Stop building AI agents: build one AI COO, a small team and a library of skills instead

I don't know about you, but I got totally caught up in the AI bug. I started building every little skill and employee I could, and I ended up with twenty-five skills, no idea how to use them, and nothing really working.

You didn't do anything wrong if that's you too. You went with what everyone was doing. You're told to build a separate agent for every single job and then wire them all together. That's exactly why your AI agents still feel stupid, still hallucinate and aren't really getting the work done, because the smartest people are building AI systems that do the complete opposite.

So I stopped building agents and changed how the whole thing works.

What I do instead

I mainly communicate and work with one employee, my COO. I allocate a lot to my COO, and it hands the work out to a small team of employees underneath it. Each of those employees pulls from one shared library of skills, rather than me running loads and loads of different agents.

A skill is a process or an SOP in my business, written down so Claude can follow it. An employee is a folder with an agent file that says who it is, what it connects to and the index of skills it uses.

This is where Anthropic's own guidance points. Their article When to use multi-agent systems (and when not to) says to start with one agent, and that multi-agent setups typically use three to ten times more tokens for the same task. It ends on this line: "Start with the simplest approach that works, and add complexity only when evidence supports it." Their engineers make the case for skills in the talk Don't Build Agents, Build Skills Instead.

The way I set it up is the Incuv8or AI Framework. It's six steps, in this order.

THE INCUV8OR AI FRAMEWORK

1. BUILD YOUR AI BRAIN
   Your business, written down where AI can read it.

2. MAP YOUR ORG
   The AI Org Mapper turns your brain dump into a COO, departments & the
   skills you need, & saves the chart in your brain.

3. BUILD YOUR COO
   The one employee you talk to. It knows your org chart.

4. ADD EMPLOYEES ONLY WHEN THEY PASS THE TEST
   Your COO drafts each one with you: a folder with an agent file &
   skills index. Anything that fails the four-part test is a skill in
   one shared library.

5. HAND OUT THE WORK FROM YOUR TASK MANAGER
   Same statuses on every list, a field for the employee, a dispatcher
   that picks it up.

6. ADD ROUTINES WHERE THEY'RE NEEDED
   Only for jobs that run on a schedule.

One rule runs across every step. Nothing goes live without your approval, and every correction becomes a rule the employee keeps.

The foundation is the most important part. Get the data and the structure right, and everything else becomes easy.

Why one COO and a small team works better than lots of agents

When every job has its own agent, you have to remember which one does what, open the right one and brief it every time. When one of them gets something wrong, you have to work out which one it was.

When most jobs are skills, the COO does that remembering for you. You give it the task, and it decides whether to do the work itself, split it across the employees underneath it, or hand it back to you if it would change something live. Each employee then looks at its own index and picks the skill that fits.

My Scriptwriter works that way. When it gets an Instagram task, it reads the post type and picks the scripting skill that fits, whether that's a talking head or a B-roll caption, then follows that skill. The formats are skills, not separate employees.

What you need before you start

• The Claude desktop app on a paid plan, with the Claude Code tab.
• Obsidian, or any folder on your computer, for your AI brain.
• ClickUp for Step 5. Any task manager with statuses, custom fields and comments will work, but the steps say ClickUp.

You don't need to know how to code.

Step 1 - Build your AI brain

Everything starts here. Your AI brain is your business, written down where AI can read it: who you are, what you sell and your clients. Your COO and every employee read from it before they do anything, so if it isn't there, they guess.

The full setup, with every prompt, is in How to set up your AI brain in Obsidian. When it's done, open the Claude desktop app, click the Claude Code tab and start a new chat. Click the + at the bottom of the chat, choose your AI brain folder and paste this.

Read my AI brain. Then tell me what I sell, who my client is & what's
going on with them right now. Tell me which files you read to answer.

What you'll see: the right answer, with the names of the files it read. If it's vague or wrong, fix the documents before you move on.

Step 2 - Map your org

The next thing you need to do is map your org. I mapped all of the different things I do in my business and how AI could help me do them, and that gave me a list of skills, which are really just the processes I already run.

You don't have to work that out on your own. I built a skill for it, the AI Org Mapper, and it interviews you about your business one question at a time.

Install it, then start a new chat in the Claude Code tab and attach your AI brain folder. Then type this.

Hey Claude, I want to use /org-mapper to map out my AI org chart.

It asks you what a typical week looks like, what you do on repeat, what you outsource, what you hate doing, and what you find easy and challenging. Voice notes are perfect for this.

What you'll see: your COO at the top, your departments underneath, the employees each department needs and the skills each employee uses, plus the top five to build first. It's saved in your AI brain as AI ORG CHART.md. Number one is always the COO.

The download, the install steps and the full session are in How to map your AI org chart before you build any AI employees.

Then turn the chart into folders. In the same chat, paste this.

