.md
It does not replace CLAUDE.md. It takes the load off it.
CLAUDE.md is always on. Every line in it gets sent with every prompt you write, even when the prompt has nothing to do with those lines. Coding rules ride along while you write an email. That is wasted context, every time.
BASE domains are triggered. They cost nothing until they match. So the split is simple:
Keep in CLAUDE.md: where things live in your repo, and anything you need before you know what you are doing.
Move into BASE: every behavior rule. "Always do X." "Never do Y."
Doing this on my own setup cut my always-on context from 381 lines to 217, and nothing was lost. The rules still arrive, they just arrive only when they apply.
Starter prompts
Paste these into Claude Code after you install BASE.
Show me every BASE domain I have, with its trigger words and how many rules each one holds.
Read my CLAUDE.md. Tell me which sections are behavior rules that belong in a BASE domain, and which sections have to stay always-on.
Check every rule in my BASE domains. Does the folder, file, or tool it points at still exist?
Find rules that say the same thing twice across my domains, and tell me which copy to keep.
My rules are not firing when I expect them to. Show me the trigger words for that domain and what is missing.
Show me every star command I have with base commands list, then tell me which ones I have never actually typed.
I keep asking you to be less wordy. Turn that into a star command so I can type it instead of explaining it every time.
Star commands: the half of BASE nobody uses
Domains fire on their own. Star commands are the opposite. You type one and it changes how Claude answers for the rest of the session.
A star command is a named stance. *blunt strips the warm-up and leads with the answer. *audit assumes something is broken and goes looking for it. *hold makes Claude defend its position instead of folding the second you push back. You stop explaining how you want the answer and start naming it.
See what you already have:
base commands list shows every star command and its rule count.
base commands show audit prints one command's full rules.
base commands add --name X --description "..." --rule "..." creates one by hand.
A fresh install ships with 15, and most of them are workflow modes like *dev and *plan. The pack below adds 21 more that change how Claude thinks, not what it works on.
The mode pack: 21 commands
Behavior modes. Each one sets a single stance and holds it until you change it.
*blunt Answer first. No warm-up, no hedging, no "let me know if you need anything" closer.
*analytical Tables and numbered findings instead of prose. Facts separated from inferences.
*steelman Builds the strongest version of an argument, then the strongest counter. No thumb on the scale.
*audit Assumes something is wrong. Finds problems before it acknowledges what works, tagged by severity.
*mentor Teaches the concept before applying it. Bridges from what you already know and checks comprehension.
*editor Rewrites and tightens existing text. Adds nothing new. The draft is the scope.
*plainspeak Strips jargon and always lands the "so what". What this produces and what changes for you.
*hold Defends its position until real evidence moves it. Repetition and confident tone do not count.
Combo lenses. These stack two or three behavior modes into one combined stance instead of running them in sequence.
*vet audit plus operator. What breaks and what it costs, before you commit.
*counsel steelman plus operator plus blunt. Strongest case and sharpest counter, in business terms, delivered straight.
*weigh analytical plus operator. A trade-off table with cost, risk, and time columns.
*teach mentor plus plainspeak. Builds the mental model with no jargon.
*bottomline plainspeak plus operator. The plain-language "so what" for the business.
*dissect analytical plus audit. Forensic structured review, severity-tagged, no business framing.
*strategy discuss plus steelman plus operator. Open exploration, strongest cases both ways, no premature action.
*retro meta plus audit. What is broken and what is worth systematizing. Needs a meta command that does not ship with BASE.
Procedural commands. These run a flow instead of setting a stance.
*ping Fires a one-line message to another live Claude session by name. No doc, no ceremony.
*docs Updates your user-facing docs to match what the session actually changed. Skips docs that are already correct.
*end One-shot end of session. Sweeps for unregistered build work, syncs the graph, writes the handoff.
*dispatch Writes a build spec then boots a crew to build it. Needs a squad command that does not ship with BASE.
Write your own: the format
Star commands live in ~/.base-gbl/commands.toml. Three fields: a name, a description, and a list of rules. Here is a full one.
[[command]]
name = "blunt"
description = "No hedging, no warm-up, answer-first, terse"
rules = [
"BLUNT MODE: Answer first, explain only if asked. No hedging, no warm-up, no qualifiers. One sentence beats a paragraph, because Claude defaults to warm thorough responses that bury the answer in context.",
"Do not soften bad news. State problems, gaps, and failures directly, because hedged delivery wastes time parsing what you actually mean.",
"Skip all preamble, trailing summaries, and let-me-know-if-you-need-anything closers. They are filler that adds zero information.",
]
The thing that makes a rule work is the because clause. "Be concise" is a preference Claude will drift off within three turns. "Be concise, because you default to warm thorough responses that bury the answer" tells it what failure it is correcting, and that survives.
A combo is the same shape, with rules that name its sources so Claude pulls them at runtime.
[[command]]
name = "vet"
description = "What breaks and what it costs (audit + operator)"
rules = [
"VET MODE, a stacked combo lens. Pull both sources with base commands show audit and base commands show operator, then apply their rules together as one combined stance.",
"Scrutinize for problems, gaps, and failure modes, and frame every finding through cost, risk, time-to-value, and opportunity cost. Lead with what breaks and what it costs.",
]
Edit a source mode once and every combo that names it inherits the change, because the combo holds a pointer instead of a copy.
Installing a pack, and three things that will bite you
Save your commands to a file, then import it. Import is append-only, so it can never damage what you already have.
Save the pack as operator-modes.toml anywhere on disk.
Run base commands import ~/operator-modes.toml
Verify with base commands list. It prints what was added and what was skipped.
Type *blunt in any prompt and watch the next answer change.
Three things that caught me:
Import skips names you already have, silently by design. BASE ships with a *debug, so a pack containing its own *debug gets skipped and you never notice you are still running the old one. Run base commands remove debug first if you want yours to win.
Extra fields get stripped. BASE stores name, description, and rules. Anything else you put in the TOML is dropped on import. This is why a combo has to name its sources inside its own rule text instead of in a separate field. The prose is load-bearing.
A combo pointing at a command that does not exist fails at runtime, not at import. It imports clean, then half-runs the first time you type it. Check that every source a combo names actually appears in base commands list before you trust it.
Back up first, every time: cp ~/.base-gbl/commands.toml ~/.base-gbl/commands.toml.bak
How to update your CLAUDE.md
Read it out loud, section by section. For each one ask: could a keyword or a file path reliably tell Claude to load this?
If yes, it belongs in a BASE domain. Brand rules, tool rules, writing rules. All of it.
If no, leave it. Your folder layout is the main one. You need it before any keyword could match.
Replace what you moved with one pointer line so you can still find it later. Something like: "Coding rules live in the BASE DEV domain. Run base domain get DEV."
Check the trigger words first. Before you delete a section, make sure the domain actually fires on the words you really type. A rule that never triggers is worse than a rule that is always on.
How to update your knowledge domains
See what you have:base domain list shows every domain, its triggers, and its rule count.
Look inside one:base domain get DEV prints that domain's rules.
Add a trigger word:base domain add-trigger --domain DEV --keyword typescript
Make a new domain:base domain create --name PDF --keyword pdf --keyword invoice
Check your work:base doctor should say HEALTHY. Run it after every change.
Two habits that will save you real time:
Audit by asking what still exists, not by asking whether the rule sounds smart. Dead rules read perfectly fine and only fail when they fire.
Check for overlap before you add. If a rule already lives somewhere else, merge it instead of writing it twice. Two copies drift apart and you will not notice.
Humanizer: Strip The AI Tells Out Of Everything Claude Writes
Claude writes clean and it writes obvious. This skill runs two passes over the draft, kills 35 documented AI patterns, and rewrites it in your voice once you feed it your own writing.
What it does
Humanizer is an agent skill that rewrites AI-generated text so it stops reading like AI, without touching the facts. It works off 35 documented patterns pulled from Wikipedia's Signs of AI writing, maintained by WikiProject AI Cleanup. Inflated language, em dashes, forced word pairs, passive voice, chatbot sign-offs like "I hope this helps," all of it goes.
Why it's needed
Every email, post, and landing page you hand to Claude comes back grammatically perfect and instantly recognizable. Your prospects have read a thousand of those. The moment they clock the rhythm, the message stops being from you and starts being from a tool, and that is the whole ballgame when you are trying to sell something.
How it works
Two passes, and that second pass is the part most rewriters skip.
Pass one: rewrites the text without treating the original structure as fixed. It is allowed to break the shape, not just swap synonyms.
Pass two: checks the new draft against all 35 patterns AND against the original claims, then makes final adjustments.
Feed it your own writing samples and it matches your voice instead of a generic human one. On files it preserves code, data, and frontmatter, so you can point it at a real document without it mangling the parts that are not prose.
How I use it
I gave it 36 of my own YouTube transcripts, so it learned how I actually talk. Now anything going in front of a client or a prospect runs through it first. It is also wired into my bottleneck report workflow: that pulls a prospect out of the pipeline, drafts the email, then fires the humanizer skill and the subject line skill so the output sounds like me every single time instead of once in a while.
Install it
One command, global:
npx skills add blader/humanizer --global
On Claude Code 2.1.142 and newer you can also run /plugin marketplace add blader/humanizer. Claude Desktop needs the repo downloaded as a ZIP and uploaded as a skill.
What it will not do
It does not invent facts. Names, dates, quotes, and citations have to come from your source material or from you. It cleans how the thing reads, it does not fill in what the thing does not know. MIT licensed, 37,200+ stars.
Anthropic locked Claude Design behind their cloud. This is the open-source clone that runs locally and uses the Claude Code CLI you already have. Mac, Windows, Apache 2.0.
What it does
Open Design (nexu-io/open-design) is the open-source answer to Claude Design. Same idea, none of the lock-in. Local-first, Apache 2.0, runs on whatever coding-agent CLI is already on your machine. Claude Code, Codex, Cursor Agent, Gemini CLI, GitHub Copilot CLI plus 11 more.
Pick a skill (saas-landing, dashboard, mobile-onboarding, magazine-poster) and a design system (Linear, Stripe, Notion, Apple, plus 68 others). Fill the 30-second discovery form. Your CLI streams an artifact (HTML, PDF, PPTX, MP4) into a sandboxed iframe. Save it to disk. The file is yours.
Why it's needed
Claude Design is closed-source, Anthropic-locked, and cloud-only. Two weeks after it shipped, Open Design dropped: same workflow, your laptop, your keys, your filesystem. Artifacts land as real files under .od/projects/<id>/. Git them. Ship them. Deploy them to Cloudflare Pages from the app.
If you already pay for Claude Code, this is free leverage on top of your existing subscription.
What's in the playbook
Install on Mac and Windows with the gotchas (Apple Silicon-only binary, Windows SmartScreen bypass) on the same page as the steps.
Build from source path for Intel Macs, Linux, and WSL2. Plus the Docker variant if you do not want to touch your local Node.
Agent setup: wire up Claude Code (or any of 16 supported CLIs), or fall back to BYOK with Anthropic, OpenAI, Azure, or Gemini keys.
Five day-one workflows: SaaS landing, investor deck, mobile onboarding, email plus social burst, dev handoff.
Five gotchas that will trip you up if you read the README at 1am.
Who should read it
Agency owners, freelancers, and operators who charge for design deliverables. Anyone who already runs Claude Code and wants to ship landing pages, decks, mobile flows, and email campaigns from the same machine. Skip it if you only push pixels in Figma.
CC Switch: One-Click Provider Swap for Claude Code
Anthropic almost killed Claude Code on the Pro plan last month. CC Switch lets you swap your provider in 50ms from a menu bar icon. Bedrock, OpenRouter, NIM, custom. Never restart.
What it does
CC Switch (farion1231/cc-switch) is a Tauri plus Rust desktop app that lives in your menu bar. Click the icon, pick a provider, your Claude Code keeps running on the new endpoint. No terminal restart. No environment variable surgery. 50ms swap.
Supports Claude Code, Codex, OpenCode, Gemini CLI, OpenClaw, and Hermes Agent. 50 plus preset providers including Anthropic, Bedrock, OpenRouter, NIM, plus full custom config for your own proxy.
Why it matters
On April 21 2026, Anthropic briefly delisted Claude Code from the $20/mo Pro plan. The Reddit thread hit 2,600 upvotes in a day. The Register and XDA covered it. Half the community started routing Claude Code through Bedrock, OpenRouter, and self-hosted proxies overnight.
CC Switch became the cleanest way to do that without losing your mind. It also doubles as MCP server management and Claude Skills sync, bi-directional, so both your Claude Code config and your CC Switch config stay aligned.
If you ever burn through your Pro plan limit and want to keep coding without paying API rates, this is the tool.
Best for
Pro plan users who hit the rate limit weekly and want a clean fallback to Bedrock or OpenRouter
Teams managing multiple CLIs (Claude Code plus Codex plus Gemini) from one machine
Anyone running a custom proxy (Cloudflare AI Gateway, LiteLLM, internal routing) who wants a one-click switcher
MCP server tinkerers who want a GUI instead of editing JSON
Install
macOS: download the latest .dmg from the GitHub Releases page. Windows: .exe installer. Linux: .AppImage.
First launch will ask for permission to read your ~/.claude/ config. Grant it. The app reads your existing Claude Code setup, lists detected providers, and lets you add new ones with a few clicks.
Shipping releases multiple times a day right now. Keep auto-update on.
The first physical Claude Code product. $30 ESP32 desk gadget with a 2.16-inch AMOLED screen, pairs over BLE, animates a pixel cat that gets busier as you approach your rate limit.
What it does
Clawdmeter (HermannBjorgvin/Clawdmeter) is a DIY ESP32-S3 desk gadget with a 2.16-inch AMOLED screen. It pairs to your Mac over Bluetooth, reads your Claude Code usage live, and displays Current and Weekly progress bars on its tiny screen. A pixel-art mascot called Clawd animates faster as you near your rate limit.
Five days old. 1,000 plus stars. TechCrunch covered it. CNX Software covered it. The first physical product built specifically for Claude Code.
Why it exists
If you live in Claude Code all day, you check your usage dashboard 15 times a day. That is 15 context switches that wreck your flow.
Clawdmeter puts your usage on a small object sitting next to your keyboard. Glance, see the bars, get back to coding. No browser tab. No dashboard refresh. No mental tax.
The hardware is $30 in parts from AliExpress. The firmware and Mac pairing app are open source. The whole thing was built by one developer in Iceland using Claude Code.
Hardware you need
ESP32-S3 dev board with built-in 2.16-inch AMOLED display (around $25 on AliExpress)
USB-C cable for flashing and power
Optional: 3D-printed case (STL files in the repo)
Total: $30 to $40 including shipping. Same parts the creator used.
Setup
Flash the firmware from the repo using esptool or the Arduino IDE. Download the Mac companion app from GitHub Releases. Pair over BLE, grant the app permission to read your Claude Code usage logs, mount the device on your desk.
Linux and Windows companion apps are on the roadmap.
Your Claude Code bill is not about what Claude writes back. It is about what you send in. ccGlass is a free local proxy that shows you every token leaving your machine, and the hidden leak that is draining your wallet.
What it does
ccGlass (jianshuo/ccglass) is a free, open-source local proxy that sits between Claude Code and the model. Every request your agent makes passes through it, and a live web dashboard shows you exactly what was sent: input tokens, output tokens, cache hits, cost, latency, and the full payload of every single call.
It works with Claude Code, Codex, and Kimi. One command to start, and a browser tab streams every turn as it happens.
Why it matters
Most people think their AI bill comes from what Claude writes. It does not. I ran 10 requests and watched it live: 1,707,749 tokens going in, 4,341 tokens coming out, $11.57 spent. That is roughly 393 tokens sent for every 1 token Claude wrote back. You pay for the send, not the answer.
The hidden leak ccGlass exposed on my own setup: over 500 tool definitions loaded on every single call. Every MCP server you connect and forget about re-sends its entire tool menu into context on every request, whether you use it or not. You cannot fix what you cannot see. ccGlass makes it visible.
What you'll learn about your token spend
You pay for input, not output. The cost is dominated by what you send the model, not what it returns.
Connected MCP servers are your biggest leak. Each one injects 100s of tool definitions into every call. ccGlass shows the exact count per request.
Your cache hit rate is money. A 65 to 70 percent hit means a third of your context is re-processed at full price every turn. Watch it climb as you trim.
Failed calls still cost you. ccGlass flags errors (like a 404 on a model variant) so you stop paying for round-trips that never landed.
The fix it points you to
Once you can see the leak, the fix is simple: disconnect the MCP servers you are not actively using. Your tool count drops, your input tokens collapse, your cache hit rate climbs, and the same work costs a fraction. ccGlass lets you re-run and watch the numbers move in real time, so you can prove the savings instead of guessing.
Best for
Anyone burning through their Pro plan limit weekly who wants to know where the tokens actually go
Heavy MCP users (GoHighLevel, Slack, Canva, Apify, and the rest) whose context is quietly bloated by unused tool menus
Agency operators running Claude Code on client work who need to control and explain their AI spend
Anyone whose agent "feels dumb" and wants to see what context it actually received
Install
Grab it from the repo and start the proxy, then point Claude Code at the local endpoint. Open the dashboard in your browser and run a normal session. Every request streams in live with its full token breakdown.
Free and open source. It reads your traffic locally, nothing leaves your machine beyond the calls you were already making.
I charged a client ten grand for this exact build. Stack NotebookLM with Claude Code and you get a citation-backed AI assistant for free.
I charged a client ten thousand dollars to build them a personal AI assistant.
It reads every document in their business. SOPs, contracts, sales call transcripts, onboarding docs. Ask it anything and it answers with exact citations. Not hallucinations. Actual references.
Today I published a full tutorial showing how to build the exact same thing for free. Two tools. No coding. No monthly fees.
What I cover
Sales recon notebooks. Feed call transcripts in and generate presentations for your next close.
SOP notebooks. Give your team something they can chat with instead of constantly bugging you.
Client onboarding notebooks. Keep project knowledge updated as things evolve.
Why this combo works
The secret sauce is stacking NotebookLM, which is free from Google, with Claude Code. NotebookLM stores and processes the data. Claude Code automates the whole thing.
Together, they act like a sixty thousand dollar employee that never forgets anything.
The kicker: NotebookLM does not cost you extra tokens. You are basically getting a second AI brain for free on top of your Claude subscription.
Stop bouncing between Midjourney, Google AI Studio, and Canva. One MCP server turns Claude Code into a thumbnail and ad image machine.
Is it possible to use Claude Code to generate high-quality ad images, YouTube thumbnails, and social media content? Yes.
It's called the Nano Banana MCP server, and paired with Canva's Magic Layers, it might be the best image creation stack available right now.
The full workflow
Generating images from scratch
Editing existing ones
Iterative refinement
Style transfers
Combining it all with Canva for scalpel-level editing on every asset you create
No more context switching between Midjourney, Google AI Studio, and your code editor. One terminal. One conversation. Production-ready images at any aspect ratio.
I also show the old way I used to do this with HiggsField and custom GPTs, so you can see exactly why this stack is a game changer.
Google is not the only search anymore. Audit your site for ChatGPT, Perplexity, and Gemini visibility with 13 subskills and a client-ready PDF.
What it does
Audits websites for AI search visibility (ChatGPT, Claude, Perplexity, Gemini, Google AI Overviews) with 13 specialized subskills covering citability, crawler access, llms.txt, brand mentions, schema, technical SEO, and content quality. Generates client-ready reports including PDFs with charts.
Why it's needed
Traditional SEO still matters, but AI search is where traffic is going. If your site isn't getting cited by AI, you're invisible to a growing slice of users. This skill audits the gap and tells you exactly what to fix.
When to use it
Onboarding a client website. Pre-launch checks. Auditing your own sites. Building reports you can hand off without editing.
Done guessing what to post? This plugin researches competitors, writes hooks, scripts videos, and learns from your winners inside Claude Code.
This video is going to show you how to turn Claude Code into a real content engine, not just a script writer.
Most creators are stuck in reaction mode. Every day they wake up wondering what to post, what angle to take, what hook might work, and whether the content will land. That is not a system. That is survival mode.
Go Viral Bro fixes that.
Inside the video, I walk through the exact Claude Code workflow I built to research what is already working in your niche, break down your competitors, generate stronger content angles, write better hooks, script full videos, analyze your winners, and feed that data back into the system so it gets smarter over time.
This is the same kind of loop that used to require a team of researchers, writers, strategists, and VAs. Now you can run it through Claude Code.
What you'll learn
Install the Go Viral Bro plugin step by step
Set up onboarding and training so Claude understands your brand, audience, competitors, and posting style
Use /viral discover to scrape competitor content and trending ideas across YouTube, Instagram, Reddit, and GitHub
Transcribe videos, study verbal hooks, and break down visual hooks so you understand why content wins
Use /viral angle to generate stronger content ideas based on contrast, not guesswork
Use /viral script for long-form scripts, short-form scripts, LinkedIn posts, hooks, talking points, filming cards, and CTAs
Use /viral analyze to study your own top-performing content and see what's actually working
Use /viral update-brain to feed winning data back into the system so your agent improves with every post
You are not just installing a plugin here. You are building a repeatable content pipeline.
This is for you if
You create content for your business and want a system
You are tired of guessing what to post
You want Claude Code to help with growth, not just coding
You want better hooks, angles, and scripts
You want your content process to improve over time instead of starting from zero every day
The goal is simple: stop posting based on vibes. Start posting based on a system that learns.
Claude Skill Social Post: Voice-Cloned Auto-Poster
Trains on your Facebook voice, then ships a 14-day calendar to FB, IG, Threads, and X. Author's first post hit 72K reach, 358 likes, 443 comments, gained 700 community members.
Built by Hao (駱君昊). Feed it your existing posts, it learns the rhythm of how you write, then it drafts a two-week posting calendar and pushes the posts out across Facebook, Instagram, Threads, and X.
What makes this one worth pulling in is the proof. The author posted his own first auto-generated post and it broke through hard: 72K reach, 358 likes, 443 comments, and 700 new community members. Day two flopped. He wrote a postmortem about why and shipped it inside the same repo. Most "AI content" repos hide their losses. This one teaches from them.
What it does
Voice fingerprint from your existing Facebook content
14-day content calendar generation
Cross-post to Facebook, Instagram, Threads, and X
Day-2 flop postmortem doc included as a teaching artifact
Pair it with
Run this after Go Viral Bro generates your script. Go Viral Bro handles the research and the script. This one handles the voice, the calendar, and the multi-platform push.
X Article Publisher: Markdown to X (Twitter) in One Command
Write in Markdown, publish a fully formatted X Article in 2-3 minutes instead of 20-30. Handles headings, bold, links, images, tables, dividers, and Mermaid diagrams automatically.
By wshuyi. If you've ever tried to publish a long-form article on X, you know the pain: copy from Markdown, paste into X, watch every bold, link, and heading disappear, then spend 20 minutes re-formatting by hand. Each image is a five-click upload. Each Mermaid diagram is impossible.
This skill kills that workflow. You point Claude Code at a Markdown file. It opens X Articles, types the content, applies the formatting, uploads the images in the right positions, renders Mermaid as images, drops the cover. Done in 2-3 minutes.
What it does
Markdown to X Articles, end to end
Cover image auto-upload
Inline image upload at exact positions in the article
Tables rendered as images for visual fidelity
Mermaid diagrams rendered to images
Divider support and cross-platform clipboard
10x time reduction per article (20-30 min down to 2-3 min)
Pair it with
Use Go Viral Bro to research your angle and generate the long-form script. Use this skill to ship it to X Articles in minutes. Claude Skill Social Post handles the short-form cross-post across FB, IG, Threads, and X feed. Together: research engine, long-form distribution, short-form distribution.
Generate leads on LinkedIn without paying for Sales Nav tools. Five skills handle DMs, replies, posts, and connection requests through your Claude plan.
I've been using an unpackaged version of this in my Claude workspace to generate leads. After getting some requests to share it, I turned it into a plugin with skills.
It uses Chrome via the browser extension to run. No hidden costs, no fees, you just need a Claude Code plan.
What it does
DM Connections. Send personalized openers to your existing connections.
Answer DMs. Reply to unread messages with contextual, on-brand responses.
Create Posts. Generate LinkedIn content from proven templates.
Grow Network. Send strategic connection requests to ICP matches.
Higgsfield Seedance 2.0 Skills: Video Prompts on Autopilot
Stop wrestling with video prompt syntax. Fifteen skills feed your concept to Claude, Playwright hits Higgsfield, and your Seedance render kicks off untouched.
What it does
A pack of 15 dedicated skills that turn Claude Code into a production-ready video prompt engine for Seedance 2.0. You describe the concept. Claude handles camera control, lighting, consistency, and the full production pipeline. One prompt in, fully engineered video prompt out.
What you can generate
Brand ads
Product shots
Cinematic scenes
Anime
E-commerce videos
Basically any style you can describe. The ninja scene demo was generated 20 minutes before filming. That is the speed.
Why it matters
You are combining the smartest LLM on the market with the best video generator and offloading the prompt engineering entirely to Claude. No more wrestling with prompt syntax. No more hunting for camera direction. The skills already know what Seedance 2.0 needs and Claude writes to spec automatically.
The automation layer
Wired into Playwright CLI for browser automation. Once Claude builds your prompt, Playwright logs into Higgsfield, selects Seedance, pastes the prompt, uploads reference images, and starts generation. You keep building inside Claude Code while videos render in the background.
Paste the repo link into Claude and ask it to install the skills
Give Claude your video concept and style (cinematic, product, e-commerce, anime, whatever)
Tell Claude: "Do your magic using Playwright"
That is it. You get a production-ready video without touching Higgsfield directly.
Bonus: the Higgsfield deal
Sign up today and Higgsfield is running 70 percent off with unlimited Seedance 2.0 access. One of the best video generation stacks on the market, cheaper and faster than almost anything else.
Higgsfield AI MCP: Every AI Image + Video Model Inside Claude Code
Stop juggling Kie API and a stack of standalone API keys. One MCP server connects every AI image and video model to Claude Code, terminal-first.
What it does
Higgsfield just dropped an MCP server that plugs every AI image and video model on their platform directly into Claude Code. Nano Banana Pro, Seedream 4.5, Soul, Seedance, Kling. All running from your terminal, all through one credit pool.
You write 3 words. Claude picks the right model, writes the prompt, fires the gen.
What you can generate
YouTube thumbnails
Skool community covers
Custom characters and logos
AI avatars for client ads
Product ad creative
Cinematic 16:9 hero cuts and 9:16 shorts (rendered in parallel from one prompt)
Image AND video, side by side, no tab-switching, no re-prompting from scratch every time.
Why it crushes Kie API and standalone APIs
Most people stack 4 to 5 separate API subscriptions to access these models. Different billing accounts. Different keys. Different docs. Different rate limits.
Higgsfield consolidates the entire stack. One platform. One credit pool. One MCP. The cost math alone is a no-brainer once you compare your current API spend to a single Higgsfield plan.
Why MCP is the unlock
Think of MCP as a toolbox. Every model becomes a tool inside it. Claude is the brain that figures out which tool fits the job. You stop being the prompt engineer and start being the orchestrator.
Feed Claude your brand context once. Every image and video Claude generates after that lands on-brand without you re-explaining the rules.
Paste it into Claude Code and ask Claude to install the MCP for you
Run /mcp, click the Higgsfield entry, hit Allow to authenticate
Start generating: ask Claude for a thumbnail, an avatar, a 16:9 hero, a 9:16 short, anything
That is it. Claude is now connected to every model on Higgsfield, no API keys to manage on your side.
Stack with the Seedance skills
Pair this MCP with the Higgsfield Seedance 2.0 Skills already in this toolkit. The skills hand Claude pre-engineered prompt templates for cinematic, brand story, motion design, e-commerce, fight scenes, and more. Combined with the MCP, Claude picks the skill, writes the prompt, fires the gen end-to-end.
