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 AI Bottleneck Finder asks you a handful of questions, then tells you the one constraint capping your revenue right now plus two builds you could ship this week. Five minutes, free, no call required.
Take the free AI Bottleneck Finder →
The best Claude Code skills are not the ones with the most GitHub stars. They are the ones that own a role nothing else in your setup owns.
If you are handing Claude Code a pile of random skills and expecting sharper output, you will keep getting the same mediocre results no matter how many you add. Ten skills that all roughly do the same thing is like hiring ten support reps and never telling any of them who handles what. They all answer the same ticket, and the answer is different every time.
I’m Charles J Dove, and I run Charlie Automates. I have been using Claude Code for over a year. In the last 4 months alone I installed 73 skills across 316 sessions. These 5 are what is left, ranked by actual usage, and the order matters more than the list.
Why Does Install Order Matter More Than the Skill List?
Install order matters because each layer inherits what the layer above it already knows. A skill installed underneath a context layer gets your rules, your projects and your voice for free. The same skill installed on a bare setup has to be told all of that every time you use it.
That is why BASE goes first and the design skills go fourth. It is not a preference. Everything below the context layer inherits from it, so putting the context layer in last means every skill you installed before it was built against a Claude that did not know who you were.
The 5 roles, in order: context, skill creation, execution, design, and voice. One skill per role. If two skills claim the same role, you have a routing problem, not a capability problem.
Skill 1: BASE, the Context Layer
BASE is a knowledge graph for your rules, and it belongs at the top because it determines the result of everything under it.
It registers Claude Code hook events. Every time a keyword shows up in your conversation, the matching context gets injected, every single time, without a tool call. That happens through domain rules. Right now, writing this, my GLOBAL domain is active with 14 rules because it is always on, and my CONTENT domain fired because I used content words. Nothing else loaded.
Compare that to a CLAUDE.md file. Everyone knows what that is by now: a static document that dumps its entire contents into every session with no awareness of what actually applies. BASE matches on keywords and file paths instead, then suppresses what is not relevant to what you are discussing.
The usage number tells the story better than I can. 5,922 CLI calls in 4 months. It is not a skill I remember to use, it is the layer everything else runs on top of.
Just like your .claude folder, BASE creates a .base folder: one surface for your domains and rules, another tracking your projects and sessions. Get it, plus the starter commands the creator gave me, on the BASE resource page.
Skill 2: SkillSmith, the Skill That Builds Skills
SkillSmith is the skill you use to create every other skill, and it ships with 4 commands: discover, scaffold, distill and audit.
The problem it solves is structural drift. Build a skill with vanilla Claude and you get a single SKILL.md file. No checkpoints, no separation between routing logic and business logic, and things get messy the third time you extend it.
Here is what SkillSmith produced for Listing Cinema, a real estate skill of mine that turns photos into videos: a checklists folder for quality-checking output, a frameworks folder for the workflows inside the skill, a tasks folder with one task per video type, a templates folder showing what every output should look like, and the SKILL.md on top. Same structure every time, which is the entire point.
The second use is more valuable than the first. A skill you download from the internet was built for the person who built it, not for you. Their routing logic assumes their file paths, their business assumes their offer. Run audit on a downloaded skill and SkillSmith reverse-engineers it into something built around your setup. Run distill on a book, a course or a transcript and it turns that source into a working skill.
With BASE above it, SkillSmith also knows your business and who you are while it builds. That is the compounding I mean by order. Full walkthrough on the SkillSmith resource page.
Skill 3: PAUL, the Execution Layer
PAUL is an orchestration layer for building things, and it stands for Plan, Apply, Unify, Loop. If you know GSD, it lives in the same category.
The problem it solves is context rot. Long sessions degrade. Subagents hand back work that costs more to clean up than to redo. Plans get abandoned without closure, and the real state of the project drifts quietly away from what you think it is.
PAUL’s fix is to write the given, when, then acceptance criteria before a single task exists. Then it applies the plan, verifies against its own task list, and unifies: reconciling planned against actual and closing the loop.
Here is the file structure from a dashboard app I built with it. A PROJECT.md with the full scope and the reason for building it. A ROADMAP.md with every phase. A flows file for anything special the app needed. And a STATE.md, which is the one that earns its keep. It tracks completion per phase, the loop position, the attached GitHub repo, and any deferred issues. Mine read 95% overall and 90% through phase 5, with the last part not yet unified. That is the number I would otherwise be keeping in an Apple note.
The app took 5 phases. Inside each phase: plan, apply, unify, then plan the next part. Rinse and repeat. And when you step away mid-build, paul pause writes a handoff, then paul resume archives it, reads the state and knows exactly what comes next. Get it on the PAUL resource page.
Skill 4: Front-End Design Plus UI UX Pro Max
Slot 4 is 2 skills, because I never run either one alone. Front-End Design ships with Claude Code from Anthropic. UI UX Pro Max you download yourself. I have taught my setup to fire both in tandem on any design work.
That pairing is a standing rule in my BASE domains, which means it is never optional. I do not remember to invoke it; the keyword fires and both load. That is the payoff of having installed the context layer first.
One honest note on design, since this is the section where people expect me to sell animation. The landing pages that convert best are not the ones with the coolest 3D scroll effects. charlieautomates.com has case studies, testimonials, an about section, and one main CTA to take the bottleneck quiz. It gets to the point fast. These 2 skills are how I built it.
