Find the one bottleneck killing your revenue.
30-minute free strategy call with Charles. We diagnose your highest-leverage AI bottleneck, install Charlie OS live on your Mac or Windows in one hour, and map your exact next 30 days, all on the same call.
Can you build an entire team of AI executive officers that runs your business while you sleep?
My honest answer is about 60% yes, with real caveats. I’m Charles J Dove, and I run Charlie Automates. In this post I build a working AI CMO from scratch in Claude Code, and I tell you exactly which 40% does not work yet.
What is an AI CMO, actually?
An AI CMO is a markdown file. That is not a joke and it is not a simplification.
An AI executive officer is a text file containing instructions and frameworks that describe how a role operates, plus wired connections to the tools that role needs. Claude reads the file, adopts the role, and acts through the connected tools. Agents, skills, and slash commands are the same object described three ways. Once you internalize that, the mystique drops away and this becomes an engineering problem with a known shape.
The hard part was never the AI. The hard part is deciding which data to feed it.
The 60/40 split: augmentation versus automation
This build focuses on AI augmentation, meaning AI that helps you make better decisions, rather than AI automation, meaning AI that takes actions unattended. That distinction is why my answer is 60% and not 100%.
Here is the rule I would hand anyone before they write a line of this:
If you cannot right now list the five highest-ROI processes in your business, do not automate anything. Most people try to automate a process they have never actually defined. You end up with an agent faithfully executing a workflow that was broken before AI touched it, only now it runs at machine speed.
Define the process first. Augment second. Automate last, and only where the process is proven.
What every AI officer requires
An AI executive officer requires exactly three components, and the third is the one people skip. This build uses Claude Code as the runtime, a written spec for the deliverable, and two frameworks running in sequence: SEED and SkillSmith.
| Component | What it is | Why it matters |
|---|---|---|
| An agentic coding tool | Claude Code, running in VS Code | The runtime. Nothing else in this class is close right now |
| Clarity on the deliverable | Knowing what you want and what data to give it | You cannot direct an agent toward an outcome you have not named |
| The right frameworks | SEED and SkillSmith | Without a proven skill structure, output quality is random |
I ran the ideation on Opus 5 and switched to Fable 5 for the actual build. Ideation rewards depth of reasoning. Scaffolding rewards speed.
If Claude Code itself is new to you, start with the Claude Code crash course and come back. The rest of this assumes a working setup.
Why frameworks decide the outcome
SEED and SkillSmith do two different jobs, and running them in sequence is the entire method.
SEED handles ideation. It runs a guided interview that turns a vague idea into a structured spec Claude Code can act on cleanly. You are not writing a prompt. You are answering questions until the shape of the thing is unambiguous.
SkillSmith handles construction. It scaffolds the skill file with a consistent internal structure: frameworks, sources, tasks, checklists, templates, and the top-level skill definition that governs how the whole thing operates.
Without the right framework underneath a skill, it does not matter how good your idea is. The output will not be what you wanted. Both are free in the Founder’s Toolkit.
Step 1: Ideate the CMO with SEED
I gave SEED a short prompt describing a CMO officer and, critically, what it should be allowed to touch:
- Apify to scrape my Instagram and YouTube data
- The
/viral:discovercommand to audit competitors and my own content - GoHighLevel to read CRM analytics, prospects, and deals
- GA4 to read website analytics
That list is the actual job. A CMO who cannot see content performance, pipeline, and site analytics is not a CMO, it is a copywriter.
SEED came back with a problem statement that named the real issue: every workflow started with a separate command, and all the connective reasoning lived in my head. The solution it proposed was one office to talk to, with a persistent brain underneath.
Then it asked the question that shaped everything after it: is the CMO a strategist I consult, or an operator that acts?
I chose consultant. That single answer cut the scope in half and removed most of the risk.
Step 2: Set approval gates before you build anything
Approval gates define what an agent may do without asking you, on a four-level scale: consult, recommend, draft, and send. My CMO is capped at draft. It never sends.