Get access to the MCP
Want the Higgsfield MCP wired straight into Claude with image and video models on tap? Get Higgsfield MCP access here (my affiliate link). One subscription, one credit pool, every model in this resource.
Build Video Games With Claude Code: One Prompt to a Deployed Game
You do not need Unity, a game engine, or a single line of your own code. Describe a game to Claude Code, hand it screenshots of games you love, and get a playable, shareable URL the same day.
What it does
Claude Code plus the Higgsfield MCP now builds full browser games end to end. You describe the game in plain English, Claude plans it, generates every sprite, tileset, texture, and sound with Higgsfield's models (including Fable 5), writes the game code, and deploys it to a live URL you can share or drop on your site. No Unity, no game engine, no boilerplate you have to touch.
The whole thing runs from your terminal through one credit pool. You are not the programmer or the pixel artist here. You are the director.
The reference trick: feed it games you love
This is the part nobody shows you, and it is the single biggest quality unlock. The game pipeline accepts reference images, mood boards, character art, screenshots of existing games, and locks their look into a single STYLE FORMULA that gets stamped byte-identical into every asset it generates. That is why a game built from references reads like one coherent world instead of a pile of mismatched AI art.
So do not describe the art style in words. Show it. Where to pull references:
Pinterest. Search "pixel art platformer", "isometric game art", "GBA RPG tileset", or the name of a game plus "art style". Pinterest boards are the fastest way to assemble a 6-image mood board in one look.
Google Images. Search "[game name] screenshot" or "[game name] gameplay" and grab 4 to 6 frames that capture the palette, lighting, and mood you want. Steam store pages and press kits are clean, high-res sources.
Your own favorite games. Screen-grab a game you already own while you play (a Celeste run, a Hollow Knight cave, a Vampire Survivors wave). Your own screenshots are the most honest reference for "make it feel like this."
One character, many angles. If you have a hero you want, give it front, side, and back so Higgsfield can register it as a consistent element across every scene.
Then say: "Build me a game that plays like Vampire Survivors but looks like these," and attach the images. The reference speeds up every decision without replacing a single step of the build.
How the pipeline actually runs (plan → build → deliver)
Under the hood the game skill runs a fixed three-phase pipeline, and knowing it helps you steer:
Plan. Claude fixes the game profile (genre, players, controls, platforms), writes an asset manifest (one row per sprite, tile, texture, sound), and derives the STYLE FORMULA from your references. If your brief does not pin a style, it stops and asks you to approve the formula first. Approve it, because that one line governs the entire look.
Build. It fires every asset generation in parallel while it writes the game code from proven skeletons, wiring each asset in as it lands. Desktop and mobile browsers with gamepad support are the default target, so your game works on a laptop, a phone, and a Steam Deck out of the box.
Deliver. It deploys the finished game and hands back a real, shareable play URL (and lists it in the Higgsfield games catalog). That link is the deliverable, live the same day.
Honest framing: these are browser games built from generated 2D/3D assets and hand-written game logic, not AAA engine titles. That is exactly why one person with a good reference board can ship one in an afternoon.
Paste the command into Claude Code and ask Claude to install the MCP for you
Run /mcp, click the Higgsfield entry, hit Allow to authenticate
Assemble your reference images, then say: "Build me a game like [favorite game] but [your twist], matched to these screenshots"
Within that same day you can have a deployed game with a link you can send anyone.
Playbook: from a game you love to a deployed one in your style
Pick the feeling, not the clone. Choose one game whose vibe you want (Vampire Survivors' escalation, Flappy Bird's one-tap tension, Stardew's cozy loop). Name it to Claude as the reference; it pre-fills the mechanics questions.
Build the mood board. Pull 4 to 6 reference images from Pinterest, Google, or your own screenshots. Aim for one consistent palette and art style across all of them, mixed styles produce mixed output.
Give the twist. One sentence on what makes it yours: the theme, the character, the setting. "Vampire Survivors, but you are a barista and the enemies are Monday-morning customers."
Approve the STYLE FORMULA. When Claude posts the derived style, read it and confirm. This is your one lever on the whole look, so use it.
Let it run, then playtest. Assets generate while the code is written. Open the delivered URL, play it, and send back specific fixes ("enemies too fast", "make the jump floatier"). Iterate in plain English.
Ship the link. Post the play URL to your community, embed it, or use it as a lead magnet. A game people can actually play beats another PDF.
What you need
Claude Code as the builder (it plans, codes, and orchestrates the whole run).
Higgsfield MCP for the game pipeline and every image, video, 3D, and audio asset (one credit pool, models including Fable 5).
A reference board of 4 to 6 images. This is optional but it is the difference between "fine" and "how did you make that."
Tired of one-shot PRDs blowing up your context? PAUL runs client builds in plan, apply, and unify loops with human checkpoints between phases.
What it does
PAUL is a free, open-source Claude Code plugin built for iterative development with human checkpoints. Plan the work, apply the changes, unify the plan against reality, then loop. Keeps every project on rails without burning context on sprawling one-shot PRDs.
Why it exists
GSD has been the go-to Claude Code plugin since it dropped. After using it on real client builds, I found 7 structural problems that kept showing up. PAUL solves every single one of them.
In the video I break down the 7 problems from a non-technical angle, walk through installing and using PAUL at a high level, and share my honest take on when to use GSD vs PAUL. Not about pushing you to one or the other, about giving you the perspective to pick the right tool for your project.
When to use PAUL vs GSD
PAUL: iterative client builds, longer projects, anything that needs human checkpoints between phases
GSD: tight well-defined features that fit inside one PRD and one automated Ralph loop
Your CLAUDE.md is bloated and loading every rule on every prompt. CARL splits rules into domains and only fires what the task needs.
What it is
CARL (Context Augmentation & Reinforcement Layer) is a plugin for Claude Code that replaces a static CLAUDE.md with dynamic, keyword-triggered rule sets. Instead of loading every rule on every prompt, CARL organizes rules into domains and only loads what matches the task you are working on.
Result: fewer tokens, less noise, better outputs. Claude only sees rules that actually matter.
How it works
Your rules split into domain files:
Global - always loads. Coding standards, safety, formatting
Development - loads when you mention fix, bug, build, test
Content - loads when you mention script, video, YouTube, post
Clients - loads when you mention client, deliverable, proposal
Any custom domain you create
Each domain has trigger keywords defined in a manifest file. When your prompt matches, that domain injects into context. Everything else stays out.
Example
Prompt: Help me write a YouTube script about CARL
CARL detects YouTube and script, loads Content (19 rules) + Global (9 rules). Total loaded: 28 rules. Dev rules, client rules, outreach rules do not load because they are irrelevant.
Install
Run in your terminal:
npx carl-core
Choose Global (applies to all Claude Code projects, recommended) or Local (current project only). Restart Claude Code. Done.
What gets created
.carl/
├── manifest Controls domains and trigger keywords
├── global Rules that always load
├── commands Star command definitions
├── context Context aware rules
└── development Example domain
Key concepts
Domains: a file of related rules. You name it, define keywords, write the rules. CARL decides when it loads.
Manifest: controls activation. RECALL keywords trigger loading, EXCLUDE keywords block, ALWAYS_ON loads every session, STATE is active or inactive.
Rule format: simple key-value pairs like CONTENT_RULE_0=Write in Charles's voice direct practical no BS.
Star commands: manual overrides like *carl (open interactive help) or *brief (force short output). Create your own for any workflow.
Why this matters
Small CLAUDE.md files are fine at first. They break once you add project rules, client preferences, content guidelines, and workflows. Static files do not scale. Everything loads every time. Tokens wasted, attention diluted.
CARL scales with you. Add unlimited domains. Claude only loads what it needs.
Context rotting halfway through a big build? GSD turns your project into spec files and fresh subagents so each phase starts with 200k clean tokens.
The problem
Claude has a context window. The longer your conversation, the more it forgets. By the time you are deep into a multi-phase project, Claude is working with degraded context and making dumb mistakes.
Most people "solve" this by starting fresh conversations (losing context), copy-pasting requirements repeatedly (tedious), or hoping it works (it will not). GSD solves it properly.
What GSD actually does
A spec-driven development system. Instead of one long conversation that degrades, GSD:
Extracts requirements upfront into PROJECT.md
Breaks work into phases via ROADMAP.md + STATE.md for tracking
Plans each phase atomically with detailed PLAN.md and verification steps
Executes with fresh subagents, each phase gets 200k clean tokens
The key insight: curated documents beat raw conversation history.
The core workflow
/gsd:new-project - guided questioning to extract what you are building. Creates PROJECT.md.
/gsd:create-roadmap - breaks the project into logical phases. Creates ROADMAP.md and STATE.md.
/gsd:plan-phase 1 - generates atomic task plans for phase 1. Each task has clear instructions and verification criteria.
/gsd:execute-plan - fresh subagent picks up the plan and executes. No context degradation. Clean commits per task.
Repeat plan-phase + execute-plan per phase.
Commands you will actually use
/gsd:help - shows all commands
/gsd:progress - where am I, what is next
/gsd:plan-phase N - plan a specific phase
/gsd:execute-plan - run the current plan
/gsd:add-phase - add work to the end
/gsd:insert-phase N - slip urgent work between phases
/gsd:pause-work - save state, pick up later
/gsd:resume-work - continue where you left off
/gsd:map-codebase - for existing projects, documents what is there
When to use GSD
Use it for: new apps from scratch, multi-phase features, anything over one session, projects where you cannot afford confusion.
Skip it for: quick fixes, single-file changes, simple scripts.
Rule of thumb: if you would normally lose track of what you are building halfway through, use GSD.
Why it works
Context engineering: PROJECT, ROADMAP, STATE, PLAN files give Claude exactly what it needs, nothing it does not
Fresh subagents: each phase starts with full context capacity
Atomic commits: one task = one commit = clean git history you can roll back
Built-in verification: each task has success criteria, not just instructions
Installation
npx get-shit-done-cc --global
That is it. Commands are now available in Claude Code.
The bottom line
GSD treats complex builds like they deserve to be treated: with structure, documentation, and fresh context for each phase. Stop hoping Claude remembers what you are building. Make the system remember for you.
You define the work, GSD 2 ships it overnight. Autonomous agent with budget ceilings, crash recovery, and 20-plus model providers baked in.
An autonomous coding agent that breaks your project into milestones, slices, and tasks, then executes them one by one with fresh context windows, automatic git commits, and real-time cost tracking.
You define the work. It runs the work.
It is not locked to Claude Code either. It supports 20+ providers, including Anthropic, OpenAI, Google, OpenRouter, and GitHub Copilot. You can even assign different models to different phases, so research uses one model and execution uses another.
Built for developers who want to ship full projects overnight instead of prompting one task at a time.
YC president Garry Tan built 12 Claude Code modes for CEO-level review, paranoid code checks, QA, and release engineering. Better constraints, better output.
A skill system for Claude Code created by YC President Garry Tan. Instead of one generic assistant, you get 12 specialized modes like:
CEO-level product review
Paranoid code review
Automated QA against a live URL
Release engineering
Each mode constrains Claude to think like a specific role with specific objectives.
Want Claude to drive GIMP, Blender, or any desktop app that ships zero API? This framework scans the code and auto-generates a real CLI wrapper.
A framework that auto-generates command-line interfaces for any desktop application so AI agents can control software that does not have an API.
Point it at GIMP, Blender, LibreOffice, or dozens of other apps, and it analyzes the codebase, builds a full CLI with testing, and publishes an installable package.
The generated CLIs make real calls to real software, not toy wrappers.
Stop wiring n8n nodes one at a time. Claude reads the node catalog, picks the right nodes, and hands back an importable workflow JSON.
Watch the video to see how I use the n8n MCP server to build workflows for me. Claude reads the n8n node catalog over MCP, picks the right nodes, wires them together, and hands you back a working workflow JSON you can import.
Install 34 marketing skills in one command, then lock Claude to your brand in CLAUDE.md so every page it builds ships on-brand. No AI slop.
Short video from day 3 of the 90 Days Making AI Easy series. 34 marketing skills. One command to install. The real move is what you do after you install it.
What the playbook covers
Install the plugin in one command. Literally copy and paste.
Stack it with the built-in /frontend-design skill so your pages look good and convert.
Get Claude to generate a brand kit PDF with your colors, fonts, tone, and audience.
Lock it into CLAUDE.md so Claude stays on brand every single session.
The result
You tell Claude to build a landing page and it already knows your brand. No repeating yourself. No off-brand output. No AI slop. This is the exact setup I use for client work at C&C Strategic.
Build real voice agents for law firms, trades, or accountants without touching the ElevenLabs dashboard. Plain English prompts, Twilio numbers, done.
In this module you will learn how to build fully configured voice agents using Claude Code and the ElevenLabs MCP server without ever touching the ElevenLabs dashboard.
What you'll learn
How to install and connect the ElevenLabs MCP server to Claude Code
How to create voice agents using plain English prompts
How to add custom knowledge bases to your agents
How to connect a Twilio phone number so your agent can take real calls
3 ways to monetize voice agents for your business or clients
What we build
Intake receptionist for a law firm
Customer support agent for a construction company
After-hours agent for an accounting firm
What you need
Claude Code (Pro or Max plan)
ElevenLabs account (free tier works to start, Creator plan for production)
Twilio account (optional, only if you want a live phone number)
Ralph was built to spawn fresh agents per task. The plugin broke that. Use my custom script to run Ralph properly without context rot.
In this episode I go deep on the Ralph plugin and compare it head-to-head with my favorite Claude Code plugin, GSD (Get Shit Done). Ralph was originally designed to spawn fresh AI instances for every task to keep quality consistent. The current Ralph plugin leads to context degradation instead. I show how to use Ralph properly with my custom script, then walk through GSD's fresh-subagents-per-phase approach.
Live demo at the end: building a Kanban Pro app using Ralph the right way, with proper context management. If you enjoyed the video, thumbs up and subscribe for more Claude Code content.
Skip the API-key juggling for every SaaS tool you use. Claude in Chrome clicks, types, and navigates your already-logged-in browser sessions.
Claude in Chrome is a browser extension that lets Claude see, click, type, and navigate websites directly. Turn Claude into a hands-on operator that uses your active browser sessions to perform tasks without API keys or complex re-authentication.
What you'll learn
Automating visual testing for frontend development
Interacting with authenticated business tools through your already-logged-in browser
Debugging live applications with real-time console and network monitoring
The 17 specialized MCP tools for managing tabs, recording interactions, and running JavaScript for advanced automation
The playbook
The PDF below is the full playbook: install steps, every tool, real workflows you can copy, and how to wire it into Claude Code.
Anthropic just shipped a research preview that gives Claude full operator access to your Mac. Native apps, iOS Simulator, design tools, anything with a UI.
Computer use is a built-in MCP server inside Claude Code that lets Claude take screenshots, click, type, and scroll on your real desktop. Once enabled, Claude can drive any native macOS app, the iOS Simulator, system settings, design tools, anything a CLI or browser cannot reach. The whole loop runs from the same terminal session that wrote your code.
What you'll learn
Enable the computer-use MCP server in three steps from /mcp
Grant the two macOS permissions (Accessibility + Screen Recording) once and forget about it
Per-app approval, sentinel warnings, and the three control tiers (read, click, full)
Safety model: machine-wide lock, terminal excluded from screenshots, global Esc kill switch
Troubleshooting checklist when computer-use does not appear in /mcp
Requirements
macOS only. Claude Code v2.1.85 or later. Pro or Max plan (no Team or Enterprise yet). claude.ai authentication, no third-party providers. Interactive session, the -p flag is not supported.
The playbook
Nine-page PDF playbook: enable flow, permission tiers, runtime behavior, five real workflows, safety guardrails, and a troubleshooting checklist. Built from the official Anthropic docs.
Run these two together and Claude stops guessing at taste. UI UX Pro Max locks the system, Anthropic's built-in front-end design skill executes it. 67 styles, 96 palettes, 161 reasoning rules.
What it does
A comprehensive design system engine with 67 UI styles, 96 color palettes, 57 font pairings, 13 stack presets (React, Next, SwiftUI, Flutter, etc.), and 161 reasoning rules. Generates a tailored design system for your project before you write a line of code.
Why it's needed
Going from "build me a landing page" to "build me a landing page that matches the vibe of a beauty spa" takes a lot of upfront taste decisions. UI UX Pro Max encodes those decisions and produces a full system in seconds.
When to use it
Starting a new frontend project and you want the design language locked before code. Great for client work where you need to justify choices. Also solid for avoiding the same default aesthetic on every project.
Run it with Anthropic's front-end design skill
These two are a pair, and I never run one without the other. Anthropic ships /frontend-design built into Claude Code, no install needed. It handles the execution: production-grade components, spacing, motion, the parts that make a page feel built instead of generated.
UI UX Pro Max handles the decisions. Style, palette, type pairing, the system your project is going to live inside.
Run them in tandem and each covers the other's blind spot. The system skill has taste but no hands. The design skill has hands but picks the same safe defaults every time. I have this wired as a standing rule in my BASE domains so it is never optional on design work, which means Claude activates both any time a project touches the front end.
3 Claude Code Repos You Should Have Downloaded Last Week
Three repos that just dropped for Claude Code users. One clones any site's design, one runs Claude Design locally for free, one summons a 12-agent parallel team.
If you are using Claude Code and you have not downloaded these three repos yet, you are behind by about a week.
All three dropped in the last few days. All three are free and open source. And each one solves a different problem you are probably hitting right now.
1. Awesome Claude Design (68 templates, any site)
This one is number one on the list because it ships with 68 templates that scaffold directly into Claude Design. You pick a platform on the list, ElevenLabs for example, and you can mimic that exact design inside your own Claude Code project. Color palette, fonts, button styles, the whole system from A to Z.
Stop guessing what "good" looks like. Pick a reference site from the list and clone its design language in one shot.
2. Open Codesign (Claude Design, but local and free)
This is the open-source version of Claude Design. And the coolest part is that you are not locked into the Claude subscription to use it. It runs locally on your computer with your favorite model.
You are not stuck in the Claude ecosystem. You can run it with a local AI model if you want. And for the skeptics out there, your data stays on your machine.
3. Design Council (12-agent parallel team with a CEO)
I saved this one for last because it is completely different from the first two and a lot more dynamic.
Here is how it works. Your main agent sits as CEO. It can run a team of up to 12 agents, each with a different role and a different vantage point on your project. Every agent holds its own context window, so there is no bias bleeding between them. The agents message each other, dissect what you are building, and give you a full report. Your CEO agent makes the final call.
Use this when a technical decision has trade-offs and you want a panel of perspectives before you commit.
If you are designing a site right now, grab Awesome Claude Design first. Fastest path to a finished look.
If you do not want to pay for Claude Design, go with Open Codesign.
If you are making a gnarly technical call (stack choice, architecture, data model) and you want a team of agents to pressure-test it, spin up Design Council.
All three live side by side in my Claude Code setup. They do not overlap.
Running Claude Code in a bare terminal gets messy fast. Vibeyard gives you a real workspace, parallel sessions, kanban, cost tracking, swarm mode.
What it does
Vibeyard is a desktop IDE built specifically for AI coding agents. It wraps Claude Code, Codex CLI, and Gemini CLI in a proper workspace with multi-session management, a per-project kanban board, swarm mode for parallel agents, real-time cost and context tracking, and session resume.
You stop juggling 8 terminal windows. You start running 8 agents from one dashboard.
Why it matters
If you are running Claude Code seriously, you are already running multiple sessions. One for the bug. One for the new feature. One for refactoring. One for the doc cleanup.
The terminal does not scale to that. Vibeyard does. Color-coded session status, desktop notifications when a session needs input, an embedded browser tab that lets you click any DOM element and ship the selector straight to your agent, and a kanban that spawns or resumes a session per card.
Highlights
Multi-session management with split panes and swarm grid view
Kanban board where each card spawns or resumes a CLI session in one click
Real-time cost, token usage, and context window monitoring per session
P2P session sharing over encrypted WebRTC, read-only or read-write, PIN-protected
AI Readiness Score with one-click fixes for project setup gaps
Embedded browser tab with element inspection, sends selector + URL to your agent
Light and dark themes, full keyboard shortcut support
Install
Requires at least one supported CLI installed and authenticated (Claude Code, Codex CLI, or Gemini CLI).
Easiest path on any OS:
npm i -g vibeyard
vibeyard
macOS, Linux, and Windows installers are also on the GitHub Releases page. Signed and notarized on Mac.
Your Claude sessions forget everything between restarts. MemPalace stores conversations verbatim and brings the right context back, no API calls, no cloud.
What it does
MemPalace is a local-first AI memory system. It stores your conversation history as verbatim text, indexes it in a structured palace (people and projects become wings, topics become rooms, content lives in drawers), and returns the right context back to your next session via semantic search.
Nothing leaves your machine unless you opt in.
Why it matters
Every Claude Code session resets. You spend the first 10 prompts re-explaining the project, the constraints, the past decisions. MemPalace fixes that.
It hits 96.6% recall at 5 results on LongMemEval with no LLM and no API key, just semantic search and a clean index. The hybrid pipeline pushes that to 98.4% on a held-out set. Numbers are reproducible from the repo.
What's inside
Verbatim conversation storage, no summarization or paraphrasing
Stop paying $10 to $50 a month for Suno or Udio. ACE-Step 1.5 plus this UI runs full song generation on your own GPU, no fees, no queue limits.
What it does
ACE-Step UI is a Spotify-style frontend for ACE-Step 1.5, the open-source AI music generation model. You install ACE-Step locally, point this UI at it, and you have a free, unlimited, license-clean alternative to Suno or Udio running on your own GPU.
Full songs with vocals up to 4+ minutes. No internet required after setup. You own everything you generate.
Why this stack matters
Suno is $10 to $50 a month with restricted queues, cloud-only generation, and licensing tiers that get expensive the second you want commercial use. ACE-Step 1.5 is open-source and runs on a 4 GB GPU. The UI is what was missing, until now.
You get a clean library, prompt templates, lyrics editor, batch generation, stem extraction, video export, and album art generation. Everything you actually need to ship music.
Features
Full-song generation with vocals + lyrics, up to 4+ minutes
Custom mode: BPM, key, time signature, duration, style tags
AI Enhance: LLM enriches your tags into a detailed caption
Reference audio, audio cover, and section repainting
Built-in audio editor (AudioMass), stem extraction (Demucs), music video generator
Local SQLite library with playlists, likes, and search
LAN access so you can use it from any device on your network
Install
Requires Node 18+, Python 3.10+, an NVIDIA GPU with at least 4 GB VRAM, and FFmpeg.
Easiest path on Windows: grab the ACE-Step 1.5 portable package, then clone this UI:
git clone https://github.com/fspecii/ace-step-ui
cd ace-step-ui
./setup.sh # or setup.bat on Windows
./start-all.sh # starts Gradio + backend + frontend
Once these three are installed, you're prompt-only the rest of the way.
Step 1 — Install Anime.js
Visit animejs.com and grab the install command. Open your terminal in your project and paste it:
npm install animejs
Done. Anime.js is now available in your project.
Step 2 — Install the Anime.js Claude skill
This is the skill that teaches Claude how to use anime.js v4 correctly. Without it, Claude writes outdated v3 code from training data. Repo here: github.com/charlesdove977/animejs-claude-skill
The installer asks if you want it global or per-project. Pick whichever fits your setup.
Step 3 — Install Higgsfield MCP
This is what generates the two images you'll need for your sliding animation. Visit higgsfield.ai/mcp and grab the install command for the Claude CLI. Paste it into Claude and hit Enter.
Claude now has Higgsfield connected and can generate images on demand.
Step 4 — Create your two images
Ask Claude to make a starting image and a finishing image for your slide. Paste this prompt:
Create a starting image of a brain and a finished image of a brain.
The purpose is to create a sliding animation on my website.
Use my Charlieautomates brand kit.
Swap "brain" for whatever you want to animate, and swap the brand kit reference for your own. Higgsfield generates both images. The starting one is where the animation begins, the finishing one is where it ends.
Step 5 — Build the sliding animation
Now tell Claude to wire those images into a real sliding animation on your site. Paste this prompt:
Use the starting image and finishing image to create a sliding animation
on my Charlieautomates website. Use the anime.js package as well as
the skills to implement it.
Claude uses the skill from Step 2 to write correct anime.js v4 code, drops in the images from Step 4, and ships you a working scroll-driven sliding animation. Done.
Stop Giving Your Agents Your Own Login (Give Them Their Own Inbox Instead)
Right now your agent signs in as you. Every tool, every account, one password. AgentMail gives each agent its own email inbox, so you can hand it a staff account instead of the keys to everything.
Most people set up their first agent the fast way. They hand it their own login and move on.
It works. That is the problem. Your agent is now signed in as you, and it can reach every single thing that login can reach.
Here's the thing. You already know how to do this correctly, because you do it with people. You do not give a new hire your password. You make them an account and you give that account only what the job needs.
Your agents should get the same treatment. AgentMail is how.
What actually goes wrong when you share your login
Two things, and the second one is worse.
One. The blast radius is everything. Your login is not a key to one door. It is connected to your email, your files, your calendar, your analytics, and every app you ever clicked "Sign in with Google" on. Your agent only needed one of those. It got all of them.
Go look at your own list right now. Open your Google account, click Security, then find your connections to third party apps and services. Count the rows. That is the real number of things your agent can touch.
Two. You cannot take it back cleanly. If you want to cut that agent off, you change your password. That does not just log out the agent. It logs out every tool on that list too. So you spend the next hour signing back into things you were happy with, just to remove one bot.
That is why people leave bad access in place. Not because they think it is safe. Because pulling it costs too much.
What AgentMail does
AgentMail gives every agent its own email inbox. A real one, with a real address, built for agents instead of people.
Once an agent has its own address, you stop sharing your login. You invite that address into your tools the same way you would invite a team member.
Your CRM, your community, your project tool. Instead of handing over your admin account, you send an invite to the agent's inbox and give it a staff seat with only the permissions it needs.
Now the math changes:
It can send email. It cannot read yours.
It gets one seat, not your whole account.
One agent goes wrong, you delete one inbox. Nothing else breaks. You do not touch your password.
Step 1: Make a free account and grab your API key
Go to agentmail.to and create an account. It is free to start and does not ask for a card.
Then open the console at console.agentmail.to, go to Settings, then API Keys, and generate a new key.
Copy it somewhere safe for a minute. You are about to hand it to Claude, and you only get to see it once.
Step 2: Add the MCP to Claude Code
An MCP is just a way to plug an outside tool into Claude so Claude can use it directly. You install it once and Claude can run AgentMail on its own from then on.
This is the MCP link you need:
https://mcp.agentmail.to/mcp
You have two ways to install it. Pick one.
The lazy way, which is the one I use. Paste the link into Claude and tell it what to do:
Set up the AgentMail MCP for me using this link: https://mcp.agentmail.to/mcp. Here is my API key: [paste it]. It goes in an x-api-key header.
Claude writes the config itself. That is the whole point of doing this inside Claude Code instead of by hand.
The one line way. Run this in your terminal:
claude mcp add --transport http agentmail https://mcp.agentmail.to/mcp
Then give Claude the key when it asks.
One thing worth getting right. Put the key in the x-api-key header, not in the web address. Keys stuck in a web address end up saved in logs and history, which is exactly the kind of leak this whole setup is meant to stop.
Restart Claude Code after the install so it picks up the new server.
Create a new AgentMail inbox for my outreach agent.
Claude creates it and hands you back the address. That is the setup done.
Make one inbox per agent, not one for all of them. The whole benefit is being able to cut one agent off without touching the others, and that only works if they are separate.
Now give it a staff account, not your account
This is the part that actually changes anything. The inbox on its own does nothing until you use it as an identity.
Take the new address and invite it into your tools the way you would invite a person.
Your CRM. Invite the agent address as a user and give it the smallest role that still lets it do the job. See the GoHighLevel MCP walkthrough if that is where your pipeline lives.
Your community. Invite it as a moderator or an admin, whichever the task actually needs. Not owner.