You could use SkillSmith to distill both into a single design skill. That is a good weekend project and I have not done it yet. Both are on the UI UX Pro Max resource page, with more on the anti-slop design stack.
Skill 5: Humanizer, the Most Underrated of the Five
Humanizer makes Claude’s output sound like you, and it is the skill people skip until the first time they send AI-flavored copy to a real prospect.
I built mine by giving Claude 36 of my own YouTube transcripts, so it learned how I actually speak. On top of that it runs 35 proven patterns from Wikipedia’s WikiProject AI Cleanup, which is a catalogue of the tells that make text read as machine-written. No em dashes. None of the hedging vocabulary. None of the three-item rhythm every model falls into.
Where it pays off is inside other workflows, not on its own. My bottleneck report workflow pulls prospects from the pipeline, reads what they submitted, and writes a custom email that activates three skills at once: the email skill for structure, the subject line skill for the open, and Humanizer for the voice. The output looks the same every time and does not read like AI.
Get it on the Humanizer resource page.
What Does the Full Stack Look Like?
The full stack is 5 roles filled by 5 skills, installed top down so each layer inherits the one above it.
| Order | Skill | Role it owns |
|---|---|---|
| 1 | BASE | Context layer: rules, projects, sessions |
| 2 | SkillSmith | Skill creation with consistent structure |
| 3 | PAUL | Execution and orchestration |
| 4 | Front-End Design + UI UX Pro Max | Design that does not look generated |
| 5 | Humanizer | Output that sounds like you |
Install all 5, then build the sixth with SkillSmith. That sixth one is yours: the skill for the thing only your business does. Nobody on the internet is going to publish it, because it only makes sense inside your operation.
Key Takeaways
- One skill per role. 10 skills that overlap produce different output every run. 5 that own distinct roles produce the same output every run.
- Order beats selection. Everything under BASE inherits what BASE knows, so a context layer installed last wastes every install before it.
- 73 installed, 5 kept across 316 sessions in 4 months. This list is a survivor list, not a roundup.
- Downloaded skills need reverse-engineering. They were built for their author’s setup. Run SkillSmith’s audit before you trust one.
- PAUL’s
STATE.mdis the artifact that matters. Completion percentage, loop position, repo, deferred issues, all in one file Claude reads before it touches the project. - Humanizer earns its slot inside other workflows, chained behind an email skill and a subject line skill, not run on its own.
- Build the sixth skill yourself with SkillSmith, for the process only your business runs.
This stack is one layer of a bigger idea. If you want the architecture the whole thing sits inside, read the 3-step system to build an agentic OS.
Where to Go From Here
Install Charlie OS in one hour.
The AI Bottleneck Protocol, live on your Mac or Windows in 60 minutes, with this whole stack already wired. Then we map your exact next 30 days on the same call. Free first, fit-checked.
All 5 skills are free on charlieautomates.com:
- BASE, the context layer
- SkillSmith, the skill that builds skills
- PAUL, the execution layer
- UI UX Pro Max plus Front-End Design
- Humanizer
If you want to go deeper on any one of them, join CC Strategic AI on Skool. Entry is free, there is a Zero to Dangerous Claude course inside that walks the PAUL workflow specifically, and Premium members get weekly calls where we work through real setups.
At my agency CC Strategic we install stacks like this for business owners, and new builds drop every week on my YouTube channel @charlieautomates.
FAQ
What are Claude Code skills?
Claude Code skills are packaged instruction sets that load when their trigger conditions match, giving Claude a defined workflow for a specific kind of task instead of improvising one. Each skill lives in your .claude/skills folder with a SKILL.md entry file, and can carry supporting tasks, templates, checklists and frameworks alongside it.
How many Claude Code skills should I install? Install one per role, not one per idea. I installed 73 in 4 months and kept 5, because overlapping skills produce inconsistent output. The 5 roles worth filling are context, skill creation, execution, design and voice.
What is the best Claude Code skill to install first? BASE, the context layer. Every skill you install afterward inherits what BASE already knows about your rules, your projects and your business, so installing it first compounds across the whole stack. Installing it last means everything before it was configured against a Claude with no memory.
Do I need SkillSmith if I can just write a SKILL.md? You can write one file by hand. SkillSmith gives you the same structure every time: checklists, frameworks, tasks and templates alongside the entry file. That structure is what stops routing logic and business logic from tangling as the skill grows, and it is what makes a skill auditable later.
Can I use a Claude Code skill someone else published? Yes, but audit it first. A shared skill encodes the author’s file paths, offer and preferences. Run SkillSmith’s audit command on it and rebuild it around your setup before you rely on it.
What does PAUL do that a good prompt does not? A prompt gets something built. PAUL adds the shape around it: acceptance criteria written before any task exists, a roadmap of phases, a state file tracking completion and loop position, and a unify step that reconciles what was planned against what was actually built. It also survives you stepping away, through pause and resume.
Is Humanizer just a “remove em dashes” prompt? No. It runs 35 documented AI-writing patterns from Wikipedia’s WikiProject AI Cleanup against your draft, and it writes from a voice fingerprint built out of your own material. Mine was trained on 36 of my YouTube transcripts. Removing em dashes is the smallest thing it does.
Are these Claude Code skills free? All 5 are free and linked on the Charlie Automates Founder’s Toolkit. Front-End Design ships with Claude Code from Anthropic; the other four are downloads.