How I Built an AI Team in Claude (and Why I Cut It Down)

I used to think AI agents were something other people built. People who could code. I'm a strategist, not a developer, and "build an autonomous agent" sounded like something I'd pay someone for.

Then I built one. I've been building AI employees all year, since I started working with OpenClaw in January, and for a while I built a lot of them. That's where it went wrong.

This post is the whole picture: how my team started, why I cut it down, what it looks like now and how work moves through it. The prompts are written up separately and I'll link them.

How it started

When I first mapped my AI team, I did what you'd do for a real company. I sat down, listed every job my businesses need done, and put them in departments with someone in charge of each.

I settled on ten directors. Under them, managers who each owned one thing. Under the managers, employees who each owned one job. Every job got its own employee, with its own skill, its own routine and its own place in ClickUp.

It looked great on paper. In practice I had more employees than I could remember, half of them I didn't use, and every new idea became another employee to build and look after.

What it looks like now

Now I basically only ever talk to one AI employee, my COO. I brain dump to it and it hands the work out. Underneath it, my business is split into departments, and each department has the fewest employees it needs.

Me
└── COO (Opus) - the one I talk to; hands out the work
    │
    ├── BUSINESS OPERATIONS
    │   ├── Ops Manager (Sonnet) - the gritty admin
    │   ├── Inbox Assistant - files my email & drafts replies
    │   └── Meeting Notes - files call notes & pulls out actions
    │
    ├── CLIENTS
    ├── MARKETING
    │   ├── Blog Manager - writes my blog posts from my briefs
    │   ├── Scriptwriter - helps me write my Instagram scripts
    │   └── ManyChat Manager - builds the keyword flows
    │
    ├── SALES
    ├── PRODUCT
    ├── EDUCATION
    ├── TECHNOLOGY
    │
    └── PERSONAL LIFE
        └── Personal Finance Manager

Skills library (shared) - every process written down once,
used by whichever employees need it

A department is a grouping, not an employee. I have eight because I run more than one business: my freelance clients, Tech Queen Academy, and the products I'm building. Some of mine don't have an employee yet, and that's fine. They get one when there's work that needs it.

If you're a freelancer starting from scratch, you'll probably need fewer. Your departments should fit your business, not mine.

If you haven't mapped yours yet, that's the first step: Before You Build Any AI Employee, Map Your AI Org Chart.

What an AI employee actually is

I used to say an AI employee is three things: a skill, a routine and a workflow. Those three pieces are all still in my setup, but they're not the employee.

An employee is a folder. Inside it is its agent file: who it is, what its job is, which model it runs on, what it's allowed to touch, and the index of the skills it uses.

A skill is one process, written down. How I write a blog post. How I write a teaching reel script. How I check a website's legals. Skills live in one shared library, and more than one employee can use the same one. My writing skills, like my voice, are used by the Blog Manager and the Scriptwriter.

A routine is a schedule that wakes an employee up. My Inbox Assistant runs every hour. Routines come last, once the employee does the job well when you hand it to them.

The workflow is where the work arrives. A field in ClickUp that says which employee owns a task, and a status that says whose turn it is.

I wrote each of those pieces up on its own:

• Skills - the three types, and the template
• Routines - the three types, and the templates
• The Workflow - the ClickUp setup, and how a task moves

All of it sits on your AI brain.

Every employee reads your business from your files, every run, because they start with no memory. If your business isn't written down somewhere Claude can read it, none of this works. Not set up yet: My AI brain in Obsidian.

When a job gets its own employee

This is the bit that cut my team down. A job only gets its own employee if it needs at least one of these four things. Otherwise it's a skill, held by an employee that already exists.

1. Its own context. A big, focused area of knowledge that would crowd everything else out. My Blog Manager only thinks about my blog.
2. Different permissions. It's allowed to touch things the others shouldn't. My ManyChat Manager works in my logged-in ManyChat account. My Inbox Assistant works in my mailboxes.
3. A different model. My Ops Manager runs on Sonnet, which is faster and lighter, so routine admin doesn't use up Opus.
4. To run in parallel. It needs to work at the same time as the others, like my Meeting Notes filing a call while the Blog Manager writes.

My website legals check, my compliance check and my security audit used to be three employees. None of them passes the test on its own, so now they're skills.

Most people need a COO and three to five employees to start.

How a task moves

Every task in my ClickUp has a field called AI EMPLOYEE and a status. When I want something done, I write what I want on the task, set AI EMPLOYEE, and set the status to w/ agent to do. A small script checks ClickUp every 10 minutes and starts Claude on that task with that employee's agent file.

When it's done, the work is at the top of the task, my brief is kept underneath, and the status is w/ samantha for review. If I want changes, I comment and send it back. If I ask a question, it answers me in the comments instead of redoing the work.

If I give a task to the COO, it decides: do it itself, split it across the employees it's allowed to use, or hand it back to me if it would change something live.

Every change gets a comment on the task and a line in my log. That's how nothing gets done twice, nothing gets lost, and I can always see where something is without asking.

And one loop runs without me touching it. Every night the Knowledge Base Updater reads the day and writes it into each client's log and status page. Every morning the Morning Brief reads what it wrote and tells me what today is.

The rules

1. One COO at the top. I talk to the COO. The COO hands out the work.
2. Employees are earned. A new employee has to pass the four-part test. Everything else is a skill.
3. Skills are written once. One shared library, and each employee has an index of the skills it uses.
4. Nothing public without me. No employee posts, sends, publishes or deploys on its own. Everything stops at w/ samantha for review.
5. The COO drafts, I approve. New employees and skills are drafted by the COO, one at a time. I test each one before it takes any work.
6. Everything is written to the vault or the task, never only to a chat. A chat is gone by morning.
7. Passwords never go in a file or a chat. The Mac's Keychain holds them.

What it actually feels like

Not like having a team of robots. More like having a very good operations person who never sleeps and never has an opinion about how I run things.

I open the laptop. There's one page telling me what today is. There's one view with the things that need me. My inbox has only people in it, with the replies already drafted. The client status pages are true as of last night, and I didn't write them.

And when I have a new idea for an employee, I don't build it. I ask my COO whether it passes the test.

Start with one

Don't start the way I did, with ten directors. Follow The Incuv8or AI Framework: build your AI brain, map your org, then build your COO and give it that chart. Your COO drafts your first employee with you, you test it, and you add the next one when that one works.

Map your AI org chart · Build your AI COO

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

And if what you want is the whole system - the pricing, the positioning, and the tech and AI that lets you run a business on three days a week - that's what I teach.

Watch the free training →