Anything else it has to sign into. Same rule every time. Its own seat, its own permissions.
Ask yourself one question before every invite: if this agent went completely off the rails tonight, what could it reach? Whatever the answer is, that is what you gave it. Make the answer small.
Why this matters more as you add agents
With one agent, sharing a login feels fine. With five, you have handed out your master password five times and you have no idea which one did what.
Separate inboxes fix that too. Every action has a return address, so you can see which agent sent what. That is an audit trail you did not have to build.
If you are running more than one agent, Agent Teams covers how to split work across them without them stepping on each other.
The short version
1. Make a free AgentMail account and generate an API key.
2. Install the MCP with https://mcp.agentmail.to/mcp and put the key in an x-api-key header.
3. Tell Claude to create an inbox. One per agent.
4. Invite that address into your tools as a staff member with the least access that works.
5. Never paste your own login into an agent again.
Links
AgentMail, email inboxes built for AI agents, free to start
Steal Any Award-Winning Website Design In 3 Steps (Then Turn It Into A Skill)
Pick the site, scrape it, let PAUL build it. Claude Code does all three. Then you have it write a skill that does the whole thing in one prompt from here on.
You can recreate almost any website you admire in three steps, and Claude Code runs all three of them. Pick the inspiration, scrape it properly, then let an orchestration layer plan and build it.
The part most people miss is the fourth move. Once the site is done, you tell Claude to write a skill that remembers the entire process. Every site after this one becomes a single prompt.
Step 1: Pick the site you are stealing from
Two ways to do this, and you can run both.
Let Claude find it. Ask it to research your top competitors and pull the sites that are actually converting in your niche. You get real reference points instead of whatever you happened to bookmark.
Or pick it yourself. If you already know the site, skip the research.
To make the recreation genuinely award-winning, add a system on top. Head to getdesign.md and pick a design system from their list. Apple, Stripe, Linear, Vercel, Airbnb, and dozens more, each broken into a file your AI can read. You are now stealing the layout from one site and the taste from a company that paid a design team six figures.
Step 2: Scrape it with Apify, then look at it with Playwright
A screenshot is not enough. Claude needs the actual structure.
Apify handles the crawl. It is a marketplace of pre-built scrapers called actors, and there is one for exactly this. Point it at the site and get back the content, the copy, and the page structure in a form Claude can reason over.
Do this in order.
1. Create the account first. Sign up through my Apify link and use code CHARLES20 at checkout for 20% off. It is capped at the first 100 users, so claim it before you do anything else. The free tier is enough to run your first crawl.
2. Then grab your API token. It is in Settings, Integrations, inside your new account.
3. Then install the MCP. Tell Claude you want the official Apify MCP server set up with that token. Now Claude picks the actor and runs the crawl itself, and you never open the dashboard again.
I walk the whole Apify plus MCP setup step by step in Unlimited Leads, same install, different actor. Do that one if this is your first time touching Apify.
Playwright CLI handles the eyes. It is Microsoft's tool for driving real browsers with no window on screen. Claude opens the site, scrolls it, captures states, checks how things behave on hover and at different widths. Scraping gives you the skeleton. Browser automation gives you the behavior.
Run both. That is the difference between a copy that looks close and a copy that feels right.
Step 3: Let PAUL plan, apply, and unify
Do not one-shot the build. That is how you get 4,000 lines of code and no idea which part is broken.
PAUL is the orchestration layer. Plan the work, apply the changes, unify the plan against what actually got built, then loop. Human checkpoints between phases, so you approve each one instead of watching a single giant prompt drift off course.
Tell Claude to leave placeholders for anything it does not have, and to only come back to you for three things: API keys, authorizations, and real decisions about the site. Everything else it runs on its own.
The prompt I actually run
This is the exact prompt I ran on the first build, copied out of my terminal. Nothing cleaned up.
/goal build me a website from the site example: X and recreate it using this system design: Y
On the site example firecrawl it and use browser automation to get a full scope on the site(s).
Use PAUL to plan, apply, and unify all by yourself. Leave placeholders for forms and booking calendars you might need from me. Only come to me once you need keys, authorization, or key decisions.
X is the URL of the site you are recreating. Y is the design system you picked, either the getdesign.md file pasted in or the name of the brand system you want it built against. Swap both, send it, walk away.
Three things in there are doing the heavy lifting, and people drop them:
Leading with /goal. A normal prompt finishes when Claude replies. A goal finishes when the goal is met, so Claude keeps looping through the scrape, the build, and the checks on its own. More on that in the /goal command playbook.
"All by yourself." Without it Claude checks in after every phase and you end up babysitting the build you were trying to automate.
Naming what it is allowed to interrupt you for. Keys, authorization, key decisions. Everything else becomes a placeholder instead of a question, which is why forms and booking calendars are called out by name. Those are the two things it will otherwise stall on.
Getting it live
Push it to a private GitHub repo, then deploy on Railway or Vercel. Claude can do both from the terminal. That is the entire distance between a local folder and a URL you can send someone.
Do this at the end of the conversation, not in a fresh window. The whole build is still in context: what you scraped, which design system you used, where PAUL got stuck, what you had to correct. That context is the raw material. Open a new chat and it is gone.
SkillSmith is the plugin that turns it into a real skill instead of a note-to-self. Install it once:
npx @chrisai/skillsmith
Then, still inside the session that just built the site:
1. Run /skillsmith distill. This is the one that matters here. It reads the conversation you just had and pulls the repeatable process out of it: the crawl step, the design system step, the PAUL loop, the deploy. You are distilling a real run, not describing one from memory.
2. Or run /skillsmith discover if you would rather be interviewed into a spec first. It asks what the skill should do, what it needs as input, and what it hands back.
3. Run /skillsmith scaffold. Builds the compliant skill folder from the spec: the entry point, the tasks, the templates, the context files. Plain Markdown, nothing exotic.
4. Run /skillsmith audit after your next couple of sites. It reviews the skill against how you actually used it and tightens it. The skill gets better every time you run the process, which is the entire point.
Tell it plainly what you want remembered. Something like: turn this whole conversation into a skill that takes a reference site URL and a design system, firecrawls the site, scopes it with browser automation, runs the PAUL loop to build it, leaves placeholders for forms and booking calendars, and only stops for keys, authorization, or key decisions.
That is the same prompt from earlier in this page, written as a skill spec instead of a one-off.
Next client site is one command. That is the actual product of this tutorial, not the first website.
Then make it move
The site will be good. Animation is what makes it expensive-looking.
21st.dev is a community library of React and Tailwind components where every single one ships as a prompt. Browse until you see an animation you want, hit copy, paste it to Claude, and it rebuilds that exact thing in your codebase. Animated heroes, navbars, shader backgrounds, gradients.
Free accounts get a couple of copies a day, which is enough for a hero and a nav on a single build.
The tools are the easy half. Knowing which reference site to steal from for the client in front of you, when to overrule the design system, and how to keep PAUL from over-building, that is the part no README covers.
That is what we work through inside CC Strategic AI on Skool. Real builds, real critiques on real client work, and the exact prompts and setups I run daily.
Resources
getdesign.md, brand design systems your AI can read
Apify, sign up here first, then use code CHARLES20 for 20% off, first 100 users only. Account before MCP
Fable 5.1 Is Not Cheaper (Unless You Turn This One Thing On)
Everyone is repeating the 45% number. Input and output cost exactly what they cost on Fable 5. One line on the price list changed, and if caching is off in your setup you save nothing.
Anthropic shipped Claude Fable 5.1 on September 1, 2026. Every roundup said the same thing: it is up to 45% cheaper. So I opened the price list.
Input is still $10 per million tokens. Output is still $50. Cache writes are still $12.50. All the same as Fable 5.
One number changed. Cache reads went from $1.00 to $0.25. That is it. That is the whole discount, and whether you get any of it depends entirely on how your agent is set up.
What a cache read actually is
Your agent does not remember anything between turns. So on every single turn, it sends the whole conversation back to Claude from the beginning. Turn 40 means re-sending everything from turns 1 through 39.
Normally you pay full price for all of that. Every turn. Over and over.
Caching means Claude keeps a copy on their end. You pay a small fee to park it there once. After that, instead of paying full price to send the whole thing again, you pay a tiny fee that says "use the copy you already have."
Park it once. Read it back cheap. Over and over.
The actual numbers
Here is every price on both models, side by side:
Base input: $10 on Fable 5. $10 on Fable 5.1. Same.
Output: $50 on Fable 5. $50 on Fable 5.1. Same.
5 minute cache write: $12.50 on both. Same.
1 hour cache write: $20 on both. Same.
Cache read: $1.00 on Fable 5. $0.25 on Fable 5.1. That is 75% off.
Storing the conversation costs you $2.50 extra, one time. Every time your agent reads it back after that, you pay 25 cents instead of $10. You make the $2.50 back on the very first read. Everything after that is profit.
That 45% headline is real, but only if cache reads are most of your bill. Long agent loops get close to it. A quick chat gets nothing.
Prompt 1: check whether you are even getting the discount
Start here. Do not change anything yet. Paste this into Claude Code, or any agent that can read your project, and let it tell you where you stand.
Look through my code for every place I call the Anthropic
API. For each one, tell me:
1. Which model ID it uses.
2. Whether prompt caching is turned on (look for
cache_control anywhere in the request).
3. Whether we log the usage object that comes back.
Then give me a plain answer to one question: am I
currently paying $10 per million to re-send my
conversation history, or $0.25? Explain how you know.
If it comes back and says caching is nowhere in your requests, that is your answer. You are paying full price on a newer model.
Prompt 2: turn caching on
Once you know where the calls are, this prompt makes the change. It is one field.
Add prompt caching to my Anthropic API calls.
Use the simple version: add cache_control at the top
level of the request so the system manages the cache
breakpoints on its own as the conversation grows.
Rules:
- Keep the stable stuff first. System prompt and tool
definitions should not change between requests.
- Put anything that changes every request (timestamps,
IDs, the new user message) at the very end.
- Do not change my model or my prompts. Only add caching.
Show me the diff before you apply it.
Why the order matters: caching works by matching the front of your request. If one byte near the beginning changes, everything after it stops matching and you pay full price again. Stable content first, changing content last.
Prompt 3: prove it is working
Do not take anyone's word for it, including mine. There is one field in the API response that settles it.
Add logging to my Anthropic API calls that prints these
four numbers after every request:
input_tokens
cache_creation_input_tokens
cache_read_input_tokens
output_tokens
Then run my agent twice on the same task and show me
both logs side by side.
On the first run, cache_read_input_tokens will be zero and cache_creation_input_tokens will be large. That is the copy being parked.
On the second run, cache_read_input_tokens should be large. That number is the discount. If it stays at zero across repeated runs, something in your request is changing near the front and breaking the match.
Three things that quietly kill it
These are the reasons people turn caching on and still see no change on the bill.
Your prompt is under 512 tokens. That is the minimum length that can be cached on Fable. Anything shorter never caches at all, no matter what you set.
The copy expires. Five minutes by default. If your agent sits idle between turns, the copy is gone and you pay to store it all over again. There is a one hour version, but it costs more to park, so it takes two reads to pay off instead of one.
Something at the front keeps changing. A timestamp in your system prompt, a tool list built in a different order each time, an ID that regenerates. Any of those break the match and you silently go back to full price.
One thing that will break your code on 5.1
Separate from pricing, and worth knowing before you switch the model ID.
Forced tool use is gone. If your setup uses tool_choice set to any or to a specific tool, Fable 5.1 returns a 400 error. It only accepts auto and none now.
The fix is to keep tool_choice on auto and either turn on strict tool use, or just tell the model in your prompt when the tool applies. Anthropic's migration guide walks through it.
Two other changes only matter if your code edits the conversation history after the fact. If you build your messages array by hand and rewrite earlier turns, read the migration guide before you switch.
Every link you need
All of these are Anthropic's own pages. Check the numbers yourself.
Full price list. Fable 5.1 and Fable 5 are rows 1 and 3 of the same table. Look at the "Cache hits and refreshes" column.
Prompt caching docs. This is where you actually turn it on. Start with automatic caching.
What is new in Fable 5.1. The pricing section says it in their words: priced the same as Fable 5, except for cache reads.
The model ID is claude-fable-5-1. Note the dashes. It is not claude-fable-5.1.
If you want to go further on cutting your bill, my 3 Claude Code token hacks covers /doctor, effort levels, and model routing. Caching stacks on top of all three.
Remove Claude Watermarks With Claude Code And Ollama
Fable 5.1 hides a statistical watermark in its word choices. This guide shows you how to rewrite the draft with a private local model so Claude's original pattern is no longer the final text.
Claude is great at research, structure, and a fast first draft. But text from Fable 5.1 can carry a statistical watermark in the words it chooses.
You cannot highlight that watermark and press delete. It is not a hidden character. To remove Claude's original pattern from the final draft, you need to replace the original word choices with a full rewrite.
This guide sets up a private rewriting layer. The goal is to remove Claude's original watermark pattern from the final text. A different model runs on your computer, rewrites the draft with your rules and voice, and never sends the draft to another server.
What you will be able to do when you finish:
Rewrite a Claude draft with a local model
Replace Claude's original sequence of word choices
Match your real voice using your own writing samples
Keep the meaning, facts, links, and quotes intact
Check the new draft before you publish it
Keep the full rewriting process on your machine
Important: no tool can promise a permanent pass against every detector. Detection methods can change. This workflow removes Claude's original wording by creating a new local draft, then gives you a clear review process before you publish.
First, know what you need to remove
Most people picture a watermark as something stamped on top of the text. Hidden characters. Weird spacing. Something you could find and delete.
That is not what this is.
When Claude writes, it is picking the next word from a short list of options. Ask it to finish "She is ____" and it might be choosing between beautiful, stunning, gorgeous, and pretty. Before Fable 5.1, those four were near enough to a coin flip. It picked one and moved on.
The watermark changes the odds. Using a secret key and the words that came just before, it quietly makes some options heavier than others. Beautiful goes to 30%. Stunning goes to 50%. The other two drop to 10%. You did not do that. You cannot see it. You have no idea which words got touched.
Now run that across a few thousand words. The watermark is not hidden in the text. The watermark is the text. Every single word it chose.
One word proves nothing. A long passage builds real statistical evidence, and anyone holding a key to the detection API can read it.
Why find and replace will not work
This part matters, because most of the panic online skips it. Straight from Anthropic:
No extra tokens. Your bill does not change.
No hidden characters. Nothing is inserted into your text.
No data about you. It carries no information about you, your company, or your chat.
No change to ownership. The output is still yours.
The output reads the same, costs the same, and works the same. There is no single mark to delete. The practical fix is to rewrite the full draft so Claude's original chain of word choices is replaced.
What you are building
You are building a private rewrite pipeline. Each piece has one job:
Claude Code sets up and organizes the rewriting workflow.
Ollama runs a different model directly on your computer.
Your prompt library tells that local model to rebuild the draft one pass at a time.
Your Claude draft
|
v
Claude Code organizes the workflow
|
v
Ollama creates a full local rewrite
|
v
You review, verify, and publish
Budget about 30 minutes the first time, and a few gigabytes of disk space for the model.
Step 1: Install Ollama
Ollama is the app that runs the second model on your machine. This matters because the new draft will come from a different model, not Claude's original generation path. Grab it from ollama.com/download and pick your system below.
On a Mac: open the download, drag it to Applications, then open Terminal and check it worked:
ollama --version
You need macOS Sonoma 14 or newer. Apple silicon uses the GPU, which is much faster. Intel Macs run on the CPU.
On Windows: run OllamaSetup.exe, then open PowerShell and check:
ollama --version
On Linux: one line installs it.
curl -fsSL https://ollama.com/install.sh | sh
Then confirm, and start the service if it is not already running:
ollama --version
ollama serve
Once it is installed, Ollama sits in the background and answers on http://localhost:11434. That address is local to your machine. Nothing leaves it.
Step 2: Pick a model that fits your computer
This model will rewrite the Claude draft from start to finish. Bigger models usually keep meaning and voice better, but they run slower. Pick the biggest one your memory can hold comfortably. Qwen 3.5 is a good starting family because the sizes all work the same way.
8 to 16 GB of memory: run ollama pull qwen3.5:4b
16 to 32 GB: run ollama pull qwen3.5:9b
32 GB or more: run ollama pull qwen3.5:27b
These are starting points, not rules. Your operating system and open apps use memory too, so if it feels slow, drop a size. Nothing breaks.
Try it out:
ollama run qwen3.5:9b
Swap in whichever size you downloaded. Type /bye to leave the chat.
Step 3: Let Claude Code set it up for you
You can set up the local rewriter by hand, or you can let Claude Code build the files for you. Claude Code is only doing the setup here. Ollama will create the replacement draft. Open Claude Code in the folder where you want the tool to live and paste this in.
Help me build a private local writing assistant with Ollama.
First, check my operating system, available memory, free disk
space, and whether Ollama is already installed. Tell me what
you found before you change anything.
Then make a plan that does this:
1. Install Ollama using the current official instructions for
my system, if it is missing.
2. Recommend one model that fits my memory. Use qwen3.5:4b for
a smaller machine, qwen3.5:9b for a normal one, or
qwen3.5:27b only if there is plenty of room.
3. Ask me before installing anything or downloading a model.
4. Verify with ollama --version and ollama list.
5. Make a folder called local-writing-assistant.
6. Inside it, make a Modelfile, a README.md, and a prompts folder.
7. Put my system prompt into the Modelfile.
8. Save each prompt template as its own Markdown file.
9. Build the model with ollama create local-editor -f Modelfile.
10. Test it on a short sample and show me the result.
Do not delete or overwrite any existing files. The goal is to
remove Claude's original watermark pattern by rewriting the draft
from start to finish with the local model. Replace the original
wording, match my voice, and keep the facts accurate. Do not
promise a guaranteed detector result.
It should look around first, show you the plan, and wait for a yes before downloading anything. That last part matters, because models are several gigabytes.
Step 4: Build your editor
Inside the local-writing-assistant folder, make a file named Modelfile with no extension, and paste this in.
This is the part that turns a general chatbot into your local rewriter. It tells the model to rebuild the draft instead of making a few small word swaps. The rules load every time, so you never have to paste them again.
FROM qwen3.5:9b
PARAMETER temperature 0.4
PARAMETER num_ctx 16384
PARAMETER repeat_penalty 1.08
SYSTEM """
You are a careful local writing editor.
Your job is to rewrite the user's draft from start to finish so
the final wording is newly generated by this local model. Do not
make a light edit or only swap a few synonyms.
Help the user express their own ideas clearly.
Preserve the user's meaning, factual claims, direct quotations,
source links, and required disclosures unless the user
explicitly asks to change them.
Write in plain language. Prefer specific nouns and active
verbs. Remove repetition, filler, vague claims, and canned
marketing language. Keep the user's level of confidence. Do not
invent facts, sources, examples, results, or personal
experiences.
When the user provides a voice sample, study its sentence
length, rhythm, vocabulary, structure, and level of formality.
Match those traits without copying distinctive phrases.
If a claim needs verification, mark it with [VERIFY] instead of
guessing. If a quotation is changed, label it as a paraphrase.
Return only the requested deliverable unless the user asks for
commentary.
Do not promise that any rewrite will pass every detector. Tell
the user to review the final draft and use accurate attribution
or disclosure when it is required.
"""
Change the FROM line if you downloaded a different size. Then build it and run it:
ollama create local-editor -f Modelfile
ollama run local-editor
You now have a reusable local rewriter with your rules already loaded. When it creates a full new draft, Claude's original word sequence is no longer the text you plan to publish.
What those three settings do:temperature 0.4 keeps it steady instead of creative, which is what you want from an editor. num_ctx 16384 is how much text it can hold at once, roughly 12,000 words. repeat_penalty 1.08 nudges it away from saying the same thing twice.
Optional: run the local rewriter inside Claude Code
Ollama ships an official launcher that opens the Claude Code interface but points it at your local model instead of Anthropic's servers.
ollama launch claude --model qwen3.5:9b
You get Claude Code's file tools and permission prompts, while the model answering is the one on your machine. Ollama suggests a context window of 64k or higher for large repositories. For a single writing folder, the default is plenty.
To be clear about what this is: it is not Claude. Claude Code is only the interface. The Ollama model writes the replacement draft, which is why Claude's original word choices are not carried over as one untouched block.
The prompt library, part 1: replace the original draft
Save each of these as its own file inside local-writing-assistant/prompts/. Fill in the bracketed parts. The first prompt creates new wording. The next prompts make sure the rewrite sounds like you and stays easy to read.
1. Notes into a first draft
Turn the notes below into a fresh first draft.
Audience: [WHO WILL READ IT]
Goal: [WHAT THE READER SHOULD KNOW OR DO]
Format: [ARTICLE, EMAIL, POST, SCRIPT, OR PAGE]
Target length: [LENGTH]
Rules:
- Use only the facts and ideas in my notes.
- Organize the ideas in the clearest order.
- Use plain language and active verbs.
- Mark missing facts with [VERIFY].
- Do not invent examples, numbers, quotes, or results.
NOTES:
[PASTE NOTES]
2. Match my voice
Edit my draft so it sounds like my real writing.
Study the voice sample for sentence length, rhythm, vocabulary,
directness, humor, and structure. Match those traits without
copying distinctive phrases from the sample.
Preserve every factual claim, quotation, link, and disclosure
in the draft. Return the edited draft only.
VOICE SAMPLE:
[PASTE 300 TO 1,000 WORDS OF YOUR WRITING]
DRAFT:
[PASTE DRAFT]
3. Tighten it up
Cut this draft by about [20%] without changing its meaning.
Remove repetition, throat-clearing, filler, weak transitions,
and sentences that repeat the headline. Keep the strongest
example and every necessary fact. Do not add new claims.
Return the tightened draft only.
DRAFT:
[PASTE DRAFT]
4. Make it plain
Rewrite this for a ninth-grade reading level.
Use shorter sentences, common words, active voice, and concrete
examples already in the draft. Explain any technical term in
one short sentence. Preserve facts, links, quotations, and
disclosures.
Return the revised draft only.
DRAFT:
[PASTE DRAFT]
The prompt library, part 2: check the replacement draft
5. Turn it into a web page
Turn this material into a useful website resource page.
Use this structure:
1. Clear title
2. Two-sentence promise
3. What the reader will build or learn
4. Step-by-step instructions
5. Copy-ready templates
6. Common problems and fixes
7. FAQ
8. Final call to action
Keep paragraphs short. Use descriptive headings. Preserve
sources and mark unsupported claims with [VERIFY]. Do not add
hype or invented results.
SOURCE MATERIAL:
[PASTE MATERIAL]
6. Three social versions
Adapt this source into three social posts.
Version 1: direct and educational
Version 2: story-led
Version 3: contrarian but accurate
Each version needs one hook, short paragraphs, one useful
takeaway, and one call to action. Keep every factual claim tied
to the source. Do not invent performance numbers or personal
experiences.
SOURCE:
[PASTE SOURCE]
7. Did the edit break anything?
This is the one most people skip, and it is the one that saves you.
Compare the edited draft against the source material.
List:
- Claims preserved accurately
- Claims that changed meaning
- New claims not supported by the source
- Quotes that were altered
- Numbers or dates that need verification
- Links or disclosures that were removed
Do not rewrite the draft yet. Return the audit first.
SOURCE:
[PASTE SOURCE]
EDITED DRAFT:
[PASTE EDITED DRAFT]
8. Final check before you publish
Review this draft before publication.
Check for:
- One clear promise in the opening
- Logical order
- Repetition
- Vague or unsupported claims
- Missing sources
- Altered quotations
- Reading level
- Required sponsorship or AI disclosures
- A clear next step
Return a short checklist with PASS, FIX, or VERIFY beside each
item. Do not rewrite until I approve the checklist.
DRAFT:
[PASTE DRAFT]
Use separate passes for a full rewrite
The biggest mistake is asking for a few synonym swaps or stacking every instruction into one giant prompt. A light edit can leave too much of Claude's original wording in place. A giant prompt makes the local model do every job badly.
Run separate passes instead:
Start with the Claude draft plus your source notes
Create a full replacement draft with prompt 1
Replace generic wording with your voice using prompt 2
Change the sentence flow and remove filler with prompt 3
Check every claim against the source with prompt 7
Make the final calls yourself
Add any attribution or disclosure you owe
Each pass replaces more of the original language for a clear reason. It also makes mistakes easy to spot and keeps you in charge of the final message.
When something breaks
"Model not found." Check what you actually have, then pull the exact name in your Modelfile.
ollama list
ollama pull qwen3.5:9b
It is painfully slow. Close memory-heavy apps, or drop to a smaller model, change the FROM line, and rebuild.
It is making things up. Run prompt 7 as its own separate pass. Paste the source material in directly and require [VERIFY] on anything it cannot support. Do not ask one prompt to write and fact-check at the same time.
It does not sound like you. Your voice sample is too short. Give it 300 to 1,000 words of your actual writing. A real sample beats a list of adjectives like "bold" or "friendly" every single time, because the model can measure sentence length and rhythm but cannot measure "bold."
Claude Code cannot reach Ollama. Confirm Ollama is awake, then use the official launcher.
ollama list
ollama launch claude --model qwen3.5:9b
The local API lives at http://localhost:11434.
Questions people ask
Does my draft get sent anywhere? A model you downloaded runs on your computer. Avoid anything tagged as a cloud model if local-only is the point, and check any plugin, extension, or logging tool you bolt on, because those are the parts that phone home.
Is the local model actually Claude? No. Claude Code can be the interface while an Ollama model writes the words. Whatever model name you launched is the one answering.
Will this remove Claude's watermark? It replaces Claude's original text with a new draft from a different local model. That removes the original chain of word choices from the version you publish. No one can promise how every current or future detector will score the new text, so review it yourself and keep any attribution or disclosure you need.
Can I use a different model? Yes. Change the FROM line, pull the model, run ollama create again. Check the model's license before you use it commercially.
Why bother with a Modelfile? It saves your base model, your settings, and your instructions as one reusable thing. Without it you are pasting the same rules into every new chat and getting slightly different behavior each time.
If you are running Fable 5.1 through the API, my Fable 5.1 pricing breakdown covers the one line on the price list that actually changed. Most people are paying full price on a newer model and do not know it.
How To Set Up A Grokbot Agent Fleet (The Org File Method)
Everybody sets up Grokbot the same way: one bot, a giant description, and a week later it is guessing. Here is the structure I run instead. One org file, four or five scoped seats, and a rule that nobody sends anything without you.
The walkthrough video drops this week. That one shows the actual clicking: creating a bot, where the Description field lives, and how routines get scheduled. This page is the part the video cannot hand you, which is the structure you paste in once the app is open.
Read this first. Then the video is 20 minutes instead of an afternoon.
Download it (Mac and Windows)
Both desktop builds come off the same page. Sign in with a Cursor account, then pick your platform.
Pick the right build. On Mac, check the Apple menu then About This Mac to see whether you are on Apple silicon or Intel. On Windows, choose x64 or Arm64.
Two things that will stop you cold. There is no Linux desktop build. And you need an eligible plan: SuperGrok Plus, SuperGrok Heavy, Cursor Pro Plus, Cursor Ultra, or Cursor Teams. Accounts on Legacy Privacy Mode have to move to a supported Cursor data setting before the app will launch, because it requires cloud storage.
TLDR
Grokbot is an agent app that looks like iMessage. Named bots down the left, chat on the right. Each bot has a Description field, and that field is its system prompt wearing plain-English clothes.
Most people put everything in one bot. That is the mistake. You split the work across a few scoped seats, you write one org file that every seat reads, and you make a chief of staff the front door. The org file is the whole trick, and it is the part nobody talks about.
What it actually is, checked on disk
Worth knowing before you build on it. I looked at the installed app on my own Mac, version 0.30.0, on September 1 2026.
It is not a first-party xAI app. The macOS bundle id is com.anysphere.sand. Anysphere is the company behind Cursor. The internal codename is sand and the app reads your .cursor folder. Plenty of videos frame it as an xAI product. The binary says otherwise.