Read my AI ORG CHART.md. Create a folder called ORG in my AI brain & move
the org chart into it. Inside ORG, create a COO folder, one folder for
each department on my chart & a folder called SKILL LIBRARY.

Don't create any employee folders or skills yet. Show me the folders
when you're done.

What you'll see: an ORG folder in your AI brain with your org chart, a COO folder, your departments and an empty skill library.

My business has eight departments under the COO: Business Operations, Clients, Marketing, Sales, Product, Education, Technology and Personal Life. Mine is bigger because I run multiple businesses. I am developing software, I have an educational and coaching platform, and I have my freelance business. If you're starting out with one stream of business, your departments will be fewer, so tailor them to your business.

Step 3 - Build your COO

Your COO is the one employee you talk to, and it knows your org chart. Start a new chat, attach your AI brain folder and paste this.

Draft ORG/COO/agent.md with me: who it is, its job, which model it runs
on, what it connects to, what it can touch, how it decides what to do
with a task & its skills index. It reads my AI ORG CHART.md so it knows
my departments. It never sends, pays for, deletes or publishes anything,
& it never touches a password. Ask me one question at a time & show me
before you save.

What you'll see: a COO agent file you've read before it was saved. If Claude asks which employees the COO has and the answer is "none yet", that's fine. The COO does the work itself until you approve one.

I've turned my COO into a template with none of my business in it. It's in The First AI Employee I'd Build as a Non-Techie Founder: An AI COO.

Step 4 - Add employees only when they pass the test

This is where most people go wrong, and where I went wrong the first time. Don't go and build twenty-five skills or twenty agents. Your COO drafts each employee and its skills with you, one at a time, starting with your top five, and you test and approve each one.

Before it drafts anything, it runs the test. Anthropic gives three reasons to add another agent: protecting context, running work in parallel and specialisation. Their Claude Code docs on subagents add limiting which tools an agent can use and sending work to a cheaper model. The test I use is built from both. A job only gets its own employee if it needs its own context, different permissions or tools, a different model, or to run at the same time as the others. Otherwise it's a skill in your shared library, held by an employee you already have.

You're my COO. Read ORG/COO/agent.md & ORG/AI ORG CHART.md.

Take number 2 on my top five. First check it really needs to be its own
employee: its own context, different permissions or tools, a different
model, or running at the same time as the others. If it doesn't, tell me
which employee should hold it as a skill instead.

If it does, create its folder in its department & draft its agent.md &
its skills index. Then interview me about the first skill, one question
at a time, like you're writing an SOP for a new hire, & save it in SKILL
LIBRARY. Show me everything before you save.

Test it with a real job in a chat. When it does the job the way you would, approve it and move on to number 3. Most people need a COO and three to five employees to start.

What goes inside each employee's folder is in How to build a toolkit for your AI employees, and what goes in a skill file is in How to Build a Claude Skill.

Step 5 - Hand out the work from your task manager

The next piece is the workflow. I use ClickUp so I can manage when and what the COO is doing.

Every list has the same statuses, and every task has a field called AI EMPLOYEE. I write what I want in the task, set AI EMPLOYEE to the COO and set the status to "w/ agent to do". A small script called the dispatcher checks ClickUp every 10 minutes, starts the right employee on the task, and when it's done the task comes back to me at "w/ samantha for review".

The prompts to set up the statuses, the field and the dispatcher are in steps 4 and 5 of The First AI Employee I'd Build as a Non-Techie Founder: An AI COO.

What you'll see: a task you gave the COO coming back to your inbox with the work on it, done by the right employee with the skill that fits.

Step 6 - Add routines where they're needed

Routines come last. A routine is a schedule that wakes an employee up to do its job, and it's only for jobs that run on a schedule, like a morning brief or sorting your inbox. Everything else goes to the COO as a task.

Add routines once your employees are doing the job well when you hand it to them. If you're not sure whether a job needs one, I wrote up the questions I use: Claude routine or AI employee?

The catch, and the fix

The catch is that your team is only as good as its skills. If a skill is vague, the employee will pick it and still get the job wrong, and it's tempting to fix that by explaining it again in the chat.

The fix is to fix the skill, not the chat. Every correction you give goes into the skill or the employee's learned rules, so it doesn't make the same mistake twice. The COO template already has a Learned rules section for this.

The second catch is that anything that changes something live, like a website or a routine, comes back to you as a session at your computer. Nothing live changes without you.

What this looks like as a business

Every business owner you support has seen the same videos and either has a pile of half-built agents or is too overwhelmed to start. Setting up their AI brain, mapping their org, building one COO with a small team and writing their processes into skills is a systems job, and it's work they can't do for themselves. You're paid for knowing what to build and in what order.

Start with one

Don't build another agent this week. Check your AI brain with the Step 1 prompt, run the AI Org Mapper, read your top five, and then build your COO.

If you want the whole framework in one place, it's in You don't need an AI operating system to have an AI team.

If you haven't got your AI brain set up, that's the foundation for all of this: start here.