This is the security layer, and it belongs in the skill file itself rather than in your memory of how you intended to use it. An agent with CRM access and send permission is one bad inference away from emailing your entire pipeline. An agent capped at draft produces the same output and leaves the irreversible action with you.
Build the gate before the capability. It is far harder to retrofit restraint onto an agent that already has permissions.
Step 3: Feed it your existing data
The biggest quality jump does not come from a better prompt. It comes from the agent knowing your business.
If your context currently lives in ChatGPT or the Claude desktop app, you can move it. Open Settings and pull your saved memories directly, or open a fresh session and ask: “What are the core things you know about me? I am building a CMO officer. Put everything together in one document I can hand to another setup.”
Then read what comes back and decide what is worth migrating. Most of it is not.
On my machine this layer is already handled. I run a persistent operator profile at the top of my Claude folder holding my projects, my goals, and my working context, so any agent I spin up already knows who it works for. That is the difference between an agent that gives you generic marketing advice and one that gives you advice about your business. If you want the structural version of this, I wrote up the 3-step system to build an agentic OS, which is the foundation everything here sits on.
Step 4: Scaffold with SkillSmith
SkillSmith asked how the CMO should be invoked. Single front door, /cmo, which is the right answer nearly every time. Many entry points means you have to remember which is which, and you will not.
It then scaffolded the full structure:
- Data sources. Which MCPs it uses, when to use each, and why not the others.
- Frameworks. Email architecture, funnel math, hook scoring, list growth.
- Checklists. Approval gates, data integrity, draft quality.
- Tasks. The work it is responsible for, in files it updates itself as things change.
- Templates. Six distinct output formats so results are consistent every run.
- The master skill file. The top-level governance. You never edit this by hand. Run
/skillsmith tasks auditand let Claude repair its own skill.
That last point is worth sitting with. The agent maintains its own definition. You audit, it fixes.
The process is slow and boring to watch. That is usually the signal it is working. The boring things get the results.
What the CMO actually produced
I ran /cmo and asked for a social audit plus an HTML report. It returned three deliverables built from three months of YouTube and Instagram data plus my live GoHighLevel pipelines: a content audit, a CRM audit, and a write-back to its own brain files. Two of the findings were things I did not already know.
Content audit. A full three-month rundown of YouTube and Instagram with linked videos, so I could go back and rewatch what performed. It flagged one Instagram carousel as off-ICP, meaning it drew engagement from people who will never buy. That is a real insight; raw engagement numbers would have called that post a winner.
CRM audit. It read my GoHighLevel pipelines and told me which deals closed and which stages had gone stale. Two sections of my pipeline had contacts sitting untouched. I had not looked at that pipeline in weeks.
Brain update. After delivering, it wrote the findings back into its own brain files. The next run starts from what it already learned rather than from zero.
From there the loop is obvious: “Draft a campaign for the stale section of the pipeline.” It has the contact data, it has the funnel math framework, and it is capped at draft, so nothing goes out without me.
How do you run an AI CMO 24/7?
You have three options, and only one of them genuinely runs unattended.
| Approach | How it works | The catch |
|---|---|---|
| Manual | You type /cmo when you want it | Zero infrastructure, zero autonomy |
| Cron routines | Scheduled local triggers, for example every Monday 9am | Your machine has to stay on 24/7 |
| Hermes or OpenClaw | Skills pushed to a hosted runtime via open APIs | Real 24/7, requires setup and hosting |
Cron is the honest middle option and its limitation is physical, not technical. A scheduled job on your laptop does not fire when the laptop is closed. If you need true always-on, the skills have to live somewhere that is always on.
The 40% that does not work yet
I said 60%, so here is the rest of it plainly.
- It has no judgment about your market that you did not give it. It analyzes what you connect. It does not know what you have not told it.
- Onboarding is not optional if you want good output. I skipped it in this build for demonstration purposes and the first run showed the gaps. Spend the time filling in the brain.
- Stale data produces confident wrong answers. My CRM analysis was thorough and partly about a pipeline I no longer use. The agent could not know that.
- Draft-only is a real ceiling. It is the correct ceiling for anything touching customers, but it means a human is still in every outbound loop.