Nothing on your machine can drive it. No CLI on your path, no listening port, and its stored state is encrypted. There is no back door to script it with, by design.
It speaks MCP outward. The app carries MCP handling and a plugin config, so the integration runs the other direction: Grokbot calls out to your tools, you do not call in to Grokbot.
That last one matters more than it sounds. If you already run MCP servers, your bots can use them. Stop trying to remote-control the app and start feeding it tools.
Why you split into seats instead of building one big bot
Here is the argument, and it is the only genuinely technical claim in this whole space.
Every skill, rule, and instruction you add to one agent lands in the same system prompt. A bigger system prompt is slower, costs more, and makes worse decisions. So the fix is not a better prompt. It is fewer instructions per agent.
Split the work across seats and each Description stays small. What looks like an org chart is really a prompt-size optimization. That is why the shared business context lives in one file every seat reads, instead of being copy-pasted into five Descriptions that then drift apart.
Setup order (do it in this order)
Build the chief of staff first. One bot, before any others. Pin it to the top of your list so it is the default chat.
Brain dump into it. Who you are, what you sell, what you want off your plate. Do not format it. Talk.
Reverse prompt it. Ask it how it would set up the fleet. It can create its own bots, so this is a config step, not a conversation.
Reject half of what it proposes. See the guardrail below.
Have it write each final Description back to you and save those in a file. Otherwise the app is your only copy and you lose the specs the first time you rebuild.
The reverse prompt, and the two sentences that save you
Paste this as step 3:
Based on everything you know about me, how would you set up
Grokbot? Which bots should we set up, what are their roles and
responsibilities, and what routines should each one run? For every
bot you propose, tell me which tier it serves and what the specific
return is. If you cannot name the return, do not propose the bot.
Those last two sentences are the guardrail. Without them you get a bot per goal. You listed twelve things you want, so it hands you ten bots, each optimizing something, none of them aimed at the thing that pays you.
Forcing every proposal to name a tier and a return makes it rank instead of enumerate. If it cannot answer, that bot does not get built.
The org file: one document every seat reads
This is the part that separates a fleet that works from four bots that contradict each other.
Write one file. It holds your ranking, your rules, your rate card, and one block per seat. Shared context lives here once. Each bot's Description stays short and points at its own job.
Fill this in and keep it in your workspace, not only in the app.
# [Your company] org (for your agents)
Source of truth. Updated [date].
[Your name]. [Role]. [Work email]. Timezone [zone].
## Rank
1. [The one thing that pays you] (Tier 1). First two hours of every day.
2. [Long-term goal] second.
Never rank on what is interesting to build. Named return required
before adding any feature, page, offer, or surface.
## Offer stack
Tier 1 (only headline): [one offer]. Funnel: [step] -> [step] -> [step].
Tier 2, keep running, do not expand: [list].
Tier 3, partnership or side bet: [list].
## Dated items (countdown only)
- [Thing you keep avoiding]: [hard date].
- [Seat kill date]: [date]. [What counts as a pass].
## Standing rules (every seat)
Draft, propose, or queue. Nobody sends, publishes, posts, or spends
without asking me. Approval once is not approval next time.
No bot ever uses my login. Personality is internal tone only.
Anything a third party reads is my voice.
## Rate card (do not invent, do not soften)
[Your real numbers.]
## Connectors
Connected: [list]. Missing: [list]. Do not invent data from a
missing connector.
## Routing
Everyone reports to [chief of staff] first, even if I ask in their chat.
## What no agent should do
[Your hard nos.]
The seat template
One block per bot, inside the org file. Keep it this short.
## [Name]
Title: [Function]
[First-person description. Two or three sentences of personality,
then the job. Personality is internal tone only.]
[The one routine. Day, time, timezone, what it outputs, who gets it.]
May do alone: [read, research, draft, queue]
Must ask first: [send, publish, post, spend, anything irreversible]
Return: [the metric]. Kill: [date and the pass condition].
Two fields people skip and then regret. Return stops you keeping a seat because it feels productive. Kill gives it a date to prove itself by. A seat with no kill date never gets deleted, it just quietly wastes tokens.
The credential model (this is the important one)
Never give a bot your login. Not your email, not your community admin account, not your CRM.
Instead, give each bot its own inbox through an agent email service, then invite that address into your tool as a scoped team member with only the permissions that seat needs.
Your community bot gets a moderator seat, not your owner account
Your CRM bot gets contacts and opportunities on, and settings, workflows and payments off
Revoking one bot never touches anything else
Keep a register of which address is invited where, with the permission level and the date. When you retire a seat, remove its access first, then delete the bot.
Two things to check before you trust it
Both of these bit me, and neither is obvious from the app.
1. Draft-only is a policy, not a control. If you connect an email account, the send and reply tools are usually live in the tool list even when your instructions say draft only. The bot follows the rule because you wrote it, not because anything blocks the call. Treat that seat as capable of sending until you have actually disabled those tools and confirmed it.
2. Check your auto-approval settings. Agent apps ship settings that pre-approve categories of action so you stop getting prompted. Open your settings file and read what is in there. An auto-approval covering browser logins means a subagent can walk through a sign-in flow without asking you first. Know what you have approved before you leave the thing running unattended.
Routines are cron, and you can reverse prompt those too
A routine is a scheduled run. Same idea as a cron job, minus the syntax.
Do not write them by hand up front. Build the seats first, then ask each bot directly:
What routines would you run, how often, and what does each one output?
Asking each seat after it exists gets sharper answers than asking the chief of staff to schedule everyone before anyone has a job. Then stagger the times so two bots are not producing work in the same ten minutes.
What it will not do
No model picker. No reasoning level, no local option. That is deliberate, and it is why the app feels simple.
No local control. You cannot script it from your own machine.
It does not replace a customizable agent setup. Keep your local one for cheap repetitive work, local models, and anything touching your codebase. Attach it as a tool the chief of staff can reach and the grunt work burns local compute instead of paid tokens.
Migrating between agent apps is cheap, so do not treat any of this as a lock-in decision. Ask your current agent to write a handoff list of projects, paths, and rules, then paste that into the new one.
Where the video fits
The video walks the clicks: creating the first bot, the Description field, pinning the chief of staff, connecting a plugin, and scheduling a routine. It ships this week and gets linked here the day it lands.
Build the org file before you watch it. The clicking takes fifteen minutes. Deciding what each seat is allowed to do is the part worth thinking about.
The Night Shift: Run 4 Claude Code Agents At Once (Herdr + BASE)
One Claude session at a time is the ceiling. Herdr gives every agent its own always-on terminal and tells you which one is stuck. BASE briefs all of them from one place. I call the whole thing the night shift.
Herdr is built by the herdr team. BASE is built by Chris Kahler. Both are free and open source.
This page goes deep on Herdr. If you want the full BASE walkthrough, read BASE: A Knowledge Graph For Your Rules right here in the toolkit, then come back.
What the night shift is
Most people run one Claude Code session. You type, you wait, you type again. That is the ceiling, and you hit it every single day.
The night shift is four sessions running at the same time, in one window, on one screen. Each one has a code name. Mine are Wolf, Moose, Toucan, and Zebra.
Two tools make it work:
Herdr keeps every agent alive in a background server and shows you a board. Working, blocked, done, or idle. You never hunt for the stuck one again.
BASE knows your projects, your tasks, your open handoffs, and your rules. It hands each agent a brief before it starts.
Herdr is the room. BASE is the manager. You are the operator.
It is one Rust binary. No Electron, no app to install, no account to make.
Step 2: Start it where your work lives
Go to your project folder and type:
herdr
That is the whole command. Herdr starts a background server and drops you into it. If a session is already running, this reattaches you to it.
Here is the part that matters. The server is what holds your agents, not your terminal window. Close the lid, lose your wifi, restart your terminal. The agents keep going. Type herdr again and everything is exactly where you left it.
Step 3: Open four panes and start four agents
Herdr is mouse first, so you can right click a pane and pick split from the menu. If you like keys, press ctrl+b first (that is the prefix), then the action key.
Split right:prefix+v
Split down:prefix+minus
Do that until you have four panes on screen.
In each pane, start your agent: type claude and hit enter. Codex, opencode, and the rest work the same way.
Herdr detects the agent on its own. The sidebar now shows all four with a live state next to each one: working, blocked, done, or idle.
Give them names you can actually say out loud. Animals work great, because "Zebra is blocked" is a sentence your brain parses instantly. Right click the pane and rename, or run:
herdr agent rename <id> wolf
To see every agent and its state from the command line:
herdr agent list
Step 4: Turn on the chime so you can walk away
The whole point is that you stop babysitting. So make Herdr tell you when an agent finishes or gets stuck.
Herdr runs fine with no config file. To add one, print the defaults into a config:
delivery picks where the popup shows up. Use herdr for an in-app toast, terminal for one that survives SSH, system for a real Mac or Windows notification, off for none.
done_path is the sound when an agent finishes.
request_path is the sound when an agent is stuck and needs you. This is the one that saves your night.
Custom sounds have to be mp3. Relative paths are read from the same folder as the config file.
Load the changes without restarting anything:
herdr server reload-config
On a Mac, run brew install terminal-notifier first if you want the system notification to be clickable.
Step 5: Install BASE and give the agents a brain
Herdr solves "where do my agents live." BASE solves "what should they work on."
Fastest path, no Rust toolchain needed. Grab the prebuilt binary for your machine from the latest release, unpack it, then run:
./base install cd ~/your-workspace base scaffold
base install puts the binary in ~/.local/bin/base, writes global config to ~/.base-gbl/, and wires the hooks into Claude Code. base scaffold creates the .base/ folder in your workspace.
If you would rather build it yourself: cargo build --release, then ./target/release/base install.
Heads up if you already run BASE: base install reseeds your global config. To move to a newer version, copy the new binary over the old one by hand instead.
Confirm it worked:
base --version
The deep dive on domains, rules, and how to strip your CLAUDE.md down lives in the BASE resource. Read that one next.
Step 6: Wire the four sessions together
Now the fun part. BASE lets your sessions talk to each other, so your main session can hand work to the other three.
In each agent pane, register that session under its code name:
base relay register --as wolf --project my-project
Do the same for moose, toucan, and zebra. Then from your main session, see the whole board:
base relay board
That prints every registered session, whether it is alive, and anything waiting in its inbox.
Two ways to hand off work:
A quick message:base relay ping --to wolf --msg "the header fix is on branch feature/nav, review it"
A real task with a brief:base relay task --to wolf --slug nav-fix --summary "fix the mobile nav" --doc /abs/path/to/brief.md
The message fires inside the target session's hooks, so that agent sees it mid-work. It keeps alerting until the agent picks it up.
Step 7: Run the night shift
Here is the actual loop I run. It takes about five minutes to kick off.
Tell your main session who is who. "I have four sessions open: Wolf, Moose, Toucan, Zebra." Now it knows the board.
Ask BASE where you left off. Active projects, open tasks, any handoffs sitting alone. It pulls this from the graph, not from your memory.
Have it rank the work based on what it knows about how you work.
Tell it to match tasks to agents and send each one a brief with base relay task.
Walk away. The chime tells you when someone finishes or gets stuck. The board shows checks and X marks at a glance.
Come back and judge the work. Last run, three of my four briefs were wrong, and the agents told me why. That is the real value. You find out your thinking was off before you burn a day building on it.
Detach any time with prefix+q, or just close the terminal window. Type herdr to walk back in.
Check on it from your phone
Because the server keeps running, you can attach to it from anywhere you can SSH:
herdr --remote user@your-machine
Same board, same agents, same states. This is how you check the night shift from bed without leaving anything running on your laptop screen.
When you actually want everything to stop:
herdr server stop
Keyboard cheat sheet
Prefix is ctrl+b. Press it, let go, then press the action key.
prefix+v split right
prefix+minus split down
prefix+c new tab
prefix+n and prefix+p next and previous tab
prefix+w jump between workspaces
prefix+shift+n new workspace
prefix+q detach and leave everything running
prefix+? show every key that is currently bound
You never have to learn these. The mouse does all of it. But the keys are faster once four panes are live.
Starter prompts
Paste these into your main session once both tools are installed.
I have four Claude sessions open in Herdr: Wolf, Moose, Toucan, Zebra. Run base relay board and tell me which ones are alive.
Pull my active projects, open tasks, and any handoffs from BASE. Rank them by what unblocks the most work.
Take the top four tasks and send one to each session with base relay task. Write a real brief for each, not one line.
Run herdr agent list and tell me which agents are blocked right now and what they are waiting on.
All four agents reported back. Compare what they found against the briefs I sent and tell me which briefs were wrong.
FAQ
Do I need to know tmux? No. Herdr uses a prefix key like tmux does, but everything works with the mouse. Click a pane to focus it, drag a border to resize it, right click for the menu.
Does this cost anything? No. Herdr is Apache 2.0. BASE is public on GitHub. You still pay for whatever agent you run inside them.
Will four agents fight over the same files? They can, so give each one its own scope. Different folders, different branches, or use Herdr's git worktree helpers so each agent gets a clean copy of the repo.
What happens if my laptop restarts? Herdr brings sessions back, and it resumes supported agent conversations by default.
Is this the same as subagents? No. Subagents run inside one session and you cannot see them. This is four real sessions you can watch, interrupt, and message. Read the Agent Teams playbook for when to use which.
Do I need both tools? Herdr alone is still a big win. BASE alone is still a big win. Together is where the night shift actually happens, because one gives you the room and the other gives you the briefs.
Make Yourself Spider-Man In New York: Selfie To 30 Second AI Video
One selfie, two character sheets, one prompt. Seedance 2.5 renders you swinging through Manhattan as a single 30 second take, and Higgsfield just opened 33 days of unlimited renders so the practice runs cost nothing.
Grab both first, both are free
33 days of unlimited Seedance 2.5
Link 1 turns on 33 days of unlimited Seedance 2.5 on Higgsfield, which is what makes this playbook free to practice. Link 2 is the pack of 15 video skills that writes prompts like the one below inside Claude Code.
Every prompt used in this video is on this page. Copy them, swap the description of the person for yourself, and run it.
Link 1 is my affiliate link. Link 2 is free forever on GitHub, no email needed.
What you are making
A 30 second action film where you are the one falling off a Manhattan rooftop and web swinging between towers. Your real face, your real clothes, one continuous scene with a new camera angle every two to three seconds.
It takes one selfie and about 20 minutes. The whole thing is three moves:
Take a selfie with your front camera. That is your only source material.
Turn it into character sheets on Higgsfield, one full body turnaround and one face grid, so the model learns your face from every angle instead of guessing.
Feed the sheets into Seedance 2.5 with the action prompt at 30 seconds, and it holds your face the whole way through.
The character sheets are the step people skip, and skipping them is why AI videos drift into someone else's face halfway through. Sheets first, video second.
What you need before you start
A phone with a front camera. Face lit from the front, plain background, no filter.
A Higgsfield account made through this link, which attaches the 33 days of unlimited Seedance 2.5 to your account.
Everything below works on the Higgsfield website alone, on your phone, with no Claude Code required. The skills just make the next hundred videos faster than this one.
Step 1: take the selfie
Front camera, arm's length, face straight on, even light. No sunglasses yet if you want the model to learn your eyes, or wear them if the sunglasses are part of your look like they are in mine.
One photo is enough. You are not trying to make this photo good. You are giving the model a face to copy.
Step 2: open Higgsfield and start an image
Go to higgsfield.ai on your phone, tap Create Image, and upload the selfie you just took as the reference image.
You are about to run two prompts against that one selfie. The next two steps are those prompts, word for word.
Step 3: generate the body character sheet
This prompt makes a four view turnaround: front, three quarter, side profile, back. It is what teaches the model your build, your clothes, and where your tattoos actually are, so they do not wander between shots.
Copy this into the prompt box with your selfie attached. Edit the physical description to match you, the skin, hair, facial hair, clothing, tattoos, and glasses. Leave the structure and the negative list alone, that is the part doing the work.
Character turnaround model sheet, four consistent full-body views of the same man in a row, evenly spaced left to right: front view, 3/4 front view, side profile view, and back view. The man is exactly the same person as the reference image in all four views, preserving his exact facial identity, bone structure and proportions. Single subject only, exactly one person per view. Pure white seamless studio background, professional character sheet presentation. The same man in his late twenties with light olive skin, oval-to-square face with a defined jawline, mature adult bone structure and adult facial proportions, straight nose, medium-full lips with a natural matte finish, wearing round dark matte-black frame SUNGLASSES with deep smoke-tinted lenses, eyes not visible behind the tinted lenses, subtle realistic reflection across the lens surface with no blown-out specular glare, a dark full mustache with a short trimmed chin beard and light stubble along the jaw and cheeks, dark brown short hair mostly covered by a washed pale sage-grey cotton baseball cap with a curved brim worn forward. Completely clean bare unmarked neck with no tattoos, no ink and no markings anywhere on the neck or throat. Visible fine skin texture with natural pores, fine lines, subtle asymmetries and texture irregularities, slight natural sheen rather than a glossy or dewy retouched finish, no digital smoothing, no beauty filter, no AI-airbrushed look, skin free of artificial glare shine or highlight blooms, matte-to-natural complexion. Lean athletic build with medium-broad shoulders and balanced proportions, dense dark blackwork tattoo sleeve covering the left forearm only, from just below the elbow down over the wrist and onto the back of the left hand, plain gold band ring on the left ring finger. Wearing a charcoal heather grey short-sleeve crew neck t-shirt, light heather grey knit athletic shorts with a drawstring waist ending just above the knee, black rubber slide sandals worn barefoot, no bag. Standing upright in a neutral relaxed pose, arms relaxed at the sides, both feet flat on the ground, full head-to-toe framing with the entire body and both feet visible in every single view, identical outfit, identical sunglasses and identical lighting across all four views. Natural anatomy, high-end but unretouched commercial photography style, soft diffused studio lighting without harsh reflections, cinematic realism, clean white seamless background, sharp focus on skin texture detail. Single subject only, only this one man in frame, no neck tattoo, no throat tattoo, no chest tattoo, no clear or transparent eyeglasses, no other people, no duplicate mismatched faces, no mannequin, no reflections of other objects, no props, no furniture, no background objects, empty seamless studio, all four views standing full-body head-to-toe not cropped not sitting, no text, no watermark, no logos, no frame borders, no labels, no distorted anatomy, no extra fingers, no beauty filter, no digital smoothing, no airbrushing, no plastic skin, no glossy skin, no babyface, no overly youthful rounded proportions.
This is what comes back:
Step 4: generate the face character sheet
The body sheet holds your build. This one holds your face. Six tight head and shoulders portraits in a grid: front, both three quarter turns, both full profiles, and a raised chin angle.
Same selfie attached, same edits to the physical description, new prompt:
Character face reference sheet, a clean even grid of six tight head-and-shoulders close-up portraits of the same man, three across the top row and three across the bottom row, evenly spaced on a pure white seamless studio background. Top row: straight front-facing neutral expression, 3/4 turn to his left, full left side profile. Bottom row: 3/4 turn to his right, full right side profile, and a slightly raised chin three-quarter view. The man is exactly the same person as the reference image in every panel, preserving his exact facial identity, bone structure and proportions, identical face in all six panels. Single subject only, exactly one person per panel. The same man in his late twenties with light olive skin, oval-to-square face with a defined jawline, mature adult bone structure and adult facial proportions, straight nose, medium-full lips with a natural matte finish, wearing round dark matte-black frame SUNGLASSES with deep smoke-tinted lenses, eyes not visible behind the tinted lenses, subtle realistic reflection across the lens surface with no blown-out specular glare, a dark full mustache with a short trimmed chin beard and light stubble along the jaw and cheeks, dark brown short hair mostly covered by a washed pale sage-grey cotton baseball cap with a curved brim worn forward. Completely clean bare unmarked neck with no tattoos, no ink and no markings anywhere on the neck or throat in any panel. Wearing a charcoal heather grey short-sleeve crew neck t-shirt. Visible fine skin texture with natural pores, fine lines, subtle asymmetries, individual facial hair strands and texture irregularities, slight natural sheen rather than a glossy or dewy retouched finish, no digital smoothing, no beauty filter, no AI-airbrushed look, skin free of artificial glare shine or highlight blooms, matte-to-natural complexion. Identical sunglasses, identical lighting and identical crop scale across all six panels, natural anatomy, high-end but unretouched commercial portrait photography style, soft diffused studio lighting without harsh reflections, cinematic realism, clean white seamless background, extremely sharp focus on skin texture and facial detail. Every panel is a tight close-up, no full body shots, single subject only, no neck tattoo, no throat tattoo, no chest tattoo, no clear or transparent eyeglasses, no other people, no duplicate mismatched faces, no mannequin, no props, no furniture, no background objects, empty seamless studio, no text, no watermark, no logos, no frame borders, no labels, no distorted anatomy, no beauty filter, no digital smoothing, no airbrushing, no plastic skin, no glossy skin, no babyface, no overly youthful rounded proportions.
This is what comes back:
Step 5: run them again, then download everything
Run both prompts two or three more times. You are unlimited during the window, so there is no reason to accept the first result.
Keep the sheets where your face is unmistakably yours and the details hold across every panel, the same cap, the same glasses, the tattoo on the correct arm. Throw out any sheet with a drifting face, a bonus person, or a tattoo that jumped sides. Download every sheet you keep.
Three good sheets beats ten mediocre ones. Garbage in the sheets becomes a stranger in your video.
Step 6: create the video
Back on Higgsfield, start a new video and upload all of the character sheets you just downloaded as references. Seedance 2.5 takes up to 50 reference files in one render, so use every good sheet you have.
Upload every character sheet.
Paste the action prompt below.
Pick the Seedance 2.5 model. Say it out loud to yourself, because picking the wrong model is the single most common way people waste a render.
Drag the duration all the way to 30 seconds.
Generate.
The exact video prompt I used
This is the full prompt, unedited. Shot list first, then physics rules, then look, then audio, then negatives. That order matters, and the physics and negative blocks are why the webs behave instead of turning into a random cable across the frame.
Fast-paced 30s action sequence, 16:9, MULTIPLE DYNAMIC CAMERA CUTS, a new angle every 2 to 3 seconds, constant motion, nothing static. Same man as all three reference sheets: identical face, dark round sunglasses, pale sage-grey cap worn forward, charcoal grey crew tee, light grey knit shorts, black slide sandals, blackwork tattoo sleeve on LEFT forearm only. Same face in every shot.
SHOT 1 (0-2s): Tight low shot on his black slide sandals at the concrete parapet edge, toes over the drop, Manhattan 40 storeys below, wind whipping.
SHOT 2 (2-3.5s): CUT to low-angle wide, him against the golden hour skyline, tipping straight backwards off the edge, arms opening.
SHOT 3 (3.5-5s): CUT to bird's eye directly overhead looking straight down, he falls away from camera shrinking fast, street grid far below.
SHOT 4 (5-7s): CUT to tight on his face falling, cheeks rippling in the airflow, cap pinned, sunglasses steady, dead calm.
SHOT 5 (7-9s): CUT to wide freefall alongside him, tower facade ripping past, pigeons scattering off a ledge.
SHOT 6 (9-10.5s): CUT to extreme close on his right wrist as he snaps his arm UP, a thin white strand ejecting from the wrist.
SHOT 7 (10.5-12s): CUT to follow the strand tip flying UPWARD through open air toward the top of a tall tower far ABOVE him, the tip hits high on the upper facade and splats, sticking.
SHOT 8 (12-14s): CUT to wide, line taut running steeply UP from his hand to that high anchor, his body snapping out of the fall into an arc beneath it.
SHOT 9 (14-18s): Swinging pendulum beneath the anchor, camera orbiting hard with him, city rotating, sun strobing between buildings.
SHOT 10 (18-20s): CUT to street level looking straight up as he rips past overhead, traffic and steam vents in the foreground.
SHOT 11 (20-22s): CUT to him releasing at the top of the arc, tumbling, then snapping his other arm UP and firing a second strand upward to a further tower.
SHOT 12 (22-27s): Fast swing between towers, camera whipping laterally, city smearing, a helicopter crossing the background.
SHOT 13 (27-30s): CUT to a wide hero frame, him small against the canyon, sun blazing down the avenue.
WEB PHYSICS: every web fires UPWARD from his wrist and anchors HIGH on a tower ABOVE his body, so he hangs and swings BENEATH it. The web always stays attached to his hand. NEVER a web coming from below him, NEVER anchored at street or ground level, NEVER a long horizontal cable across the frame, NEVER both ends off screen, NEVER an existing wire he grabs, NEVER sliding along the line. Always show the strand travelling before it connects.
LOOK: 24mm wide, deep focus, late golden hour, warm key from the low western sun, cool sky fill, haze, restrained teal-orange, 35mm grain, heavy motion blur on the fast moves.
AUDIO DIEGETIC ONLY, NO MUSIC: no score, no soundtrack, no instruments, no melody at any point. Real sound only: dominant wind roar, distant street ambience, horns, a siren, clothing hammering, a sharp pneumatic thwip on each web fire, a wet splat as each tip sticks, rope tension creak, heavy swing whoosh with doppler, helicopter rotor, ending on wind.
NEGATIVE: no zipline, no tightrope, no pre-existing cable, no line unattached to his hand, no superhero costume, no mask, no spandex, no chest emblem, only his own clothes. No music. No other people falling, no face drift, no clear eyeglasses, no neck tattoo, no text, no watermark, no slow motion.
Why the prompt is built that way
Four blocks are doing all of the heavy lifting, and they transfer to any 30 second scene you write next.
Timestamped shot list. Thirteen shots with second ranges is what stops a 30 second render from becoming one slow drifting camera move. You are directing the edit, not describing a vibe.
A physics block. Video models have no idea how a web works. Spelling out that the strand fires upward, anchors above him, stays attached to the hand, and never crosses the frame horizontally is the difference between web swinging and a man holding a random cable.
An audio block. Seedance 2.5 generates sound with the picture, so if you do not say what you want you get generic score. Naming diegetic sound only, wind, horns, the thwip, the splat, is what makes it read as footage.
A negative block. No costume, no mask, no emblem, no face drift. That is what keeps it you in your own clothes instead of a copyright-shaped man in spandex.
Swap it for your own scene
The character sheets are reusable forever. Once your face is captured, the sheets work for any video you make after this. You only rewrite the action prompt.
Parkour chase: keep the shot list shape, change the action to rooftop running and vaults, drop the web physics block, keep the audio block.
Founder brand film: thirteen shots becomes six slower ones, golden hour becomes window light, and the audio block becomes room tone and keyboard.
Product ad with you in it: use your sheets plus photos of the product, and let the 07-ecommerce-ad skill from the free pack write the shot list.
If you want Claude Code writing these instead of you, install the 15 free Seedance video skills and describe the video in one sentence. The skills already carry the camera, lens, and lighting language.
Four mistakes that waste the window
Skipping the character sheets. Feed a raw selfie straight into the video and your face changes halfway through. The sheets exist to stop exactly that.
Keeping a bad sheet. A drifting face or a tattoo on the wrong arm in the reference becomes a drifting face in every single shot.
Leaving the model on default. Pick Seedance 2.5 explicitly, or you get a shorter render on a different model.
Writing a vibe instead of a shot list. "Me swinging through New York" gives you eight seconds of nothing. Timestamps and cuts give you a film.
Get your 33 days
⚡ 33 days unlimited, live now
33 days of unlimited Seedance 2.5, so every character sheet and every retry costs you nothing.
I am Charles J Dove, and I run Charlie Automates. At my agency CC Strategic we build these systems for clients, and everything I test lands free on charlieautomates.com first.
Not sure what to fix first? The free AI Bottleneck Finder tells you the one constraint capping your revenue right now, plus two builds you could ship this week. Five minutes, no call required. Take the free AI Bottleneck Finder →
Seedance 2.5 Is Live: 33 Days Of Unlimited AI Video, Free
Seedance 2.5 makes one clean 30 second shot instead of five clips you have to stitch. It launched, and Higgsfield opened it up for 33 days of unlimited use. Here is the full playbook, including the 15 free Claude Code skills that write the prompts for you.
Start here, both links are free
33 days of unlimited Seedance 2.5
Link 1 turns on 33 days of unlimited Seedance 2.5. Link 2 is the pack of 15 video skills that writes your prompts inside Claude Code. You need both to run this playbook.
The skills say Seedance 2.0. They work with 2.5 too. They write prompts, and prompt writing did not change. You just get more room to use them.
Link 1 is my affiliate link. Link 2 is free forever on GitHub, no email needed.
What is Seedance 2.5?
Seedance 2.5 is the new video model from ByteDance, and it is live on Higgsfield right now with 33 days of unlimited use. The reason people are paying attention is length: Seedance 2.5 generates a single 30 second shot in one pass, where every model before it made you build a scene out of short clips and glue them together.
Five things changed from the last version:
30 seconds in one shot. Not four clips stitched. One continuous take, so the light, the room, and the person stay the same from the first frame to the last.
Up to 50 reference files per render. Photos, video clips, audio, and text, all in one request. You can hand it your real product from ten angles and it keeps that product.
Region editing. Change one thing in a finished clip, like a shirt color or a label, without regenerating the whole video.
Sound made with the picture. Footsteps, voices, and background noise come out already lined up with the motion instead of being added later.
Better prompt following. Roughly 20% closer to what you asked for, and it holds small details like hair, text, and fabric across frames.
What is the difference between Seedance 2.0 and 2.5?
Seedance 2.0 renders in 4K. Seedance 2.5 trades on length, references, and control. Both are live, so the honest answer is that each one wins at something different and you pick per job.
What you get
Seedance 2.0 (live now)
Seedance 2.5 (live now)
Longest single shot
15 seconds
30 seconds
Reference files per render
9 images, 3 clips, 3 audio
Up to 50, any mix
Fixing one part of a clip
Regenerate the whole thing
Change one region
Sound
Added after
Made with the picture
Prompt accuracy
Baseline
About 20% tighter
4K output
Yes, and it is free right now
Higgsfield has not published a launch resolution
Available today
Yes
Yes, 33 days unlimited
Pick per job. Reach for 2.0 when you need maximum sharpness in a short shot. Reach for 2.5 when you need length, a pile of references, or sound that lands with the picture.
What can you actually make with it?
Seedance 2.5 is built for finished work, not test clips. A 30 second shot is a whole ad, a whole product demo, or a whole scene, which means you stop editing and start publishing. Below is what to make, and which of the 15 free skills writes the prompt for it.
What you want to make
Skill that writes it
Why 2.5 helps
Product commercial
07-ecommerce-ad
30 seconds is a full ad slot, no cuts needed
Product spin, every angle
09-product-360
Feed it 50 photos of your real product
Founder or brand story film
12-brand-story
One take holds the same face and room
Explainer or kinetic type ad
06-motion-design-ad
Text stays readable across frames
Music video
10-music-video
Sound is generated with the picture
Creator style ad (AI UGC)
11-social-hook
The person looks the same the whole clip
Cinematic short
01-cinematic
Camera moves run the full 30 seconds
Restaurant or drink shot
14-food-beverage
Steam, pour, and melt physics hold up
Property or interior tour
15-real-estate
One walk through instead of five clips
The other six skills in the pack cover 3D, cartoon, comic panels, fight scenes, anime, and fashion lookbooks. Same idea, different look.
Why do you need the 15 free video skills?
The 15 video skills are free Claude Code files that already know how Seedance wants a prompt written: camera move, lens, lighting, pacing, and shot order. Without them you type a sentence and get soft, floaty footage. With them you type the same sentence and Claude turns it into a full production prompt before it renders.
This is the step most people skip, and it is the reason their videos look generated instead of shot. Use the skills. They are the difference.
The pack is labelled Seedance 2.0, and it works with 2.5 the same way. The skills write prompts, and prompt writing did not change between the two, so nothing in the pack is out of date. You just have more room to use it now that scenes run to 30 seconds. Grab it here: the 15 free Seedance video skills on GitHub. Free, no email, nothing to buy. There is also a full write up of the pack here on charlieautomates.com.
How do you set this up?
Setup is four steps and takes about 20 minutes, and most of that is Claude installing files while you watch. Do them in this order, because the skills cannot render anything until the Higgsfield connection is live.
Claim the 33 days.Open this link and make your Higgsfield account through it. The 33 days of unlimited Seedance 2.5 attach to the account you land on.
Install the 15 skills. Copy this repo link, paste it into Claude Code, and say: "Install every skill in this repo into my skills folder." Claude does the rest.
Install the Higgsfield MCP. The MCP is the bridge that lets Claude fire the render itself instead of handing you a prompt to paste somewhere. Full walkthrough here: Higgsfield MCP setup guide.
Check it. In Claude Code, type /mcp and read the list. If higgsfield says connected, you are done. If it does not, the MCP did not install, and nothing after this will work.
What do you type to make a video?
You give Claude three things in one message: what the video is, which skill to use, and which model to run. Name the model every time. If you leave it out, Claude picks a default and you waste a free render on the wrong setting.
Copy this shape:
Make me a 10 second product video for [product].
Use the [skill name] skill from the Seedance skill pack.
Use Seedance 2.5.
Scenes now run to 30 seconds, so raise the length when the shot needs it. Want maximum sharpness on a short clip instead? Change the last line to Use Seedance 2.0 in 4K. Everything else stays the same.
Three prompts you can copy right now
Change the bracketed part and send it. Swap the model line any time you want 2.0 in 4K instead.
Product commercial:"Make a 10 second product ad for my [matte black water bottle]. Use the 07-ecommerce-ad skill. Seedance 2.0 in 4K. Studio lighting, slow push in, product centered, no text."
Product spin:"Make a 6 second 360 turntable of [this sneaker] on a plain white background. Use the 09-product-360 skill. Seedance 2.0 in 4K."
Brand story:"Make a 12 second brand story opener for [my coffee roasting company]. Use the 12-brand-story skill. Seedance 2.0 in 4K. Warm morning light, hands working, no faces."
Claude reads the skill, writes the full production prompt, sends it to Higgsfield through the MCP, and hands the finished file back to you. You keep working while it renders.
How much does this cost?
It costs $0 to run this playbook for the next 33 days. Seedance 2.5 is unlimited during the window, the 15 skills are free on GitHub, the Higgsfield MCP is free to connect, and Claude Code runs on the plan you already pay for. The only spend is the Higgsfield subscription you start through the link, which is what turns the 33 days on.
Compare that to hiring it out. One product video from an agency runs $2,000 to $3,000 and takes 2 to 3 weeks, because you are paying a videographer, a lighting setup, an editor, and a colorist. One render on this setup takes about 3 to 5 minutes. That gap is the entire reason to learn this now.
Three mistakes that waste the window
Most people who try this get a weak result and blame the model. It is almost always one of these three, and all three take one minute to avoid.
Not picking the model. If you never say which model to use, Claude picks a default and you burn a render on the wrong setting. Name Seedance 2.5 in the prompt every time.
Skipping the skills. Hand written prompts produce soft, drifting footage. The 15 skills carry the camera and lighting language that makes a clip look shot instead of generated.
Not checking /mcp first. If Higgsfield is not connected, Claude will write you a beautiful prompt and render nothing at all.
Where this fits in a bigger system
This setup is one example of a bigger pattern: you give Claude Code a specialist file for every job you repeat, then let it run those jobs on request. That pattern is what an agentic OS is. Video is the easiest place to see it pay off, because the before and after is a $3,000 invoice turning into a 5 minute render.
I am Charles J Dove, and I run Charlie Automates. At my agency CC Strategic we build these systems for clients, and everything I test ends up free on charlieautomates.com first. For the full breakdown of how the pieces fit together, read the 3-step system to build an agentic OS, or watch the walkthroughs on the Charlie Automates YouTube channel @charlieautomates.
Get your 33 days
⚡ 33 days unlimited, live now
33 days of unlimited Seedance 2.5. No credit counting while you learn it.
Not sure what to fix first? The free AI Bottleneck Finder tells you the one constraint capping your revenue right now, plus two builds you could ship this week. Five minutes, no call required. Take the free AI Bottleneck Finder →
Seedance 2.0 In 4K Is Free This Week: Make Product Videos With Claude Code
A product video like the ones in this guide runs $3,000 from an agency. Seedance 2.0 in 4K went free this week, and 15 free Claude Code skills write the prompts for you. Here is the whole setup, start to finish, in about 20 minutes.
Both links, grab them first
Unlimited Seedance 2.0 in 4K, free until Friday
You need two things for this guide, and both are free. Link 1 turns on unlimited Seedance 2.0 in 4K on Higgsfield. Link 2 is the pack of 15 video skills that writes your prompts inside Claude Code.
⚠ The 4K window closes Friday, 08/07/26. Subscribe during the window and Seedance 2.5 is unlimited for 7 more days after it launches.
Link 1 is my affiliate link. Link 2 is free forever, no email needed.
What went free this week
Seedance 2.0 in 4K is now free to run on Higgsfield through Friday, 08/07/26. Seedance 2.0 is the video model most of the good AI product ads you have seen this year were made with, and 4K is its top output tier. Until this week, 4K was the setting people skipped because it burned credits the fastest.
Free here means unlimited. You are not counting a balance, so you can render the ugly first version, the better fifth version, and the one you actually ship, and pay nothing for the practice in between. If you subscribe during the window, Seedance 2.5 is also unlimited for 7 days after it launches.
Why does a product video cost $3,000?
A single product video from an agency costs $2,000 to $3,000 because you are paying for four people, not for footage. A shoot needs a videographer, a lighting setup, an editor, and a colorist, plus a studio day. That is the real invoice, and it takes 2 to 3 weeks.
Seedance 2.0 replaces the shoot. The 15 skills replace the person who knows how to talk to it. You still supply the taste and the product. What disappears is the $3,000 and the three weeks.
What are the 15 free video skills?
The 15 video skills are Claude Code skill files that already know how Seedance 2.0 wants a prompt written: camera moves, lens, lighting, pacing, and shot order. You describe the video in plain English, Claude picks a skill, and the skill turns your sentence into a full production prompt.
Five of the 15 are built specifically for commercial and marketing work. Those are the ones that pay for themselves.
Setup is four steps and takes about 20 minutes, most of which is Claude installing files while you watch. Do them in this order, because the skills are useless until the MCP is connected.
Turn on the free 4K window.Open this link and make your Higgsfield account through it, so unlimited Seedance 2.0 in 4K is attached to that account.
Install the 15 skills. Open the free skills pack, copy the repo link, paste it into Claude Code and say: "Install every skill in this repo into my skills folder." Claude does the rest.
Install the Higgsfield MCP. The MCP is the bridge that lets Claude actually fire the render instead of just writing a prompt you paste somewhere. Full walkthrough here: Higgsfield MCP setup guide.
Verify it. In Claude Code, type /mcp and scroll the list. If higgsfield shows connected, you are done. If it does not, the MCP did not install and nothing after this will work.
What do you type to make the video?
You give Claude three things in one message: what the video is, which skill to use, and the words "Seedance 2.0 in 4K". That third part matters. If you leave the model and resolution out, Claude picks a default and you burn the free 4K window on a 1080p render.
Copy this shape:
Make me a 6 second product video for [product].
Pick the right skill from the Seedance skill pack.
Use Seedance 2.0 in 4K.
Three prompts you can copy right now
These are the three that cover most small business work. Change the bracketed part and send it.
E-commerce ad:"Make a 6 second e-commerce ad for my [matte black water bottle]. Use the 07-ecommerce-ad skill. Seedance 2.0 in 4K. Studio lighting, slow push in, product centered, no text."
Product 360:"Make a 5 second 360 turntable of [this sneaker] on a plain white background. Use the 09-product-360 skill. Seedance 2.0 in 4K."
Brand story:"Make an 8 second brand story opener for [my coffee roasting company]. Use the 12-brand-story skill. Seedance 2.0 in 4K. Warm morning light, hands working, no faces."
Claude reads the skill, writes the full production prompt, sends it to Higgsfield through the MCP, and hands you back the finished file. You keep working in Claude Code while it renders.
What does it actually cost?
Right now it costs $0. During the free window, every Seedance 2.0 render in 4K is unlimited, the 15 skills are free on GitHub, the Higgsfield MCP is free to connect, and Claude Code runs on the plan you already pay for.
Compare that to the alternative. One agency product video is $2,000 to $3,000 and 2 to 3 weeks. One render on this setup is about 3 to 5 minutes. That is the entire pitch.
The three mistakes that waste the free window
Most people who try this get a mediocre result and blame the model. It is almost always one of these three, and all three are avoidable.
Not saying "4K". Claude defaults to a cheaper tier. Say the model and the resolution in the message, every time.
Skipping the skills. Raw prompts written by hand get soft, floaty footage. The skills carry the camera and lighting language that makes it look shot instead of generated.
Not checking /mcp first. If Higgsfield is not connected, Claude will happily write you a beautiful prompt and never render anything.
Where this fits in a bigger system
This setup is one skill inside a much larger pattern: giving Claude Code a specialist file for each job you repeat, then letting it run those jobs on request. That pattern is what an agentic OS is, and video is one of the easiest places to see it pay off, because the before and after is a $3,000 invoice.
I am Charles J Dove, and I run Charlie Automates. At my agency CC Strategic we build these systems for clients, and everything I test ends up free on charlieautomates.com first. If you want the full breakdown of how the pieces fit, read the 3-step system to build an agentic OS, or watch the walkthroughs on my YouTube channel @charlieautomates.
Start before Friday
⚡ Free through 08/07/26
Unlimited Seedance 2.0 in 4K, plus 7 unlimited days of Seedance 2.5 after it launches.
When the window closes, 4K goes back to costing credits.
Not sure what to fix first? The free AI Bottleneck Finder tells you the one constraint capping your revenue right now, plus two builds you could ship this week. Five minutes, no call required. Take the free AI Bottleneck Finder →
Build An AI CMO That Runs Your Marketing (Claude Code + SEED + SkillSmith)
An executive officer is not a chatbot. It is a skill file with your data, your tools, and your approval gates wired in. Here is the exact build: SEED shapes the idea, SkillSmith writes the agent, MCPs give it eyes on your CRM, your site, and your content.
Can you really build an executive team with AI?
My honest answer is about 60% yes.
You are not replacing a human executive. You are building something that reads your real data and tells you what to do next, every day, without you asking. That is the 60%. The judgment call at the end is still yours.
95% of people are nowhere near using AI at full capacity. The thing you have been picturing is almost certainly possible right now. The hard part was never the model. It is knowing where to start, what data to feed it, and how to give it access to your tools without handing it the keys to everything.
This build is a CMO. Same process works for any officer you want.
Augmentation first, automation second
Most people try to automate a process they have never written down. That is why their automations break.
Quick test: can you name your five highest ROI processes right now, out loud, in order? If you cannot, forget automation for a minute. Build for augmentation instead, which means AI that helps you make better decisions with the data you already have.
The CMO in this build has some automation in it. But its main job is to read the numbers and hand you a decision.
The three things every AI officer needs
An agentic coding tool. I use Claude Code inside VS Code. I would not build this anywhere else. New to it? Start with my Claude Code crash course before this one.
Clarity on what it builds and what it can touch. The deliverable, the data, the access. This is the part people skip, and it is why their agent comes out useless.
The right frameworks. Without structure, it does not matter how good your idea is. The output will miss. I use two: SEED to shape the idea, and SkillSmith to build the agent.
An agent is just a skill file
Strip away the hype and an executive officer is a Markdown file. Plain text a model can read, holding clear instructions for how the worker should operate.
Open one of mine and you find the same parts every time: the entry point that holds triggers and commands, context about your business, frameworks it follows, tasks that string the run together, templates for how the output should look, and checklists that gate risky actions.
That is the whole trick. Skills, commands, agents, they are the same thing wearing different names.
Step 1: ideate the officer with SEED
Run /seed and describe the officer you want. Do not write a spec. SEED does the heavy lifting through a guided interview.
My prompt named four things, and yours should too:
What it is. A CMO officer, packaged with SkillSmith once the idea is shaped.
What data it may pull. Apify to scrape my Instagram and YouTube.
What commands it may call. My existing /viral discover pipeline for competitor and self content audits.
What systems it may read. GoHighLevel for CRM, prospects, and deals. GA4 for site analytics.
I also told it plainly: focus on augmentation, not full automation. That one line changes the whole build.
SEED writes the plan out to a file in your project. Read it before you move on.
Step 2: build it with SkillSmith
Hand the SEED plan to SkillSmith and run /skillsmith discover. It interviews you again, tighter this time, then scaffolds the skill folder.
Two questions it asked me that matter more than they look:
Is this a consultant or an operator? I picked consultant. A consultant reads and recommends. An operator takes action, and it gets complicated and convoluted fast. Start as a consultant.
How should it be invoked? One front door. Just /cmo. Not fifteen commands you will forget by Thursday.
Model routing, since people ask: I ran Opus 5 for the ideation and switched to Fable 5 for the build.
Step 3: set the approval gates
Inside the skill there is a checklist file, and this is your safety layer.
Mine reads: consult, recommend, draft, and stop. The CMO can write a campaign. It cannot send one. Nothing leaves my hands without me looking at it.
Decide this before you wire anything up. Fully autonomous, fully manual, or a blend. Write it into the checklist so the agent cannot quietly drift past it later.
Step 4: connect the tools with MCPs
MCPs and CLI tools are the bridges between Claude Code and your software. No bridge, no data, and the agent is guessing.
GoHighLevel MCP for pipelines, contacts, and deals.
Apify for scraping your Instagram and YouTube performance.
GA4 MCP for site analytics.
Only connect what the officer actually needs. Every bridge you add is another door, so keep the list short and keep the gates on.
Step 5: give it your memory
Your agent is only as good as what it knows about you. If your context has been living in ChatGPT or Claude chats, go get it.
Open settings and pull your saved memories, or open a fresh session and ask: what are the core things you know about me? I am building a CMO officer. Put it into one document I can hand to another setup.
Read what comes back, cut what is stale, and drop the rest into the skill as context. You can skip the onboarding interview entirely if your operator profile is already solid.
Run it: what my first /cmo actually returned
Type /cmo and let it work. I asked for a full audit of my channels and an HTML report I could open in the browser.
YouTube. Three months of performance, key insights, every video linked so I could go back and rewatch the winners.
Instagram. Reels and carousels scored, including which posts landed off my ideal customer and underperformed because of it.
GoHighLevel. A full pipeline read that showed me exactly which two stages had gone stale and needed a campaign.
Then it updated its own brain with the audit, so the next run starts smarter than this one did.
This part is slow and boring to watch. The boring things are the ones that get results.
Making it run without you
Manual slash commands are a fine place to start. When you want it running on its own, you have two paths.
Routines and cron. Cron is a time trigger. Schedule /cmo for every Monday at 9am. The catch on a local setup: your computer has to stay on.
Send the skill to a host. Hermes on a VPS, or Claude managed agents. Both run 24/7 without your machine. I have full walkthroughs for these.
Where to go from here
Every officer you are thinking about building is this same process. The only two questions that change are what it can access and what data you train it on.
If you want to learn to build these yourself at your own pace, that is what we do inside CC Strategic AI on Skool. The zero to dangerous Claude Code course, the templates, and weekly calls where you show me what you built.
Not sure what to fix first? The free AI Bottleneck Finder tells you the one constraint capping your revenue right now, plus two builds you could ship this week. Five minutes, no call required. Take the free AI Bottleneck Finder →
Resources
SkillSmith playbook, the plugin that builds the agent, with the install command and the skill anatomy
SkillSmith video walkthrough: https://youtu.be/FOvN28l_q9s
SEED playbook, the idea incubator that writes the plan, with the skill link inside it
GoHighLevel MCP, the CRM bridge this build reads pipelines through
Replace Yourself With Any AI Avatar In Your Own Videos (Claude + Higgsfield MCP)
Record the video once as yourself. Then hand Claude the clip and a picture of any character, and it hands back the same video with that character performing your exact movements, your exact camera moves, in your exact room. One prompt, about 5 to 10 minutes.
Read this first
Get 24 hours of unlimited Higgsfield MCP, free
You need the Higgsfield MCP for this tutorial. Higgsfield is giving 24 hours of unlimited MCP usage to anyone who signs up through the link below. Unlimited means you can run every avatar swap in this guide without watching a credit balance.
⚠️ Sign up before Friday. When the window closes, it is gone.
My affiliate link. Grab the free window first, then come back and follow the steps.
What this actually does
You film yourself. You talk, you move your hands, you walk across the room, the camera pans. Normal video.
Then you give Claude two things: that video, and one still picture of a character. Could be an anime character. Could be a cartoon version of you. Could be a completely different person.
Claude sends both to Higgsfield, and Higgsfield rebuilds the video with that character in your place. Same movements. Same timing. Same camera moves. Same background.
You are not typing a description of a video and hoping the AI guesses right. You are handing it a performance you already recorded and telling it to repaint the actor. That is why it comes out so accurate.
The whole thing takes about 5 to 10 minutes of waiting, and you do it from a Claude chat. No editing software, no green screen, no motion capture suit.
What you need before you start
A Claude subscription. Any paid plan works. This runs in Claude Code or the Claude desktop app.
One video of you. Short. Think 5 to 10 seconds per shot.
One picture of the avatar you want to become. Just a still image. One file.
That is the entire shopping list. No API keys to manage, no other tools to install.
What an MCP is, in one paragraph
MCP stands for Model Context Protocol. Skip the name. Here is what matters: an MCP is a toolbox you hand to Claude. Once it is connected, Claude can reach into that toolbox and use the tools inside it on its own.
The Higgsfield MCP puts every top image and video model into that toolbox. So instead of you logging into a website, uploading files, picking a model, and clicking generate, you just tell Claude what you want and it does all of that for you.
Step 1: Install the Higgsfield MCP
This takes about a minute and you only do it once.
Open this link and create your account so the free unlimited window is attached to it.
Go to higgsfield.ai/mcp and copy the MCP connection link they give you.
Paste that link into Claude and say: install this MCP for me. Claude handles the setup.
In Claude Code, run /mcp, click the Higgsfield entry, and hit Allow so it can log into your account.
Ask Claude "what Higgsfield tools do you have?" to confirm it worked. If it lists things like generate video and motion control, you are connected. If you want the deeper walkthrough of the MCP itself, that is in Higgsfield AI MCP: Every AI Image + Video Model Inside Claude Code.
Step 2: Get your reference video ready
Your reference video is the performance. Whatever you do in it, the avatar does. So a few rules keep the results clean:
Keep each shot short. 5 to 10 seconds. If you have a 60 second video, that is fine, but it works best broken into separate shots.
Stay in frame. If you walk out of the shot and back in, the model has to guess what happened in between.
Good light on your face and body. The model reads your pose from the pixels. Dark footage gives it less to read.
One person on camera. Two people in frame means the model has to decide which one to replace, and it will not always pick right.
Shoot it in the aspect ratio you want to post. Vertical for Reels, TikTok, and Shorts. Horizontal for YouTube.
Save the file somewhere you can find it. You are going to give Claude the file path.
Step 3: Pick your avatar image
This is one still picture. Not a video, not five photos. One image of the character you want to become.
Where people get them:
Make one on Higgsfield. Ask Claude to generate the character first. "Make me a full body image of a silver haired anime swordsman in a dark hoodie, plain background." Then use that image as your avatar.
Use an existing character image you have the right to use.
Use a stylized version of yourself. Feed Claude a photo of you and ask it to redraw you as a 3D animated character, then swap yourself with that.
What makes a good avatar image: the whole body visible if your video shows your whole body, facing the camera, clean and uncluttered background, and sharp enough that the face is clear.
Step 4: The exact prompt
Start a fresh Claude chat, then paste this in. Replace the two paths at the bottom with wherever your files actually live on your computer.
Use the Higgsfield MCP for this.
I am giving you a reference video of me and a still image of a character. I want the character to replace me in the video, doing my exact movements.
Do this:
1. Upload both files to Higgsfield.
2. Split the reference video into its separate shots. Treat each shot as its own job.
3. For every shot, run Higgsfield motion control using the character image as the character and that shot as the motion video.
4. Set scene_control to "video" so my original background and camera movement stay.
5. Set resolution to 1080p.
6. Submit no more than 3 jobs at a time, then wait for them to finish before starting the next batch.
7. Poll until every job is done, then give me the finished video links in shot order.
Reference video: /path/to/my-video.mp4
Character image: /path/to/my-avatar.png
If you would rather do it the way it looks in the video, drop the reference video in first, send the prompt, then follow up with the avatar image and one line: use this character. Same result, it just splits into two messages.
Step 5: Let it run
Claude uploads the files, splits the video, and fires one job per shot. Then it waits.
Expect 5 to 10 minutes. A single short shot can land in 2 or 3 minutes. A longer video with six shots takes longer because they run in batches.
Do not close the chat and do not re-send the prompt because it feels slow. Re-sending starts the whole thing over and burns the runs twice. If you want a status check, just ask: what is the job status?
Step 6: Get your videos
When the jobs finish, you have two ways to collect them:
Straight from the chat. Claude hands you the finished video links. Ask it to show your recent generations any time and it pulls them through the MCP.
From the Higgsfield website. Everything you generate lands in your account, so log in and download from there if you would rather have the files in a folder.
Then stitch the shots back together in whatever editor you already use. You now have the same video you filmed, performed by somebody else.
The three settings that decide if it looks good
Most bad results come from one of these three being wrong. This is the part nobody explains.
scene_control. This picks where the background comes from. Set it to video and you keep your real room and your real camera move, which is what you want when you are replacing yourself in your own footage. Set it to image and the background comes from the avatar picture instead, which is what you want if you are trying to teleport the whole scene somewhere else.
resolution. 720p is cheaper and faster. 1080p is what you post. During the free unlimited window, there is no reason to run anything but 1080p.
Shot length. The single biggest quality lever. Short shots hold the character steady. Long shots give the model more chances to drift, and drift is what makes an avatar's face slowly stop looking like the avatar.
What people actually use this for
You do not want to be on camera. Film it yourself in your kitchen, post it as a character. Your delivery, none of your face.
One video, many versions. Film the script once, then run it with four different avatars and test which one people watch longest.
Client ads. Record the performance yourself, then swap in a character that matches the client's brand. You never have to hire and schedule talent.
Doing what you cannot film. Your movement, a character and a world you could never have shot in real life.
Things that will trip you up
Do not paste the video into the chat and expect Higgsfield to see it. The MCP runs on Higgsfield's side, so the file has to get uploaded there first. Give Claude the file path and let it handle the upload, or use the upload widget when it offers you one.
Do not fire ten jobs at once. Higgsfield caps how many run at the same time and you will get rate limited. Three at a time is the safe number, which is why it is in the prompt.
Hands are the hard part. Fast, complicated hand movement is where these models still struggle. Slow your gestures down a notch when you film.
Your audio does not come along automatically. Keep the original audio from your reference video and lay it back over the finished clips in your editor.
Only swap characters you have the right to use. Somebody else's face or a studio's character is their property. Your own likeness, an original character, or one you generated is yours.
Where people actually learn to run this
Getting one avatar video is easy. Knowing which shots to film so the swap holds up, how to keep a character consistent across a whole series, and how to charge for this instead of just playing with it, that is the part no help doc covers.
That is what we work through inside CC Strategic AI on Skool. Real builds, real client work, and the exact setups I run in my own business.
Get the free window before you start
24 hours of unlimited Higgsfield MCP usage. Sign up before Friday.
Your Email Marketing Is Landing In Spam (Here's The 4-Step Fix)
70% of my revenue comes from email. When your domain reputation drops, that revenue quietly falls into a spam folder nobody opens. Here is how to check your reputation, warm your inbox back up, and keep it healthy with two simple workflows.
If you run an online business and your email campaigns stop working overnight, it usually is not your copy. It is your domain reputation.