An AI executive officer that augments your decisions is available today and genuinely valuable. An AI executive officer that replaces the role outright is not, and anyone selling you that is selling you something.
Key takeaways
- An AI officer is a markdown file plus tool access. Agents, skills, and commands are the same object.
- Run SEED before SkillSmith. Ideate the spec, then scaffold the skill. Skipping the first step produces a well-built agent that does the wrong job.
- Choose consultant or operator up front. That one decision sets scope and risk for everything downstream.
- Set approval gates before capabilities. Draft-only is the right default for anything touching customers.
- If you cannot name your five highest-ROI processes, do not automate. Augment first.
- Cron needs your machine on. Hermes or OpenClaw for genuine 24/7 operation.
- The agent maintains its own skill file. Run
/skillsmith tasks auditinstead of hand-editing.
FAQ
Can an AI agent really replace a CMO?
No, and about 60% is the honest number. An AI CMO built this way analyzes content performance, reads your CRM, connects site analytics, and drafts campaigns from real data. It does not carry market judgment you have not given it, and capping it at draft-only means a human stays in every outbound decision. It replaces the analysis and drafting work, not the accountability.
What is the difference between AI augmentation and AI automation?
AI augmentation means AI helps you make better decisions while you retain the action. AI automation means AI takes the action unattended. Augmentation works reliably today. Automation works only where you have an already-defined, already-proven process. Most failed AI projects are automation attempts on processes that were never documented.
Do I need to code to build an AI CMO?
No. The skill file is markdown, which is plain text with light formatting. SEED runs a guided interview to define what you want, and SkillSmith writes the file structure for you. You answer questions and review output. The only technical step is connecting MCP servers, which is a paste-a-token operation.
What are approval gates and why do they matter?
Approval gates define the four levels of autonomy an agent can operate at: consult, recommend, draft, and send. They sit inside the skill file so the limit is enforced in the agent’s own definition rather than depending on your memory. An agent with CRM access and send permission can email your entire pipeline off one bad inference. Capping it at draft gives you the same output with the irreversible step still yours.
How do I get my ChatGPT memory into Claude Code?
Open ChatGPT Settings and pull your saved memories, or start a fresh session and ask it to compile everything it knows about you into one document for migration. Then review it and keep only what is relevant to the role you are building. Most saved memory is noise for a specific executive agent.
Can I build other executives the same way?
Yes, and the process does not change. CFO, COO, head of sales, head of ops. The only two variables are what tools it can access and what data you train it on. Ideate with SEED, scaffold with SkillSmith, set approval gates, connect the MCPs that role needs.
What MCPs does an AI CMO need?
At minimum: a CRM connection like GoHighLevel for pipeline data, GA4 for website analytics, and a scraper like Apify for social performance. I also wire in vidIQ for YouTube specifics. The rule is simple: connect what the human version of that role would need access to, and nothing more.
Why use two frameworks instead of just prompting Claude directly?
Because output quality tracks structure, not prompt cleverness. SEED forces the spec to be unambiguous before any code is written. SkillSmith enforces a consistent internal skill structure so results are repeatable across runs. Direct prompting produces something that works once and cannot be maintained.
Where to go from here
SEED and SkillSmith are both free in the Charlie Automates Founder’s Toolkit.
If you want to learn to build these yourself, at your own pace, with weekly calls and the full zero-to-dangerous Claude Code course, join CC Strategic AI on Skool.
If you would rather have the whole executive suite built for you, that is what I do at my agency CC Strategic, and Charlie OS is the operating system it runs on. More at charlieautomates.com, and I post every build on my YouTube channel @charlieautomates.
Ready to install Charlie OS?
Stop reading. Start shipping.
The AI Bottleneck Protocol is a 30-minute free call. We pinpoint the one task eating the most hours and the most money in your business right now. We install Charlie OS live on your machine in under an hour. We map your exact path to fix the bottleneck on the same call. If we are not a fit, you walk away with a clear diagnosis of where your business is leaking time and money. Either way, you win.