Email is about 70% of my revenue. So when sends start landing in spam, that is not a small problem. Over a quarter it can wipe out close to half of what the business should have made, and the worst part is nothing looks broken. Your emails still send. They just land somewhere nobody looks.
This page is the exact fix: check the damage, warm the inbox back up, then keep it clean with two workflows so it never happens again.
Get the tools (all links in one place)
Google Postmaster Tools, free, shows your real domain reputation (shortcut: gmail.com/postmaster)
Warmbox, my affiliate link, warms your inbox back up automatically
GoHighLevel, my affiliate link, the CRM I use to build the tagging workflow and the sunset policy
Free to start on the first one. The other two are what I actually pay for and run my business on.
What domain reputation actually is
Every time you send, Gmail and the other inbox providers grade you. That grade is your domain reputation, and it decides whether your next email lands in the inbox, in promotions, or in spam.
Google sorts it into four buckets: High, Medium, Low, and Bad.
Most people start at High. Then the slide happens. You keep emailing a list that does not open, does not click, and does not reply. A few people hit the spam button. Some addresses go dead and bounce. Google reads all of that as "people do not want this," and your grade drops from High to Medium to Low, and eventually to Bad.
Nobody sends you a warning. You just watch your numbers fall and assume the offer stopped working.
Step 1: Check your reputation in Google Postmaster Tools
You cannot fix what you cannot see. Start here, it is free.
Click the plus button and add your sending domain (the part after the @ in your from address).
Google gives you a TXT record. Paste it into your domain's DNS settings, wherever your domain lives, then hit verify.
Wait a day or two for data to fill in. Postmaster needs real sending volume before the charts show anything useful.
Then read three dashboards. Domain reputation is your grade. Spam rate is the percent of people hitting the spam button, and you want that under 0.1%. Authentication tells you if SPF, DKIM, and DMARC are passing, and if any of those fail, fix that first because no amount of warming saves a domain that cannot prove it is you.
Step 2: Warm the inbox back up with Warmbox
Once you know your reputation took a hit, you have to rebuild it. You do that by proving to Gmail that real people want your email. That is what warming means.
A good friend put me onto Warmbox and it is what I use. You connect the inbox you want to warm up, and it goes to work on its own.
Here is what it does behind the scenes:
Sends normal-looking emails from your inbox into a private network of 35,000+ real inboxes.
Those inboxes open your email, and a share of them reply to it, which is the strongest positive signal there is.
Anything that lands in spam gets pulled out and marked as important, so the provider learns it guessed wrong.
It ramps volume up on a schedule instead of blasting day one, because a sudden spike looks like spam by itself.
Setup is short. Create your account, connect your inbox with Google or Microsoft one-click login, or plug in SMTP and IMAP details for anything else (Yahoo, Zoho, Amazon SES, SendGrid, and most others are supported). Pick a warm-up schedule, then leave it alone.
There is also a live email checker inside that runs your DNS and blacklist status, which is worth running once before you start so you know nothing else is broken.
Warming is not a weekend job. Give it a few weeks of running in the background and watch the reputation line in Postmaster climb.
Step 3: Build a workflow that tags spam, unsubscribes, and bounces
Warming repairs the damage. This step stops you from causing it again.
Inside your CRM, build one workflow that watches for three events and tags the contact when any of them happens:
Marked as spam. Tag them and stop every campaign immediately.
Unsubscribed. Tag them and take them out of all marketing sends.
Bounced. Tag them, because a hard bounce means the address is dead and keeps hurting you every time you hit it.
Then use those tags as an exclusion filter on every campaign you send from now on. That is the whole point. The tag is not for reporting, it is a wall.
I run this in GoHighLevel, which is a full CRM with automations, pipelines, calendars, AI tools, and email plus SMS in one place. It is how I run a chunk of my business. Build the workflow with an email-event trigger, add an if/else for the three event types, then add a tag action and a "remove from workflows" action on each branch. Any CRM that can trigger on email events can do the same thing.
Step 4: Add a sunset policy
This is the one most people skip, and it is the one that keeps your reputation high long term.
A sunset policy is a rule that removes dead weight from your list on a timer. Mine is simple: track who opens, who clicks, and who replies. If a contact does none of those for 90 days, they come off my campaigns.
How to build it:
Stamp a "last engaged" date on the contact any time they open, click, or reply.
Run a daily check for contacts whose last engaged date is more than 90 days old.
Optional but worth it: send one last "are you still interested?" email before you cut them. Whoever clicks gets to stay.
Everyone else gets tagged as unengaged and removed from ongoing campaigns.
It feels backwards to shrink your list on purpose. It is not. A list of 2,000 people who actually read your email beats a list of 20,000 where 18,000 ignore you, because those 18,000 are the reason the 2,000 stop seeing you at all.
What to expect after you turn this on
Nothing changes on day one, so do not panic. Reputation is a trailing score, it moves on weeks of behavior, not hours.
Week 1. Warming is running and your bad contacts are getting tagged out. Postmaster still looks the same.
Weeks 2 to 4. Spam rate starts dropping because you stopped sending to people who report you. Reputation begins to move.
Weeks 4 to 8. Low climbs toward Medium, Medium toward High. Open rates rise before revenue does, because more of your sends are being seen at all.
One rule while you rebuild: do not go back to blasting the whole list to make up for lost time. That is what got you here.
Where people actually learn to run this
The tools are the easy part. Knowing when a list is worth saving, how hard to ramp your sends, and what to write so people reply instead of report you, that is the part nobody puts in a help doc.
That is what we work through inside CC Strategic AI on Skool. Real campaigns, real numbers, and the workflows I run in my own business.
Why Your Claude Code Websites Look Like Slop (And The 4 Tools That Fix It)
Claude can write the code. It cannot guess what good looks like. These four free tools hand it a real design system, real animations, and a final quality check, so what you ship stops looking like every other AI site.
Why the sites come out looking like slop
You tell Claude Code to build you a site. It builds one. And it looks like every other AI site on the internet. Same purple gradient. Same rounded cards. Same lifeless buttons.
That's not Claude being bad at its job. It's Claude filling in a blank you left. You gave it a goal, not a design. So it reached for the average of everything it has ever seen, and the average is slop.
The fix is not a better prompt. It's giving Claude three things it cannot invent on its own: a design system to copy, animations to copy, and a judge that tells it when the work is still ugly. Four free tools cover all three.
Tool 1: getdesign.md, for the design system
The hardest part of building a site is knowing where to start with the look. getdesign.md removes that problem completely.
It's a library of real brand design systems, broken down into a file your AI can actually read. Apple. Stripe. Airbnb. Linear. Vercel. Nintendo. You can even grab Claude's own design system if you want.
Each one gives you the fonts, the colors, the spacing, the button styles, all of it. Pick the brand whose vibe you want, hit copy, and paste it into Claude before you ask for a single page.
Now Claude isn't guessing what good looks like. It's copying homework from a company that paid a design team six figures to get it right.
Tool 2: 21st.dev, for the animations
Describing an animation in words is miserable. "Make it fade up but kind of springy and staggered" gets you something you didn't want, three times in a row.
21st.dev fixes that. It's a community library of React components and marketing blocks, and here's the part that matters: every single one ships as a prompt. You browse until you see an animation you like, click copy, and paste it straight to Claude. It rebuilds that exact thing in your codebase.
Animated heroes, buttons, cards, navbars, shader backgrounds, gradients. Thousands of them. Everything is React plus Tailwind following shadcn/ui conventions, so the code lands in your repo and it's yours to edit.
Free accounts get 2 component copies a day, which is plenty when you're pulling one hero and one nav for a build.
Tools 3 and 4: Impeccable and Taste, for killing the slop
These last two do the same job from different angles. They're skills you hand to Claude, and they teach it to spot its own bad taste.
Impeccable runs 58+ checks built to catch the exact defaults AI reaches for, then gives you 23 commands to fix what it finds. /polish, /audit, and /typeset are the ones you'll live in. It respects a design system you already have instead of steamrolling it, and there's a Chrome extension that overlays the problems right on the page.
Taste is open source and calls itself the anti-slop framework for AI agents. Version 2 added brief inference, which means the agent reads the room before it generates anything: your industry, your audience, the mood, how much motion fits, what layout family makes sense.
Pick either one. Or stack both, which is what I do. Impeccable catches the pixel-level tells, Taste catches the "this whole direction is wrong" problems. They don't fight each other.
How to install all four in about 5 minutes
The two websites need no install. Just open them and copy.
getdesign.md and 21st.dev: browse, pick, click copy, paste into Claude.
Impeccable: run npx impeccable install in your terminal. It also installs through Claude Code's /plugin marketplace, which is the cleaner path if you're already using plugins.
Taste: run npx skills add Leonxlnx/taste-skill.
Both skills work outside Claude Code too, in Cursor, Codex, Gemini CLI, and anything else that reads SKILL.md files.
The order you run them in
Stacking matters more than any single tool here. Run it like this:
1. getdesign.md gives Claude the design system before it writes any code.
2. 21st.dev gives it the exact animations instead of your best guess at describing them.
3. Impeccable or Taste (or both) run last, as the quality gate that catches what still looks cheap.
Design in. Motion in. Slop out. That's the whole system.
Where people actually learn to run this
Tools are the easy part. Knowing which brand system fits the client in front of you, when to overrule the skill's opinion, and how to ship the thing without a week of fiddling, that's the part nobody puts in a README.
That's what we work through inside CC Strategic AI on Skool. Real builds, real critiques on real client work, and the prompts and setups I use daily. If you want to actually master Claude Code instead of collecting tools you never open again, that's the room.
Not sure what to fix first?
Most people don't need another AI tool. They need to know which thing is actually costing them the most.
The free AI Bottleneck Finder tells you the one constraint capping your revenue right now, plus two builds you could ship this week. Five minutes, free, no call required.
Higgsfield Just Went Free: Every Top AI Model, Unlimited, For 24 Hours
For one day only, Higgsfield opened unlimited free access to every top image, video, and audio model, including Seedance 2.0. I used it to make five cinematic real estate tours that cost me $0 and sold for $1,700. Here's the offer, exactly how it works, and how to run it up before it's gone.
Available for new users. When it's gone, it's gone.
What just happened
Higgsfield opened up unlimited free access to its entire platform for 24 hours. Every top image, video, and audio model, free, run as much as you want. That includes Seedance 2.0, the state-of-the-art video model everyone in AI is talking about.
Every AI platform trains you to hold back. You count credits. You watch the balance. You skip the tenth version of a shot because each one costs. Or worse, you never learn AI at all because it feels too expensive to even try. For one day, that problem is gone. This is the first time Higgsfield has done this, it is one day only, and when it's gone, it's gone.
What I did with it (a real $1,700 day)
I wasted no time. I made five cinematic property tours for a real estate agent. All five cost me absolutely nothing to make, because Higgsfield was free. I sold each one for $350. That's a little over $1,700 in a day with zero overhead.
The whole thing ran on two free tools: the free Higgsfield access plus the free RE Walkthrough Pro skill for Claude Code, which turns a listing into a room-by-room cinematic tour. If you want the app version that one-shots the whole pipeline, that's Listing Cinema. Point is: unlimited free access is not a toy. It's a day where the tool cost drops to zero and only your hustle is left.
How the free offer works
Read this so nothing surprises you. It's simple and you're in control:
Day one is a full day of unlimited access, free. All the top models, no credits charged, before any subscription begins.
After 24 hours it continues into a Higgsfield Plus subscription that includes 7 more days of unlimited access. Same models, still no credit counting. Treat the extra week as the bonus for staying.
You can cancel in one click during the first day if you don't want to continue. No hoops.
So the move is: start free today, run everything you can, and decide inside the first day whether the extra unlimited week is worth keeping. This offer is for new users.
Every model you get free
All of these are unlimited during the offer. The headline is Seedance 2.0, but the whole shelf is open:
If you're new to AI: this is the day you finally learn it without burning credits on every mistake. Make the ugly first version, then the good tenth version, and pay nothing for the practice.
If you're a pro: run the volume real work actually takes. Countless versions of a shot, a whole campaign in a day, no watching a balance. This is when you clear the backlog you've been rationing credits on.
About Higgsfield
Higgsfield is an AI-native creative suite: one workspace that turns an idea into a finished product. Image, video, audio, sites, apps, and interactive experiences, all in one place, from the first prompt to the final result, without switching between five different apps. The free window is the cheapest possible way to find out what it can do for your work.
Start now, one day only
This isn't a standing offer. It's open for one day, and then it's gone. The people who wait miss it entirely. Start your free unlimited access, make the thing you've been putting off because it cost too much to try, and if the extra week is worth it, stay.
This one is optional, and only worth doing after you've started your free Higgsfield access above. Once you're signed up with the unlimited access, you can plug Higgsfield straight into Claude Code with the Higgsfield MCP. That lets you generate images and videos by talking to Claude instead of clicking around the dashboard.
It's a nice extra if you live in Claude Code like I do, and it is not required to use the free offer. Sign up for the free unlimited access first, then if you want this add-on: add the Higgsfield MCP for Claude Code.
Free live training: turn this into $500+ deals
Free access is the tool. Turning it into money is the skill. I run a free live training that walks through my proprietary Claude system for generating cinematic home tours and actually selling them, the same play behind that $1,700 day.
Here's what we cover:
The video generating system, the exact Claude + Higgsfield setup that makes the tours.
How to pick and source leads, real estate agents, golf course owners, and Airbnb hosts.
How to market to them so the videos actually sell.
Run Claude Code For Free With OmniRoute (The OpenRouter Method, Step By Step)
There's a tool going around that lets you keep using Claude Code after you hit your limit, for free, by routing through other companies' free models. Here's the honest, tested, beginner walkthrough: how to set it up, and the one thing nobody tells you about what's really answering your prompts.
What OmniRoute actually does (read this first)
Claude Code normally talks straight to Anthropic, and you pay for every message. OmniRoute is a small program that runs on your own computer and sits between Claude Code and the internet. Instead of sending your messages to Anthropic, it can send them to other AI companies that have free plans. When one free plan runs out of gas for the day, OmniRoute automatically hops to the next one. That is how people are running Claude Code "for free."
The honest truth nobody tells you: the free version is not Claude answering you. When I tested this myself, the model actually doing the work was Xiaomi's MiMo, a free Chinese model, even though Claude Code's screen still said "Claude." Free models are slower, weaker, and sometimes hand back an empty answer.
So here is the rule: this is great for practice, learning, and building throwaway stuff at 2am without watching a meter. Do not run client work or anything private through it. Your messages pass through other companies' free servers, and free servers usually get to keep and train on what you send. Use it on your own projects only. Okay, here is the whole thing, start to finish.
Before you start
Node.js installed (version 22 or 24). If you can run node --version and see a number, you are good.
A free OpenRouter account. We make one in Step 3. No credit card needed.
Claude Code already installed and working.
Step 1 and 2: Install it and turn it on
Open your terminal and paste this one line:
npm install -g omniroute
Give it a minute. That puts the omniroute command on your computer. Now start it:
omniroute
You will see a little logo and a line that says "OmniRoute is running!" with a web address:
Dashboard: http://localhost:20128
API Base: http://localhost:20128/v1
Leave this terminal window open. OmniRoute only works while it is running. Closing the window turns it off. Tip: you can also just tell Claude Code "open the OmniRoute dashboard" and it will pop it open for you.
Step 3 and 4: Get an OpenRouter key, add it to OmniRoute
OpenRouter is like a giant vending machine of AI models. One free key gives you access to hundreds of them, including free ones. It is the easiest single provider to start with.
Click your profile, go to Keys, and create a new key. It looks like sk-or-v1-xxxxxxxx. Copy it.
Now hand that key to OmniRoute:
In your browser, open http://localhost:20128 (the dashboard).
Open the side menu and click Providers.
Type OpenRouter in the search, scroll to the bottom, and click Add.
Paste your OpenRouter key into the box.
The important part: flip on the "free only" toggle. That tells OmniRoute to only use models that cost nothing.
Click Save.
Done. OmniRoute now has a menu of free models to pull from.
Step 5 and 6: Make your OmniRoute key, then dodge the trap
Claude Code needs a password to talk to OmniRoute. That is a separate key you make inside OmniRoute, not the OpenRouter one.
In the dashboard, go to Endpoints (or "API Keys").
Click Create key and copy it. It looks like sk-c27997f110d15f31. Keep it private; if it ever leaks, just delete it and make a new one.
Now the trap that gets almost everyone. If you are logged into your normal Claude.ai account, Claude Code ignores OmniRoute completely and quietly uses your paid account instead. You would swear it is free while it is actually billing you. Fix it. Inside Claude Code, type:
/logout
When it asks, choose to use the API key instead of claude.ai. When you are done and want your normal Claude back, just run /login again. Nothing is broken. You are only switching which door it uses.
Step 7: Point Claude Code at OmniRoute
Close Claude Code. Then open a fresh, throwaway folder and start it with these settings. Swap in your OmniRoute key from Step 5:
cd /tmp && mkdir -p omni-demo && cd omni-demo
ANTHROPIC_BASE_URL=http://localhost:20128/v1 \
ANTHROPIC_API_KEY=YOUR_OMNIROUTE_KEY \
ANTHROPIC_MODEL=auto/coding:free \
ANTHROPIC_SMALL_FAST_MODEL=auto/fast \
claude
What each line means in plain English:
ANTHROPIC_BASE_URL: "Send my messages to OmniRoute, not Anthropic." This is the actual switch.
ANTHROPIC_API_KEY: your OmniRoute password from Step 5.
ANTHROPIC_MODEL=auto/coding:free: "Use the free coding models." Without this, Claude Code asks for paid Opus by default and it breaks.
ANTHROPIC_SMALL_FAST_MODEL=auto/fast: Claude Code also fires little background helper calls. This keeps those free too.
When it boots, the top of the window will now say auto/coding:free instead of Opus. That means it worked. Type something small to test it, like build a snake game in one html file, and watch the OmniRoute dashboard light up. One more safety note: those settings only live inside that one terminal window. Close it and your normal Claude Code is exactly as it was. Nothing permanent changed.
Step 8: Prove it is really routing (optional but cool)
Want proof it is using OmniRoute and not Anthropic? Run this, with your OmniRoute key dropped in:
If you get a big list of models back, it is live. When I ran this I got 129 models. If the list is empty, no provider is connected yet, so go back to Step 4.
How to know what model is REALLY running
This part is sneaky, so pay attention. There are three places to look, and two of them lie:
Claude Code's screen shows auto/coding:free. That is just the nickname you asked for, not the real model.
Asking the AI "what model are you" usually just makes it repeat the nickname. It genuinely does not know.
The OmniRoute dashboard, Logs tab. This one is the truth. It shows the real model that answered each message. When I checked, it said mimo-auto, which is Xiaomi's MiMo.
So if you ever want to know who is actually reading your code, open the Logs tab. Nowhere else tells you straight.
The real limits (so you are not surprised)
"Unlimited" is a stretch. OpenRouter's free plan is about 50 messages a day, then it slows you down. Adding $10 of credit bumps it to around 1,000 a day.
To get closer to always-on, connect a few more free providers the same way, so OmniRoute has more places to hop when one runs dry. Easy free ones: Groq, Google AI Studio (Gemini free), and Cerebras.
Quality is the trade. Everything free is MiMo, DeepSeek, or Gemini-Flash level. Good enough to learn and mess around. Not Claude Opus. If a build comes out messy or an answer comes back empty, that is the free model, not your setup.
Free models get rate-limited a lot. A "429" error just means "too many requests, slow down." Totally normal, and it is exactly why the auto-hop feature exists.
Quick fixes
It still says "Opus" and bills me: you did not log out of claude.ai. Redo Step 6.
"model not found" error: try ANTHROPIC_MODEL=auto/best-free or auto/coding instead.
Empty answers: the free model choked. Ask again, or add another provider like Groq.
The models list is empty: no provider connected. Redo Step 4.
Dashboard will not open: OmniRoute is not running. Run omniroute again.
Bottom line, and where to go next
OmniRoute is a real, clever tool. It genuinely lets you keep coding after you hit your Claude limit, for free, by relay-racing across other companies' free models. Just go in with your eyes open: it is free models pretending to be Claude, not Claude itself, and it is for your own projects, never client or private work. Use it to practice, to learn, to build throwaway stuff. That is exactly what it is good for.
When you are ready to do the real work well, the smart move is spending less on the paid models, not dodging them. The Save Tokens in Claude Code guide on this page cuts your bill without swapping in weaker models, and the Fable 5 YES Playbook shows how to make the most powerful model pay for itself. Want the prompts, the routing setups, and a community shipping this stuff daily? Join CC Strategic AI on Skool.
The Fable 5 YES Playbook: How To Actually Use The Most Powerful Claude Model
Fable 5 is the strongest model you can touch, and it burns tokens fast. Most people torch a week of usage by Monday. This is the YES list: the five things to do so Fable pays for itself instead of draining you.
TLDR
Fable 5 is the most capable model Anthropic ships, and it costs a lot more than Opus or Sonnet to run. That is the whole tension. Point it at the wrong work and you will hit your limit by Monday with nothing to show. Point it at the right work and it does things no cheaper model can touch.
There is a video going around where I list the five things you should NOT do with Fable 5. This is the flip side. Same five ideas, but here is exactly what to do instead, in plain steps, with the prompts. Nail these and Fable stops being an expensive habit and starts being your unfair advantage.
One line to remember: the top results in AI don't come from the best model. They come from the best harness around it. Fable is the sharpest tool in the box. This playbook is how you hold it.
The one idea behind all five
Every model has a job it is best at. Fable 5 is your architect and your judge. It is brilliant at the hard thinking: planning a real system, one-shotting a build that is beyond a smaller model, and reviewing work with a sharp eye.
It is not your day laborer. The cheaper models, Opus 4.8 and Sonnet 5, are perfect for the grind: test loops, small edits, retries, brainstorming. So the whole playbook comes down to one habit. Match the model to the job, not to the hype. Use the most expensive model only where a cheaper one would give you a worse answer.
The test is one question: would Sonnet's output actually be worse here? If no, route the task down to Sonnet. If yes, that is a real Fable job.
YES #1: Make Fable the architect, not the grunt
Don't hand Fable your test loops, tiny CSS tweaks, and retry cycles. That is Sonnet work. Hand Fable the thing only it can do well: the whole build, in one shot.
Here is the clean order that saves the most money:
Ideate with Opus or Sonnet first. This is the messy part where you don't know what you want yet. Should it be a dashboard or a CLI? What is the data model? Waste all the cheap turns you want here. This is deciding WHAT to build.
Then turn on Fable to lock the plan and build. Once you know the what, Fable interviews you (see YES #5), writes the plan, and one-shots the build. This is the HOW, and it is where Fable earns its cost.
Hand the grind back to Sonnet. After the build exists, the maintenance, the small fixes, the test loops all go back to the cheaper model.
Think of it like a job site. You brainstorm with the cheap friend. You brief the expensive contractor. Using Fable 5 to run test loops is a Ferrari doing DoorDash. Let it build the whole thing, then step it back.
YES #2: Run it on medium, save max for the judge
Effort is the dial most people leave cranked to the top and never touch. That is why their week's usage dies on Monday. Higher effort means more thinking tokens per turn, and on Fable those tokens are expensive.
The move: start on medium. Fable on medium already beats Opus on xhigh for most work. Only climb the effort dial when the quality genuinely isn't there yet. And save max for one thing: the final judgment pass on a finished build.
Medium — your default for almost everything. Fast, cheap, still frontier-level.
High — climb here only when medium's answer isn't good enough.
Max — reserve it for the final review of an important build, not the daily work.
You change it on the fly with /model and the effort setting mid-session. For the deeper token math, the Save Tokens in Claude Code guide on this page breaks down the exact settings that cut your bill.
YES #3: Point it at assets, not toys
Building a snake game with Fable 5 costs the same tokens as building something that runs your business for the next six months. One is a toy you'll never open again. One is an asset. Spend the expensive model on the asset.
The best Fable jobs are the ones that keep paying you back long after the session ends. Point it at compounding work:
A CLAUDE.md audit. Have Fable review and tighten your CLAUDE.md files into a brief a cheaper model can follow without you in the room. Now every future session with Sonnet inherits Fable's judgment.
A skills audit. Have it look at your skills, kill the redundant ones, sharpen the keepers, and spot new ones worth building. See the 4 primitives guide if skills, MCPs, and subagents still blur together.
A PRD or roadmap. Fable excels at planning. Let it write the full implementation plan, then hand that plan to a cheaper model to execute task by task.
See the pattern? Each of these captures Fable's brainpower into something your cheap models reuse for months. That is how you get frontier-level results on a Sonnet budget.
YES #4: Bring in a second set of eyes
Don't let Fable grade its own homework. Every model, Fable included, is a soft grader on its own work. Ask it if the code is good and it will tell you it's beautiful. You need a reviewer that didn't write the thing.
This is called LLM as a judge, and almost nobody runs it. It is the biggest quality unlock most people are leaving on the table. Two easy ways to do it:
Run a code review with a different model. Something like Codex can review what Fable built. A neutral second model catches what the builder can't see, because it has no ego in the code.
Use a blind subagent. Spin up a fresh reviewer that never saw the build get made, hand it only the finished result, and ask it to find what's wrong. No memory of the work means no bias toward it.
Multiple perspectives is always the key. If you want to wire this checking into a loop that runs on its own, the Loop Engineering Playbook shows how the maker-and-checker split becomes a self-running system.
YES #5: Make it interview you before it writes a line
This one is the sequel to YES #1. A lazy prompt in means a confident, wrong build out, and on Fable's burn rate that mistake is expensive. The one-shot build only works because a good brief sits behind it.
So before Fable writes any code, make it interview you. After you've ideated the WHAT with Opus or Sonnet, turn Fable on and run a prompt like this:
Interview me before you write any code. Ask me one question at a time until you fully understand the build and can confidently build me the perfect system. Do not write anything until you're sure.
Now Fable asks the questions instead of guessing the answers. Auth provider? Where is it hosted? What happens on failure? You answer, and like magic it stops guessing and builds the exact thing you pictured, the one you've been trying to get right for months. The interview writes the brief. Fable one-shots the brief.
The Fable 5 cheat sheet
Tape this to the wall. Five yeses, one habit: match the model to the job.
Architect, not grunt — Fable plans and one-shots the build. Sonnet runs the grind.
Medium by default — climb only when quality demands it. Max is for the judge pass.
Assets, not toys — CLAUDE.md audits, skill audits, PRDs. Things cheaper models inherit for months.
Second set of eyes — Codex or a blind subagent reviews. Never let it grade itself.
Interview first — make it ask questions before it writes a line.
The through-line: brainstorm with the cheap friend, brief the expensive contractor, and always get the work reviewed by someone who didn't build it.
Go deeper
Once these five click, Fable stops being a scary bill and starts being the sharpest tool in your kit. The next moves are on this page: cut your spend with the Save Tokens in Claude Code guide, get the four primitives straight in Skills vs MCPs vs Subagents, and wire it all into a self-running system with the Loop Engineering Playbook.
Want the exact prompts, the routing table, the judge prompt, and a community shipping this exact stack? Join CC Strategic AI on Skool. We break down the model and effort routing, the SEED and PAUL workflow, and the review loops already inside.
Zillow Link to Cinematic Tour in Minutes: One-Shot an AI Video App With Fable 5 + Higgsfield MCP
Claude can now create full cinematic property tours from nothing but a Zillow link. I one-shotted the entire app with Fable 5 and the Higgsfield MCP, Higgsfield hosts it for me, and you can try it right now.
Claude can now create full cinematic tours, and I can't believe I'm saying this, but from Zillow link to video in minutes. And what's even crazier: the property in the demo video doesn't exist yet. It's a pre-construction listing, and the app generates the interiors anyway.
Using Fable 5 and the Higgsfield MCP server, I was able to one-shot an application build that does this whole process for me. And Higgsfield will host the app for you, so there's no deploy step, no server, no domain wrangling.
Try the app I built
The app is called Listing Cinema. Paste a Zillow link, get a cinematic room-by-room tour video back. It's live and hosted by Higgsfield right now:
Here's the part you shouldn't sleep on. If you build your own app through Higgsfield over the next 12 days and it absolutely slaps, you can enter their contest and win up to $100,000. You're going to build the app with one prompt anyway. Submitting it is free money if it hits.
Build your own app in 4 steps
Download the Higgsfield MCP server. Grab the install command from higgsfield.ai/mcp (my link, it also gets you the free MCP trial plan), copy the command, and just give it to Claude to install.
Type /model and make sure you're running Fable 5 before Anthropic decides to take it away. This is the model that one-shots the whole build.
Tell Claude to build the app using the Higgsfield MCP. Voice-drop ALL the details about your application into the first prompt: what it does, what the user pastes in, what comes out, the vibe of the UI. One dense prompt beats ten follow-ups.
Grab your link. Claude builds it, Higgsfield deploys it, and it drops the link to the fully deployed application for you. Your AI image/video application is completely done with a single prompt.
Don't forget to submit your application to enter the contest.
Save Up To 50% On Tokens: 3 Claude Code Hacks (/doctor, an Effort Table, /model)
Claude Code can burn through your whole subscription in one session. It does not have to. Three simple moves, already built into the tool, that cut your token bill without making Claude any dumber.
Claude Code can easily spend your entire subscription in one session if you let it. But that is only true if you do not know these three simple hacks. Done together, they can save you up to 50% on your tokens, and none of them make Claude worse at the actual work.
Quick plain-English primer first. A token is a chunk of text, roughly three quarters of a word. Every word Claude reads and every word it writes is counted in tokens, and tokens are what you are paying for. So "saving tokens" just means: stop making Claude read and think about stuff it does not need to. Less clutter in, less wasted thinking, lower bill. That is the whole game.
Hack 1: Run /doctor (the health check already baked in)
The easiest fix is a command that already ships inside Claude Code. You do not install anything. Just open Claude Code and type:
/doctor
It runs a full health check on your setup and shows you exactly what is eating tokens that you do not need to be paying for. Think of it like a mechanic putting your car on the lift: it finds where the clutter is (old skills, MCP servers you forgot you connected, hooks, bloated config) and tells you what you can safely remove.
Run it in a fresh session. Type /doctor before you start real work so you see a clean picture.
Read the warnings top to bottom. Anything flagged is a token cost you are paying every single message, whether you use it or not.
Trim what you do not use. A connected MCP server or skill you never touch still loads its instructions into context. Removing three dead ones can be a real chunk of your bill.
If you want to actually watch these numbers live while you work, pair /doctor with my Status Line Setup, which puts a running token breakdown at the bottom of every session.
Hack 2: Add an effort table to your claude.md
Your claude.md file is a plain text file that Claude reads at the start of every session. It is where you write your rules, your preferences, and how you want Claude to work. One of the highest-leverage things you can put in there is an effort table: a small table that tells Claude which model and which effort level to use depending on the task you give it.
Two dials to understand, at a 9th grade level:
Model is which brain. Sonnet is the fast, cheaper brain, great for daily and simpler work. Opus is the stronger, pricier brain, better for the big-ticket stuff like architecture and hard debugging.
Effort is how hard the brain thinks before it answers. More effort means more thinking tokens, which means a bigger bill. I recommend keeping everything on high, because anything above that (like max) burns a lot of tokens for a small gain on most tasks.
The effort table you can paste today
Drop this table into your claude.md. It is a starting point. Change the rows to match the work you actually do.
## Model & Effort Routing
| Task | Model | Effort | | --- | --- | --- | | Search, quick edits, writing, simple fixes | sonnet | high | | Building an app, web design, new feature | opus | high | | Architecture, hard debugging, security review | opus | max |
In plain English, that table says:
Small or routine job? Use Sonnet on high. Do not pay Opus prices to rename a variable or draft an email.
Real build? Use Opus on high. You want the stronger brain, but high effort is plenty.
The genuinely hard stuff? Only then reach for Opus on max, and only for the task that needs it.
The whole point: you stop accidentally running your biggest, most expensive brain on tiny tasks all day. That misallocation is where most people quietly bleed their subscription.
Let Claude build the table for you
You do not have to guess at the rows. Open Claude Code in your project and paste a prompt like this:
Look at the kinds of tasks I actually run in this project. Then write me a Model & Effort Routing table for my claude.md. Use Sonnet for simple and daily work, Opus for big builds and web design, and only Opus on max effort for architecture or hard debugging. Keep everything else on high effort. Explain each row in one sentence so I understand why.
Claude will read your setup, propose a table tuned to your work, and walk you through it. Tweak it, paste it into claude.md, done. You can customize this however you like, this is just my default.
Hack 3: Override on the fly with /model
The effort table is your default. But sometimes you want to change the model or effort right now, mid-session, without touching your file. That is what this command is for:
/model
Run it and you get full control over the model and effort level Claude is using at any given moment. Just in case you ever want to override what your claude.md table says, you totally can, for a single task, then switch back.
Bump down to save. Cruising through simple edits? Drop to Sonnet for the stretch and pocket the difference.
Bump up for one hard thing. Hit a nasty bug? Jump to Opus on max just for that, then /model back down so you are not paying max effort for the rest of the session.
It is live. No restart. The change applies to your next message.
Go further: a knowledge graph is the biggest token save of all
The three hacks above stop waste. If you want to go a level deeper and cut tokens on how Claude searches your project, build it a knowledge graph. Instead of Claude re-reading your whole codebase every session to find something, it queries a map it built once. The creator measured roughly 70x fewer tokens per search.
Graphify in the toolkit (the skill that builds the knowledge graph, right here on this page)
Run /doctor to find and remove the clutter you are paying for every message.
Add an effort table to claude.md so Sonnet handles daily work and Opus is saved for the heavy lifts, all on high effort.
Use /model to override that default live, up for one hard task, down for a stretch of easy ones.
Add Graphify when you want the biggest structural win: 70x cheaper searches through a knowledge graph.
Skills vs MCPs vs Subagents vs claude.md: The 4 Claude Code Primitives Explained
If you can't say what skills, MCPs, subagents, and claude.md each do in one sentence, you're paying for four tools and using none of them right. Here's the clean mental model, plus the one-question test that tells you which to reach for.
TLDR
Most people jam everything into one giant claude.md, bolt on ten MCP servers, and hoard hundreds of skills off GitHub they never run. That unintentional bloat is what burns your tokens and makes Claude feel dumb. Claude Code gives you four components with four very different jobs. Master which is which and you stop guessing.
The whole thing collapses into four words: knowledge is claude.md, repeatable action is a skill, another app is an MCP, and too big for one context window is a subagent.
The one-sentence test
Before you install anything, answer one question about what you're trying to do. The answer picks the primitive for you.
Is it something Claude should just know, every single chat? That's claude.md
Is it a task you'll run over and over? That's a skill
Does it need to reach into another app or live data? That's an MCP
Is it too big to finish in one clean conversation? That's a subagent
If you can answer that question, you never have to guess which tool to reach for again.
claude.md — your memory (knowledge)
Your claude.md file is memory. It's the stuff Claude should carry into every conversation you have with it. Think rules, facts, and context about you and how you work.
This is not the place for a task, a workflow, or a giant dump of everything you might ever need. It's the always-on baseline. Who you are, your standards, the conventions Claude should follow by default. The moment it turns into a novel, you've turned your memory into bloat that gets read on every single turn.
Use it for: your name and role, your writing voice, hard rules ("never do X"), the handful of facts Claude needs every time.
Skills — do this task (repeatable action)
A skill is a repeatable task you trigger on demand. Package it once, run it forever. Think "do this task."
The value is that you capture a process one time, with your steps and your standards baked in, and then fire it whenever you need it. You stop re-explaining the same job every session. The skill remembers how the task is done so you don't have to.
Use it for: anything you keep re-typing. Formatting a report, drafting a reply in your voice, building a carousel, running a checklist. If you've explained the same task twice, it should be a skill.
MCPs — plug into another app (live data)
An MCP plugs Claude into another app or live data. Your Gmail, your video editor, your CRM. With an MCP connected, Claude can actually do things inside those external tools instead of just talking about them.
This is the wire between Claude and the rest of your stack. No MCP, and Claude is boxed inside the chat. With one, it reads your live inbox, moves a deal in your CRM, or edits a video. The catch: every MCP you connect adds tool definitions Claude has to carry. Ten servers you never use is ten servers of dead weight on every turn.
Use it for: reaching real, live systems. Email, calendar, CRM, analytics, a video editor, the live web. Connect the ones you actually use and skip the rest.
Subagents — too big for one context window
Claude uses subagents naturally, for tasks that are too big for one conversation. It splits the work into parallel workers so your main context window stays clean.
You don't usually summon these by hand the way you trigger a skill. When a job is large or has independent parts, Claude farms pieces out to fresh workers, each with its own clean context, then pulls back only the results. Your main thread never gets clogged with the noise of the sub-work. That's why a big research or refactor task doesn't blow up your session.
Use it for: big multi-part jobs. Auditing a whole codebase, researching across dozens of sources, fanning a task out so several workers run at once while your main chat stays focused.
The cheat sheet
Tape this to the wall. Four primitives, four jobs, zero overlap.
Knowledge — claude.md. Stuff Claude should know every chat
Repeatable action — skill. A task you package once and run forever
Another app or live data — MCP. The wire into Gmail, your CRM, your editor
Too big for one context window — subagent. Parallel workers that keep your main thread clean
Why the bloat makes Claude feel dumb
Here's the part nobody tells you. Everything in your claude.md, every connected MCP, and every loaded skill gets carried in Claude's context on every turn. That context is finite. Stuff it with rules you don't need, servers you never call, and skills you'll never run, and you've spent your budget before Claude even reads your actual question.
That's why an over-loaded setup feels slower and dumber than a lean one. The fix isn't more tools, it's the right tool in the right slot. Put knowledge in memory, actions in skills, connections in MCPs, and let big jobs go to subagents. Lean setup, sharp Claude.
Go deeper
Once these four click, Claude stops feeling like a chatbot and starts acting like a real team. The next move is wiring them into loops that run on their own. Start with the Loop Engineering Playbook, then build your own agentic OS.
The Loop Engineering Playbook: Stop Prompting, Start Building Loops
Prompting is dead. The guy who built Claude Code stopped doing it. Here's the shift to loop engineering, what a loop actually is, and the exact 5 steps to build your first one.
TLDR
For three years the skill was wording. You typed a prompt, you got an answer, you typed again. That era is closing. Boris Cherny, the guy who built Claude Code, said it flat out: he doesn't prompt Claude anymore. He builds loops that prompt Claude for him. His words: "My job is to write loops."
This is the playbook for making that jump. The three eras of talking to AI, what a loop actually is, the six parts that turn a chatbot into a self-running machine, and the exact 5 steps to ship your first loop this week.
The three eras of talking to AI
Every wave moved the skill up a layer. You didn't get worse at the old thing. The leverage point moved.
Era 1, Prompting (2023). ChatGPT goes mainstream. You type, it answers. One turn. The whole skill was the wording. You are the typist, in the loop every single turn.
Era 2, Prompt engineering (2024 to 2025). Few-shot, chain-of-thought, system prompts, context stuffing. Real craft. But it's still you steering the wheel on every turn.
Era 3, Loop engineering (2026). You stop steering. You build the machine that steers. The agent finds the work, does it, checks itself, decides the next move, and repeats. Unattended. You go from driver to factory designer.
Here's the trap. If you're still prompting turn by turn in 2026, you're not behind on a trick. You're behind on the whole way work gets done.
What a loop actually is
A loop is a recursive goal. You define a purpose, the system iterates until it's done. Prompting optimizes one turn. Loop engineering optimizes the whole system.
The difference that matters: a prompt stops the second you stop typing. A loop runs on a schedule, checks its own work, and keeps a memory so it picks up where it left off. One message equals one output. A loop equals output that scales past your hours.
The six parts of a loop
This is what turns a chat box into a self-running machine. Miss the last two and you don't have a loop, you have a money fire.
Automations — scheduled triggers that kick the loop off on their own
Separate workspaces — isolated lanes so multiple agents work at once without colliding on the same files
Skills — your conventions and context written down, so the agent acts like you instead of guessing
Connectors (MCPs) — the wires into your real tools: Slack, your CRM, the live web
Sub-agents — one agent makes, another verifies. Maker split from checker. This is the heart of trust
State and memory — external notes the loop writes to, because the model forgets between runs
Your first loop in 5 steps
Don't start by leaving it running 24/7. That's how people burn money. Right order: make it work once by hand, then make it check itself, then let it run.
1. Pick a boring, repeating task. Not your hardest problem, your most repeated one. Inbox triage. Draft replies to new leads. A daily sales summary
2. Write the goal with an exit. Bad: "improve my content" (runs forever). Good: "read the 5 newest comments, draft 1 reply each, stop." A loop needs a finish line
3. Do it once by hand in Claude Code. Prompt the whole task manually and watch where it guesses wrong. That tells you what context it's missing
4. Teach it a skill and give it a checker. Write a SKILL.md with your rules and examples. Then add a second pass that asks "did this meet the goal?" before anything ships
5. Automate it last. Wrap it in a schedule that runs the loop and reports back. Keep a human gate ("draft, don't send") for the first week. Remove it only once you trust the checker
Your literal first loop: "Every morning at 8am, read my new emails, draft a reply to each lead, save the drafts but don't send, and message me a one-line summary. Stop when the drafts are done." Safe, useful, real.
Two ways to run loops
Two rungs of the same ladder. One you trigger and verify. One that triggers itself on a server you never touch.
A loop you run, PAUL. An orchestration layer for Claude Code. Plan, apply, unify, loop. Nothing is "done" until it reconciles the plan against reality and writes its own state. Pair it with SEED to turn a raw idea into a PAUL-ready plan first
A loop that runs itself, Hermes. A self-running agent on a server. Cron wakes it, it acts and verifies on its own, then reports back into Slack or Telegram. You wrote the loop once, it works while you sleep. Full setup in the Hermes 24/7 agent playbook
The one trap to avoid
The hype has a dark side called loopmaxxing, the belief that an infinite cycle eventually produces a correct answer. It doesn't. A loop with a vague goal and no checker just converts your budget into an API bill while you sleep.
Two things make a loop instead of a money fire: a clear stop condition (when is it done?) and a checker (did it actually work?). Drill those two and you have the whole skill.
Build Your Own Agentic OS (Command Center) in 3 Steps
Turn Claude Code into your own command center. Give it a brain with Graphify, let PAUL and SEED build the dashboard, then host it on Railway so it runs 24/7. Not a mockup, the real stack.
TLDR
Most people use Claude Code as a chat box. It forgets everything the second you close the tab. An agentic OS fixes that. You give Claude a memory it can search, a framework that builds your tools for you, and a home on the cloud so it runs without you.
Three steps: map your data into a queryable brain with Obsidian + Graphify, point the PAUL and SEED frameworks at it to build your dashboard OS, then keep it local or ship it to GitHub and Railway so it runs 24/7.
What you're building
Your own command center. A full agentic OS that runs on your machine, built on Claude Code. Your data, your dashboard, your memory. No SaaS subscription, no rented tool, nothing that disappears when a company shuts down.
Built on Claude Code — runs on any model, you do not need the top tier
100% yours — your analytics, your skills, your knowledge graph
Real, not a mockup — this is the exact stack I run every day
The frameworks and tools you need
Four pieces. Two tools for the brain, two frameworks for the build.
Obsidian — free notes app. This is where your data lives as plain markdown
Graphify — turns your notes and files into a knowledge graph Claude can traverse. Full walkthrough and the repo are in the Graphify tutorial video (repo link in the description)
SEED + PAUL — these are the two frameworks you use to build the OS. SEED turns an idea into a spec, PAUL plans it, builds it, and reconciles it. Both are covered in the full agentic OS build video (resources and repos in the description)
Railway — where you host the finished OS so it runs 24/7
Go to railway.app and sign in with your GitHub account
Click New Project, then Deploy from GitHub repo
Pick the repo you just pushed
Open the Variables tab. Add your keys (the same ones from your .env file). The key name goes on the left, the value on the right
Railway builds it and hands you a live link. Click it. Your command center is now live 24/7, even while you sleep
Bonus: bundle, ship, reskin
Once it works for you, it works for anyone. The whole thing lives in one repo. Clone it, drop in keys, run it. Then fork it for each client or build custom versions for your team. That is how a personal OS turns into a product you can sell.
Want the full guided build?
The video above walks the full agentic OS build end to end. For the repos, deeper breakdowns, and a community shipping this exact setup, join CC Strategic AI on Skool.
Turn Claude Code Skills into Cloud-Based 24/7 Agents
Take any Claude Code skill on your laptop. Wrap it as a managed agent. Push it to Anthropic's cloud. Hit it from a Railway dashboard 24/7. No laptop required.
TLDR
You have a Claude Code skill on your machine. It dies when you close the laptop. This tutorial shows you how to wrap that skill as an Anthropic Managed Agent, deploy it to the cloud, and trigger it from a Railway-hosted dashboard from any device, 24 hours a day. The full video walkthrough is above. The written breakdown is below.
Five steps: pick a skill, install the ant CLI, wrap the skill as an agent, push it to console.anthropic.com, wire the agent endpoint to a Railway dashboard button.
Step 1: Pick or build a Claude Code skill
You need a skill to deploy. Open your terminal and look in ~/.claude/skills/ and ~/.claude/commands/. Every file in those folders is a skill or command you have already built. Pick one.
No skill yet? Use Skillsmith to scaffold one in two minutes. It is a free open source tool built by Christopher Kahler that walks you through skill creation with guided prompts.
Pick a skill that does one thing well. Daily analytics pull, lead qualifier, carousel maker, thumbnail generator, email triage. Single purpose works best for cloud agents.
Step 2: Install the ant CLI
The ant CLI is Anthropic's tool for managing cloud agents. Install it with one Homebrew command:
Verify the install with ant --help. You should see the agent, environment, session, and vault commands.
Step 3: Wrap the skill as a managed agent
Create a new folder for your agent. Inside the folder you need three things:
agent.yaml — the declarative manifest. Name your agent, set the model, point to the system prompt
prompts/system.md — the system prompt. Paste your skill's instructions here
Any extra files the skill needs to run (templates, examples, reference data)
The fastest way to scaffold this is to copy the _template/ folder from a Claude Managed Agents starter repo, then drop your skill content into the right files.
Once the folder is ready, run:
ant beta:agents create --name "my-agent-name"
The CLI uploads the agent to Anthropic's platform and returns an agent ID.
Step 4: Push to console.anthropic.com
After the create command finishes, your agent lives at console.anthropic.com under the Agents tab. You can see it, edit the prompt in the browser, and test it with a live session.
To test it from the terminal, run:
ant beta:sessions create --agent your-agent-id
This kicks off a session and gives you a session ID. You can poll the session, send messages, and read responses without ever opening Claude Code. That is your cloud agent endpoint.
Step 5: Build a Railway dashboard and wire the button
This is where it gets fun. We are going to host a simple web dashboard on Railway that hits your agent's session endpoint with a button click.
The fastest way to plan and build this dashboard is to combine two free Claude Code skills:
SEED — typed project incubator. Walk through ideation: what does the dashboard do, what buttons does it have, what does each button trigger?
PAUL — Plan Apply Unify Loop. After SEED gives you the spec, PAUL writes the actual implementation plan and applies it
Install both from GitHub, then run /seed:seed in Claude Code to start. Walk through the prompts. When the spec is ready, run /paul:plan to plan the Next.js + Railway stack, then /paul:apply to build it. The whole thing takes about an hour for a basic 1-button dashboard.
Why Railway over Vercel? Railway gives you persistent backends, easy environment variables for your ANTHROPIC_API_KEY, and one-click deployments from GitHub. Vercel is great for static sites but Railway is cleaner when your dashboard needs to make authenticated API calls to console.anthropic.com on the user's behalf.
Once the dashboard is live on Railway, point your button at POST /api/run-agent. That endpoint creates a session against your agent ID, polls until done, and returns the result. Click button. Cloud runs skill. You never open Claude Code.
Tips and tricks I learned the hard way
Start with one skill, not five. Get the loop working end to end before you add more buttons
Use the same skill prompt locally and in the cloud. If it works in your Claude Code on Tuesday, it should work in the cloud on Wednesday. Do not rewrite
Set a budget. Cloud agents run on your API key. Set a monthly cap in platform.claude.com before you wire up automation that loops
Log every session ID. When something breaks in production, the session ID is your only window into what the agent saw and did. Save them in your Railway database
Add a webhook destination. Instead of polling, give your agent a webhook URL to POST back to when the job finishes. Way cleaner for dashboards
Skip Vercel if your agent does anything async. Vercel functions timeout at 60 seconds on the free tier. Most cloud agent runs take longer. Railway has no such limit
Add a status pill to your dashboard buttons. Idle, queued, running, done, error. Users want to know what the cloud is doing
Want the full guided build?
The video above shows you the live build of the agent and dashboard end to end. For deeper breakdowns, the full Railway template repo, and a community of people shipping this exact setup, join CC Strategic AI on Skool. We have walkthroughs on managed agents, the Railway template, and the SEED + PAUL workflow already inside.
Set a goal, give Claude the context plus the tools, then walk away. Claude keeps working until the goal is achieved. Here is how I actually use it.
If you have been using Claude Code, you have heard about the new /goal command. You set a goal and Claude keeps working until it is achieved. The problem is nobody has shown a proper use case. This is what running /goal on a real task looks like.
How /goal works
Type /goal in Claude Code, then write the goal as one paragraph.
Give Claude the context it needs: the path to the project, screenshots of what is working versus broken, any tool or account it should reach for.
State the success condition inside the goal itself. The goal is locked when you hit enter.
Claude plans the work, runs the loop, calls every tool it needs, and only exits when the success condition is met.
The prompt structure I use
Problem in one line. What is happening, where, and what should happen instead.
Project path. Absolute path to the codebase so Claude does not guess.
Evidence. A screenshot of the working state, a screenshot of the broken state.
Permission to use tools. If you want Claude to verify on a real device, say so. Claude Computer Use can drive iPhone Mirroring, the browser, the desktop, whatever the goal requires.
The success condition. One sentence. The clearer the condition, the cleaner the exit.
Why this is different from a normal prompt
A normal prompt finishes when Claude responds. A goal finishes when the goal is achieved. Claude self-checks every loop, spawns subagents, takes screenshots, opens iPhone Mirroring, polls the deploy, refreshes the page, and keeps going. The coolest part: you can step away and do something completely different while this is happening.
When to reach for /goal
A bug that needs verification on a device you do not want to keep picking up.
A deploy plus production smoke test where the success signal is a real URL responding correctly.
Anything where the work spans multiple tools (code edit, build, push, browser or device test) and you do not want to babysit each step.
Tasks that would otherwise eat an afternoon of back-and-forth with you in the loop.
What you should not skip
State success in concrete terms. "Works on mobile and formats correctly" is fine. "Fix it" is not.
Hand Claude the verification path. If the only way to know it works is opening a phone, say so and give it Computer Use access up front.
Trust the loop. Do not jump in halfway. The whole point is that you do not have to.
Status Line Setup for Claude Code
Most people configure 1 folder. There are 3. This statusline shows what's actually eating your context, in real time, by category.
Claude Code is the application. Your workspace is the operating system. If the OS is garbage, no model saves you. Better prompts will not fix it. A bigger model will not fix it either.
This drops one Python file into your ~/.claude/hooks/ folder, wires three lines into settings.json, and gives you a live statusline at the bottom of every session that breaks down exactly what your context is costing you.
What you'll see at the bottom of every session
ctx:N% (Nk) — total context window usage, color-coded as it fills
cached:Nk new:Nk — prompt cache split. When new spikes, your cache broke.
skl:Nk mcp:Nk hk:Nk — token cost from each context source. Skills, MCP servers, hook outputs.
What's in the playbook PDF
3-step install (no pip, no env vars, no node modules)
Full Python script ready to paste, ~265 lines, stdlib only
Customization callouts: change the @charlieautomates tag to your handle, optional caveman badge, repo marker
Full breakdown of how to read every number on the statusline
How to diagnose a broken cache and what to trim when one category gets heavy
About the [CAVEMAN] badge
If you've seen [CAVEMAN] on the statusline in any of my posts, that's a separate plugin I run that compresses Claude's communication style. It's optional. The PDF shows you exactly how to delete the caveman block if you don't use that plugin, so the script works clean either way.
Blank canvas killing your project? Design a site in Google Stitch in minutes, then hand the output to Claude Code for production-ready code.
Full walkthrough on using Google Stitch to design a site in minutes, then hand the output to Claude Code to turn it into production-ready code. Skip the blank-canvas paralysis and ship a clean, on-brand site without learning a design tool from scratch.
Run your business from your phone. Telegram bot setup, invoice generation, Drive management, and image edits all triggered by one text message.
Anthropic released Claude Channels and I recorded the full walkthrough for you. Setup from scratch plus the use cases I'm actually using in my business right now.
What's in the video
📱 Full Telegram bot setup, step by step, nothing skipped
🧾 Generating branded invoices from my phone and sending them via email automatically
📂 Managing Google Drive folders and files without opening my laptop
🎨 Editing images on the go with the Nano Banana MCP
⚡ Stacking all of these together so one text message kicks off multiple tasks
This is not a "look what's possible" video. I set this up live on camera and ran real business tasks from my phone. If you have been looking for a way to use Claude Code without sitting at your desk, this is it.
Skip the timeline and keyframes. Prompt Claude Code plus Remotion and you get a generated video in minutes, editable by screenshot iteration.
Just dropped a breakdown on how Claude Code + Remotion can generate videos from a prompt. No timeline. No manual keyframes. Just prompt, video, minutes.
Inside the video
What it actually does (and what it does not, yet)
The exact setup and commands
How to iterate edits fast, even from a screenshot
Where this fits in a real creator workflow
Question for you: would you use this to replace parts of your editing stack, or is it still "too early"?
3,600 unread emails cleared in 5 minutes. Claude Cowork sorts your inbox, builds a label system, and routes every future email automatically.
Just tested Claude Cowork on an inbox with 3,600 unread emails and it completely cleaned the whole thing in about 5 minutes. 🤯
Not only did it clear the inbox, it also built a full label management system so every new email gets organized automatically.
Most people are still manually sorting emails or paying for tools that barely help. Meanwhile AI can literally act like a personal inbox assistant if you connect the right tools.
If you want to see exactly how I set this up step by step, watch the video. The PDF below is the full Inbox Zero playbook.
Stop styling automation docs by hand. Feed this WeasyPrint playbook to Claude and it turns plain HTML and CSS into clean, on-brand PDFs automatically.
Quick update. You can feed this asset to Claude before you start to help you. I forgot to share this earlier.
Most people waste a ton of time making automation docs look good by hand. I use WeasyPrint. It turns simple HTML and CSS into clean, professional PDFs automatically.
I made a beginner playbook to use it with Claude. Cheers 🥂
Stop renting cold lists. Stack Apify with Claude Code and the GHL MCP, and one prompt pulls thousands of local leads straight into your CRM, tagged and in a campaign before you finish your coffee.
One prompt. Thousands of qualified leads. Apify scrapes Google Maps for any niche in any city. Claude Code reads the JSON, cleans it, and pushes every contact into GoHighLevel through MCP. Tagged, segmented, in a workflow, in a campaign. No spreadsheets, no Zapier, no VAs.
What's inside the PDF
The 5-step setup, end to end (Apify account through first campaign fire)
The exact prompt to feed Claude Code so it runs the Google Maps scraper for you
The .mcp.json snippet to wire the open-source GoHighLevel MCP into Claude Code
Three real-world plays: Local Domination, Niche Outbound, Autopilot Campaigns
Pro tips on lead scoring, deliverability, and not torching your sender reputation
Tools you'll need
Apify (free to start, $0.002 to $0.005 per lead at scale). Use code CHARLES20 for 20% off, limited to the first 100 signups, so claim it before it runs out.
A lead engine is the easy part. It's the trust layer that actually converts strangers into buyers, and that trust layer is your content. Paid ads or organic, doesn't matter. The content is what makes a cold lead say yes.
Without it, this stack just generates traffic that bounces. With it, every lead has a reason to believe you before the sales call.
The 70% rule. YouTube long-form video converts leads roughly 70% better than short-form on Reels, TikTok, Shorts, or Facebook. Short-form earns attention. Long-form earns the sale. Both belong in the stack, but the buying decision happens inside the 12-minute video, not the 12-second hook.
The full conversion stack
One prompt pulling thousands of leads is useless if the rest of the funnel leaks. Every piece below has to exist or you're throwing leads into a bucket with a hole in it.
A defined offer. One thing, one outcome, one price. If you can't say it in a sentence, prospects won't buy it.
A lead intake system. A real qualifying form on a landing page, not just an email field. Capture the data you need to route the lead.
A qualification flow. Unqualified prospects get routed out before they waste a sales call. Your time is the most expensive resource in the stack.
A VSL landing page. A video sales letter that walks the offer start to finish, so by the time someone books, they already know the price, the process, and the outcome.
A nurture sequence in GoHighLevel. Most leads don't close on contact one. The GHL workflow warms them for days or weeks until they're ready.
Where this stack actually lives
Apify plus Claude Code plus the GHL MCP is the engine. The carousel of long-form videos, the VSL page, the intake form, and the GHL nurture sequence are the chassis. Bolt them together and the difference shows up in your calendar.
This is the MCP that has made me the most money. Wire GoHighLevel into Claude Code and one morning command triages thousands of conversations, tells you who is high quality, and sends the marketing for you.
This is my most used MCP server inside Claude Code, and it is the one that has made me the most money. It connects GoHighLevel, my CRM, straight into the terminal. I have thousands of conversations running at the same time, and every morning one command goes through the whole pipeline, tells me what is high quality versus low quality, and helps me triage every prospect before I start my day.
The GoHighLevel MCP is open source, free, and exposes 269 tools across contacts, opportunities, conversations, calendars, and campaigns. You install it once, point Claude Code at your account, and from then on you talk to your CRM in plain English.
TL;DR
What it is. An open-source MCP server (mastanley13/GoHighLevel-MCP) that gives Claude Code 269 GoHighLevel tools.
What you need. A GoHighLevel account, a Private Integration token, and your Location ID. Node 18+ on your machine.
Setup time. About 10 minutes: create the token, build the server, paste two lines into .mcp.json.
What I do with it. Triage my lead pipeline and send my marketing, every morning, from one command.
Cost. The MCP is free. GoHighLevel has a free trial through my link below.
Why I run my CRM through Claude Code
Most people log into their CRM, click through tabs, and eyeball their pipeline one lead at a time. That does not scale past a few dozen conversations. I have thousands.
When the CRM lives inside Claude Code, the work changes. I do not click. I ask. "Pull every fresh lead, score them high or low quality, and tell me who is worth my time today." Claude reads the pipeline, reads each contact, reads what they actually wrote in my intake form, and hands me a ranked list. Then it drafts the outreach in my voice and waits for my yes before it sends anything.
That is the unlock. The CRM stops being a place you visit and becomes an operator you direct.
A GoHighLevel Private Integration token (starts with pit-).
Your Location ID.
Claude Code installed and working.
Step 1: Get your GoHighLevel credentials
You need two things from inside GoHighLevel. A Private Integration token and a Location ID. Do not use a normal API key, it will not work.
Log in to GoHighLevel.
Go to Settings → Integrations → Private Integrations.
Create a new Private Integration. Name it something like Claude Code MCP. Leave the webhook URL blank.
Select the scopes you want Claude to use. For the pipeline work I do, turn on contacts, conversations, opportunities, calendars, and locations (read and write on each).
Save it and copy the generated token. It starts with pit-. Keep it secret, it is a password to your account.
Grab your Location ID from Settings → Company → Locations.
Step 2: Install the MCP server
Clone the open-source server, install, and build it. One time.
git clone https://github.com/mastanley13/GoHighLevel-MCP.git cd GoHighLevel-MCP npm install npm run build
That compiles the server to dist/server.js. That file is the entrypoint Claude Code will run.
Step 3: Wire it into Claude Code
Add the server to your .mcp.json. Use the absolute path to the dist/server.js you just built, and drop in your two credentials.
The GHL_API_KEY field takes your Private Integration token, not a legacy API key. GHL_BASE_URL is optional and defaults to the GoHighLevel API endpoint, so you can leave it out.
Step 4: Confirm it's live
Restart Claude Code so it picks up the new server. Then test it with something read-only:
List my GoHighLevel pipelines and how many opportunities are in each stage.
If Claude comes back with your real pipeline names and counts, you are wired in. Now you can ask it to do anything those 269 tools cover: search contacts, move opportunities between stages, read conversations, draft and send email, book appointments.
How I actually use it every morning
I built a single command around this MCP that runs my whole lead pipeline. It works on one GoHighLevel pipeline of inbound leads, and it walks every stage with me. The shape is the same one you can build for yourself.
New leads. Fresh intake lands in the first stage. The command reads each one and pre-scores it high quality or low quality.
High quality. Real buyers with budget and a clear problem get routed to a personal-outreach queue.
Low quality. Legit responders who are not a fit yet get a lighter, batched touch instead of a one-to-one email.
Contacted. Anyone already emailed gets checked for replies, then a follow-up or an answer drafted.
Closed. Paying customers and dead leads get filed so the funnel math stays honest.
Every morning I run it, get clear on my leads in a few minutes, and start my day knowing exactly who to talk to.
Play 1: Triage the pipeline
The triage logic is just rules Claude applies while it reads. High quality means an agency or business owner above a revenue floor, or someone whose intake answer names a team, clients, or real dollar amounts. Low quality means a real human who responded but does not clear that bar yet. Spam and fake emails get filed as dead.
Claude reads the verbatim answer each lead typed into my intake form, never a paraphrase, because that one field carries the most signal. Then it hands me a ranked table: name, what they said, where it routed them, and one line on why. I approve, edit, or override. Nothing moves in the CRM without my yes.
Play 2: Send the marketing
Once leads are sorted, the same command drafts the outreach. For high-quality leads it writes a short personal email in my voice that mirrors what they wrote and points them to one next step. For the lighter bucket it batches the sends so I can review ten at a time.
Two hard rules I bake in. Draft first, always. Claude shows me every email before it sends, and only fires on my explicit "send." One clear ask. Every message ends with a single low-friction call to action, no menu of links. The MCP sends through GoHighLevel, so replies thread back into the conversation and any workflow listening for an email-sent event still fires.
Build the same pipeline yourself
You do not need my exact command. You need the pattern. Build one pipeline in GoHighLevel for inbound leads, give it clear stages from new to closed, and capture an open-ended intake question on your form. That open-ended answer is your highest-signal field.
Then point Claude Code at it through the MCP and describe the job in plain English: read the new stage, score each lead high or low, route them, draft outreach in my voice, and wait for my approval before sending. Save that as a reusable Claude Code command and you have your own morning operator.
One warning. Lead engines are the easy part. The trust layer that converts a stranger into a buyer is your content. Pair this CRM automation with real long-form video and a clear offer, or you are just moving cold leads around faster. See the Unlimited Leads playbook for the front half of this same stack.
Sick of the generic AI look on every project? Stack GSD with Anthropic's frontend design skill and ship a law-firm-grade UI that actually stands out.
In this video, I show how to level up your Claude Code projects by stacking research, planning, and building processes. Using the Get Shit Done plugin and Anthropic's frontend design skill, we skip the generic AI look and produce a professional UI that stands out.
What you'll see
Reviewing and polishing different law firm website templates
Choosing color palettes
Design enhancements to produce a clean, functional, aesthetic UI
Tips for achieving a top-notch UI on every project
Follow along for the full approach, key insights, and tips.
You think you are running agents. You are just chatting harder. This playbook shows the 5 mistakes killing your output and how to fix them.
Most people think they're using agents. They're not. They're just chatting harder.
The 5 biggest mistakes this playbook fixes
Treating Claude like a solo worker. One long session. One brain. Context rots. Output drops. The fix is orchestration, not more prompting.
Using subagents for everything. Subagents are invisible and sequential. They're fine for lookup and research. They break down when tasks interact.
Avoiding agent teams because setup feels complex. If the process matters as much as the answer, agent teams win. Parallel execution beats faster typing every time.
Letting agents overwrite each other. No task sizing. No file boundaries. Chaos. Clear scopes and directories are non-negotiable.
Staying in build mode instead of lead mode. If you're still writing code alongside agents, you're the bottleneck. Your job is coordination, review, and decisions.
Claude isn't the problem. How you run it is.
This playbook shows how to run AI like a real team, not a chat box.
Anthropic dropped a Figma killer on April 17. Most designers still think it is hype. Read the playbook, skip the noise, ship this week.
Why this playbook exists
Charles J Dove here. I wrote this playbook the week Claude Design shipped. It is what I handed the CC Strategic team on day one so nobody wastes another client call pretending Figma is still the fastest path.
Eight pages, zero fluff. The three features that matter, the exports that kill the design-to-code MCP layer, and the agency play that turns one subscription into three deliverables in an afternoon.
What you get
The 3-feature breakdown: web capture, direct handoff to Claude Code, pricing that undercuts Figma and Stitch on the same invoice.
Agency client-work play: prototype, export, approve, ship, in one afternoon.
60-second setup: how to start your first project from a text prompt, a sketch photo, a DOCX, or any live URL.
Refine and export: every way to iterate (chat, inline comments, sliders) and every way to ship (URL, Canva, PDF, standalone HTML).
Who should read it
Agency owners, freelancers, and operators who charge for design deliverables. Anyone on a Claude Pro, Max, Team, or Enterprise plan already owns this tool and should be using it this week at Charlie Automates we run every client landing page through it now.
Skip it if you do not ship design work for money.
Try it live
Open claude.ai/design in a new tab, sign in with your Claude account, and accept the research-preview disclaimer. Two panes, chat on the left, canvas on the right. Download the playbook first so you know which three features to test before the novelty wears off.
Charles J Dove here. Genspark is an all-in-one AI workspace. It runs the boring back-office side of a business from a single tab, and I did not write one line of code to set any of it up.
Think of it like a desk where every tool already lives in one place. You talk to it in plain words. It does the work. That is the whole thing.
The three jobs I use it for
Dashboards and CRM. I drop my messy spreadsheets into the chat. It does not just hold them. In about 10 minutes it builds a working CRM with a full pipeline. You do not need a tech background for this. It can stand in for a stack of paid subscriptions, not only a CRM.
Super Agent (the research one). I gave it one prompt: research three of my competitors and turn it into a comparison report I can share with my team. It ran the whole chain by itself, research, then analysis, then formatting. The work that eats your whole day was done in minutes. The report comes back clean and ready to export or send to a client.
AI Sheets. A raw list becomes a clean, CRM-ready spreadsheet with the formulas already written. Structured data without touching Excel.
Three different jobs, CRM, research agent, spreadsheet, all in one tab, no code.
Your first 10 minutes
Go to Genspark and make a free account to grab your free credits.
Open one tab. That tab is now your whole workspace.
Pick one job to start: a CRM, a research report, or a clean spreadsheet. Do not try all three at once.
For a CRM, drag your messy spreadsheets into the chat and tell it in plain words what you want. Give it about 10 minutes.
For research, hand the Super Agent one prompt, like "research my top three competitors and build a comparison report." Let it run, then export the doc.
For a spreadsheet, paste a raw list and ask for a clean, CRM-ready sheet. It writes the formulas for you.
Once one job works, replace your paid tools one at a time. That is where the savings show up.
Who this is for
Agency owners, business owners, and anyone who is tired of paying for five tools that barely talk to each other. Let us be honest, most of us are not using half the features inside the tools we already pay for.
Skip it if you love juggling tabs and monthly bills.
You want SPA feel without the SPA tax. Laravel drives the backend, Vue 3 handles UI, Inertia kills the JSON middleman.
What it does
PHP backend (Laravel) with Vue 3 frontend. Inertia bridges them so you feel like you're writing a SPA, but you're really writing server-side controllers that return "pages" like traditional server-rendered apps.
Why it's needed
Dodges the usual SPA complexity tax. No separate API to maintain, no state sync headaches, no JWT juggling. Ship features fast because the backend drives everything.
SaaS apps, dashboards, marketing sites, most "I need a modern web app" jobs. If you're hiring or your team already knows React, this is the path of least resistance.
Astro: Content-Heavy Sites / Blog Sites
Next.js is overkill for a blog. Astro ships zero JS by default and drops in React or Vue only where you actually need it.
What it does
Ships zero JavaScript by default. "Islands" architecture lets you drop in React, Vue, or Svelte components only where interactivity is actually needed. Built-in content collections for markdown-heavy workflows.
Why it's needed
Next.js is overkill for content sites. Astro is faster, lighter, and framework-agnostic in the components you do need.
Blogs, documentation sites, marketing pages, portfolios. Any site that's 80% static content with a few interactive pieces.
Tauri: Native Desktop Apps
Electron ships a whole browser per app and it is killing your binaries. Tauri uses the native OS webview so you ship 10MB, not 150MB.
What it does
Desktop app framework with a Rust backend and a web frontend (React, Vue, Svelte, whatever you already know). Uses the OS's native webview instead of bundling Chromium, which is the thing that makes Electron so heavy.
Why it's needed
Electron ships a full browser with every app. Tauri doesn't. Result: 10MB binaries instead of 150MB, less memory, better performance, stronger security model. You keep writing UI in web tech.
Desktop apps for your own workflows (dashboards, local AI agents, pipeline controllers). Cross-platform tools that need native-feeling performance. Anywhere Electron is overkill for what you're actually shipping.
TUI Applications: Ratatui / Bubble Tea / Textual
You live in the terminal, so build there. Three frameworks in Rust, Go, and Python let you ship rich keyboard-driven tools as single binaries.
What it does
Frameworks for building rich terminal interfaces with keyboard navigation, layouts, tables, forms, and real-time updates. Three picks by language:
Ratatui (Rust)
Bubble Tea (Go, from Charm)
Textual (Python, CSS-styled)
Pick based on which language you're already fluent in.
Why it's needed
Terminals are where developers live. A well-made TUI (Claude Code itself is one) is faster than a GUI for power users, works over SSH, and ships as a single binary. The three frameworks above own their respective ecosystems.
CLI tools that need more than flags and stdout — interactive installers, dashboards, wizards, REPLs, log viewers, admin tools for yourself or your team.
2 MCPs
Google Analytics + Search Console MCP
Give Claude Code live access to your Google Analytics and Search Console. Two MCP servers, fifteen-minute setup, every keyword and conversion in plain English.
I gave Claude Code access to my Google Analytics last week. The first thing it told me was that less than 3% of my traffic comes from Google search. Two of my biggest blog posts have never been clicked on from a search result. And every conversion on my site was being tracked as a zero, because I'd never told Google what a "win" looks like.
That's the honest read on charlieautomates.com from data I'd been staring at for a year and somehow missing.
This resource gives you the same view of your own business. Once it's wired in, you stop staring at dashboards. You ask questions in plain English and your AI tells you what's working, what's leaking, and what to fix this week. Setup takes about fifteen minutes once and then works forever.
What's inside
Two free MCP servers that run side by side inside Claude Code on your laptop.
analytics-mcp by Google's official analytics team. Reads everything in your GA4 property: traffic, sessions, engagement, conversions, real-time visitors, funnels. GitHub.
search-console-mcp by Charles J Dove (me). Reads everything in your Search Console: the keywords people search to find you, your rankings, your page-2 opportunities, your low-CTR fixes. MIT licensed, free. GitHub.
A ready-to-paste /seo-snapshot Claude Code skill bundled inside the search-console-mcp repo. Runs a full SEO read in one prompt.
I built the second one this week because Google's official MCP only covers Google Analytics 4, and GA4 hides organic search keywords. They're stored as "(not provided)" inside GA4. The actual keywords live in Search Console, which is a totally separate Google product with its own API. Without a Search Console MCP, your AI literally cannot see your keywords.
Open console.cloud.google.com and either pick a project you already have or click "New Project". Name it anything. I called mine mcp-analytics. Cost: zero, as long as you stay inside the free API quota. You will.
Step 2: Turn on three Google APIs (3 minutes)
Inside your new project, you need to enable three APIs. Search Console is the one most people forget.
Google Analytics Data API
Google Analytics Admin API
Google Search Console API
For each one, paste this URL pattern into your browser, replace YOUR_PROJECT_ID with your project ID, and click the big "Enable" button when it loads:
A browser window opens. Sign in with the Google account that owns your Analytics and Search Console. You will see a consent screen listing the permissions. Approve all of them.
This creates a token on your computer that lets Claude Code authenticate as you forever after. You only do this step once.
Replace YOUR_PROJECT_ID with your Google Cloud project ID. Replace yourdomain.com with your actual domain. Use sc-domain: if you have a domain property in Search Console, or paste the full https://www.yourdomain.com/ if you have a URL property.
Step 6: Restart Claude Code (2 minutes)
Close Claude Code completely. Open it back up. The MCPs load at startup, so a restart is required.
That is it. You are done. Total time including the OAuth screen, about 15 minutes.
Ten prompts to paste into Claude right now
These are the prompts I use every Monday morning. Copy and paste them straight into Claude Code.
Quick wins to start
"Show me my top 10 pages by sessions for the last 30 days, with engagement rate and conversion rate. Flag any page where engagement is above 60% but sessions are below 100. Those are hidden winners I should promote."
"Pull my top 25 organic search keywords from Search Console for the last 30 days. For each one, tell me which page is ranking and whether the position is page 1, page 2, or worse. Sort by impressions."
"What pages got organic search traffic last month but had less than 30% engagement rate? Those are pages where Google is sending visitors who do not like what they see."
The leaky-funnel scan
"Compare last 7 days to the prior 7 days. Show me which pages saw sessions drop more than 20%, and which sources sent fewer visitors. Flag anything that looks like a leak."
"For each of my top 10 traffic sources last month, what was the average engagement rate and conversion rate? If any source has high traffic but engagement under 30%, tell me what to fix."
The hidden SEO opportunities
"Find Search Console queries where my page ranks position 11 to 20 (page 2 of Google). Sort by impressions. Tell me which page is ranking for each one. These are my fastest possible SEO wins."
"Find Search Console queries with more than 50 impressions but a click-through rate under 3%. My title tag or meta description is probably weak. Tell me which page ranks for each and suggest a sharper title."
The cross-data join (the magic)
"For each of my top 10 landing pages from Google Analytics, also pull the Search Console queries that drove organic traffic to that page. Combine them so I can see this page gets X sessions from these specific keywords."
"Find pages with strong engagement in Google Analytics (above 50%) but zero impressions in Search Console. Those are good pages Google has never seen. They need internal links and a sitemap push."
The conversion truth-check
"Look at every page that received a key event (conversion) in the last 30 days. Group by page. Tell me which pages are driving 80% of my conversions. Then tell me how much traffic each one gets so I can calculate conversion rate per page."
Six walls Google will throw at you
I am going to save you a few hours by listing every wall I hit in real time.
Wall 1: Google blocks the default scopes during login. You will see a screen that says "This app is blocked." Google's default gcloud OAuth client is no longer allowed to request "sensitive" scopes like analytics.edit. The fix: create your own OAuth client inside your Google Cloud project, download the JSON, and pass it to gcloud with --client-id-file. Full walkthrough lives in the search-console-mcp repo at docs/SETUP.md. About 20% of users hit this.
Wall 2: Your terminal splits the gcloud command across multiple lines and breaks it. Zsh and Bash will sometimes wrap long commands and treat each wrapped piece as a separate command. If you see errors like zsh: no such file or directory: --scopes=..., that is what happened. Paste the command as one long line, with no backslashes, no spaces in the middle.
Wall 3: GA4's Property Access Management refuses to add a service account email. The error says "This email does not match a Google Account." If you tried to set this up with a service account, that is why. Skip the service account path entirely and use Application Default Credentials (ADC) instead. That is what the setup above does. Your own Google login becomes the authentication, and you already have GA4 admin access on your own account.
Wall 4: Your organization blocks service account JSON key creation. If you work inside a Google Workspace organization with hardened security defaults, you may see a popup that says "Service account key creation is disabled." Again, ADC bypasses this entirely. You do not need a service account.
Wall 5: The token works but tools return 403 "insufficient scopes." The token is real but does not include the scope the tool needs. Check what is on your token:
If webmasters.readonly is missing, rerun the gcloud login command with that scope in the --scopes flag.
Wall 6: Empty results. If a tool returns no data, two things are usually true. Either your site is too new for Google to have aggregated enough searches yet (Search Console has a privacy threshold and hides queries with fewer than 10 monthly impressions), or your time window is too short. Try 90 days instead of 30. The data will be richer.
Five leaky-funnel patterns to find
A "leaky funnel" is anywhere visitors fall out of your sales process. Here is how to find each pattern using the MCPs.
Leak 1: High-traffic page with low engagement. Use the GA4 MCP to find pages that get lots of sessions but where most people leave fast. These are pages that promise something the visitor does not find. The fix: rewrite the headline, add a clear above-the-fold hook, or improve the page-to-source match.
Leak 2: High-intent page with weak conversion path. Look at pages people land on with clear buying intent (pricing page, comparison page, case studies). If they have low conversion rate, your CTA is invisible or weak. The fix: one clear CTA above the fold, repeated halfway down, repeated at the end.
Leak 3: Search clicks with no follow-through. Pull Search Console queries that drive clicks but where engagement on the landing page is under 30%. These are people who searched, clicked, and immediately left. Your page does not match what they expected. The fix: rewrite the first 50 words of the page to address the exact thing they searched for.
Leak 4: Sources that send traffic but never convert. Group GA4 data by source/medium and look at conversion rate per source. If Facebook is sending you 500 clicks a month and your conversion rate from Facebook is 0%, that is a leak. The fix: install link tagging so you can prove which campaigns convert, then kill the dead ones.
Leak 5: The "(not set)" rows. GA4 sometimes shows "(not set)" or "Unassigned" rows that bounce at 100%. That is almost never real human behavior. It is broken tracking, missing UTM tags, or a redirect stripping data. Find these by asking Claude to surface any traffic-source row with bounce rate above 95%. The fix: audit your tag setup, check your redirects, make sure every campaign link has proper UTM parameters.
Pro tip: set up conversion events first
This one is going to save you more than the setup itself. If you take only one action from this whole resource, take this one.
Most Google Analytics setups have zero conversion events configured. Zero. That means your dashboard says you got 4,500 visits last month but it has no idea how many of those visits turned into anything. Sign-ups, downloads, purchases, calls booked. All invisible.
Without conversion events, every page on your site shows a conversion rate of zero, every traffic source shows a conversion rate of zero, every campaign shows ROI of zero. You are flying blind.
To fix it, mark these as "key events" in GA4 (this is GA4's word for "conversion"):
Form submissions on your lead magnet
Clicks on your primary CTA button (book a call, buy now, sign up)
Pageviews of your thank-you page
Outbound clicks to your Skool, Discord, or Stripe checkout
Scroll depth past 90% on your sales pages
You can set these up manually in GA4 Admin then Events then Create event, then mark them as key events. Or, since you just gave Claude Code write access via the analytics.edit scope, just ask Claude:
Create a GA4 key event called book_call_clicked that fires when someone clicks an outbound link to charlieautomates.com/charlie-os-vs/ on any page of my site.
If you do not have conversion events configured, do this before you do anything else with the MCPs. Otherwise the "find my leaky funnels" prompts will tell you everything has a conversion rate of zero, which is technically accurate but useless.
FAQ
Do I have to be a developer to set this up? No. If you can paste a command into a terminal and click through a Google sign-in screen, you have what you need. The terminal commands are copy-paste. The Google steps are clicking enable buttons.
Does this cost anything? Both MCP servers are MIT licensed and free. Google's Analytics Data API and Search Console API are both free up to quotas that no solo business will ever hit. Total cost: zero dollars per month.
Will my data leave Google? No. The MCP servers run on your laptop. When Claude Code asks them a question, they call Google's APIs directly using your login, get the data, and pass it to the conversation in your terminal. Nothing gets uploaded anywhere.
Can I use this with Cursor or another AI tool, not Claude Code? Yes. Both MCP servers speak the standard Model Context Protocol, which Cursor, Continue, and other modern AI tools also support. The setup is identical.
I gave Claude write access. Can it break my Analytics setup? Technically yes. The analytics.edit scope can create or modify GA4 properties, events, audiences, and custom dimensions. In practice, Claude only does what you ask it to do, and you can always undo changes through the GA4 admin UI. If you want a safer setup, drop analytics.edit from your scope list and only use analytics.readonly.
If a command fails, do not fight it. Point Claude at the repo README and let it read the current steps. Install steps change faster than any blog post.
What each one does
Claudex Loop. Claude and Codex work as a team. Let Claude build and Codex check the work. Or flip it, if Codex is your main builder. It comes with starter commands, so you do not have to wire it up yourself.
anydoc. Word, PowerPoint, Excel and OpenDocument files become clean Markdown. Claude can read Markdown well. So every Office file you own turns into food your AI can use.
Archify. Claude explains your code in a wall of text. Archify turns that into a diagram you can click through instead.
Orca. This one is different. It is not a skill, it is an app that runs your models. Codex, Claude Code, OpenCode and Pi, all side by side. You get parallel worktrees, split terminals and a design mode. There is an Orca CLI too, if you would rather skip the app.
OmniRoute. This is close to unlimited Claude Code. It runs the Claude Code harness on any model. Stack it with OpenRouter and you reach a pile of free models you can point Claude Code at.
BASE. The best second brain for Claude Code. Other tools make you update them by hand, and that burns tokens. BASE updates itself when a chat ends. It gets smarter each session. It also remembers each project and the rules you set for it.
doop. A design canvas built for more than one person. You and your agents work on the same board at the same time. Think Claude Design, but you are not in there alone.
film-studio-skills. Seven skills that give Claude a real AI film pipeline. Hand it a script or even a rough idea, and you get prompts that are ready to generate. It checks for consistency at each step. Good match for the Higgsfield MCP or Remotion.
SimpleEnglish. Want your AI to stop sounding like a LinkedIn post and start sounding like a Boeing manual? This is the clean way to do it. It uses ASD-STE100, the plain English standard the aerospace industry writes with.
threejs-game-skills. The fun one. Works with Claude Code and Codex. It builds Three.js games you can play in a browser. Not much business use, but the results are wild.
Check the licence before client work
This is the part that catches people out. Free to install does not mean free to build a business on.
Safe for paid work. anydoc, Archify, Orca, OmniRoute, SimpleEnglish and threejs-game-skills are all MIT. Use them on client projects and do not think twice.
doop is AGPL-3.0. That licence has a catch. If you change it and host it as a service, you have to share your changes as open source. Fine for your own use. Read it first if you plan to sell it.
Claudex Loop and BASE have no standard licence. GitHub flags both as non standard. Read the terms yourself instead of guessing.
film-studio-skills has no licence at all. No licence means the author keeps every right by default. Fine to try on your own machine. Do not ship client work with it until they add one.
Two things to know before you start
Orca covers some of the same ground. It has its own design mode and its own usage tracking. Install Orca first and you may not need a separate canvas next to it. Pick the one that fits how you work.
Install one, not ten. Here is how these lists fail. You install all ten in an afternoon and use none of them by Friday. Find the line above that matches your problem this week. Install that one. Use it for a few days. Then come back for the next.
Want help setting this up
If one of these fights you, bring it to the community. This stack comes up on the weekly calls all the time.
Join CC Strategic AI on Skool and post what you tried and what broke. You will get a real answer, not a link to the docs you already read.
