5 skills I taught one AI employee, and the prompts to steal them

The Incuv8or AI Framework - step 4 of 6: Add employees only when they pass the test. Before this: build your COO. Next: hand out the work from your task manager.

When I first mapped out my AI team, my content had three writers on it. There was a Scriptwriter for my talking heads, a B-roll Caption Writer for my B-rolls and a Carousel Writer for my carousels. Each one would have had its own agent file, its own instructions and its own place on the org chart.

Then I looked at what they actually did. All three read my brief, read my Brand Bible, read the blog post the content points to, and write something in my voice. The only thing that changed was the format.

So I merged them into one employee, the Scriptwriter, and taught it every format as a skill. It now has five skills, and when LinkedIn comes next, it will learn those too. I don't need a LinkedIn writer. I need my Scriptwriter to learn a new skill.

What you're building

You're building one employee with a handful of skills, and an index the employee reads first so it knows which skill to use. This is my Scriptwriter's index, simplified.

SCRIPTWRITER - WHICH SKILL TO USE

Read the task first: the Post Type, the pillar, the keyword & the
blog post, & anything I've written in the brief.

POST TYPE                 SKILL
Talking head (my idea)    talking-head-script → my own idea workflow
Repurpose talking head    talking-head-script → proven Reel workflow
B-roll (teaching)         broll-teaching-caption
Any other short script    psych-script
LinkedIn post             (next skill to teach)

No skill for it yet? Don't make one up. Tell me & hand it back.

When a task lands with the Scriptwriter, it reads the task, finds the row and follows that skill. I don't choose the skill and I don't brief it on the format. The index does that.

What Anthropic says about building this way

You'll see a lot of people building a separate agent for every job. Anthropic's own guidance goes the other way.

In When to use multi-agent systems (and when not to), published in January 2026, Anthropic says to start with one agent. They say multi-agent setups typically use three to ten times more tokens for the same task, and they end on this line: "Start with the simplest approach that works, and add complexity only when evidence supports it."

In Equipping agents for the real world with Agent Skills, they describe a skill as a folder with a SKILL.md file in it, plus any instructions and resources it needs. Claude only keeps each skill's name and description in mind until the job needs it, and then it reads the whole thing. That's why one employee can hold a lot of skills without getting worse at any of them.

Their engineers made the same case in a talk called Don't Build Agents, Build Skills Instead.

So the direction is one employee with a library of skills, and more employees only when you have a reason.

How I decide: the Incuv8or AI Framework and the four-part test

I set my AI team up with the Incuv8or AI Framework. It's six steps, and it's repeatable, so you can run the same steps on your own business.

THE INCUV8OR AI FRAMEWORK

1. BUILD YOUR AI BRAIN
2. MAP YOUR ORG
3. BUILD YOUR COO
4. ADD EMPLOYEES ONLY WHEN THEY PASS THE TEST
5. HAND OUT THE WORK FROM YOUR TASK MANAGER
6. ADD ROUTINES WHERE THEY'RE NEEDED

One rule runs across every step. Nothing goes live without your approval, and every correction becomes a rule the employee keeps. The full framework, with a prompt for every step, is in Stop building AI agents.

This post sits at step 4. Anthropic gives three reasons to add another agent: protecting context, running work in parallel and specialisation. The test I use is built on those three and on their Claude Code docs. A job only gets its own employee if it needs one of these four things.

DO I NEED ANOTHER AI EMPLOYEE?

1. OWN CONTEXT         A big, focused area of knowledge
2. DIFFERENT TOOLS     Access the others shouldn't have
3. A DIFFERENT MODEL   Cheaper for simple work, stronger for hard work
4. RUNS IN PARALLEL    It has to work at the same time as the others

Yes to any one → its own employee.
No to all four → a skill, held by an employee you already have.

My three writers failed all four. They read the same things, touched the same things, ran on the same model and never needed to write at the same time. A different format is a different process, and a different process is a skill. The full test, with the reasons behind each part, is in Do you need another AI employee?

The five skills I taught my Scriptwriter

1. Knowing which skill to use

The first skill is the index itself. It's a skill file like the others, and its whole job is to read the task and pick the right skill. It looks at the Post Type, then the pillar, then whether the task has a keyword and a blog post, which tells it the post is teaching something. If what I've written in the brief says something different, my words win.

This is the skill that makes the rest work. Without it, I'd be back to choosing the writer myself.

2. Talking heads from my own idea

When I have an idea for a talking head, I brain dump it into the task. The Scriptwriter uses this skill to turn my brain dump into a script spine, which is the structure of the video. I review the spine, and only then does it write the script and a short caption.

The rule at the top of this skill is the one I care about most: AI structures, I supply the substance. It never invents a result, a number or a story. If something's missing, it asks.

3. Repurposing a Reel that's already worked

This is the second workflow in the same talking head skill, built from my business coach's system. When a Reel has already proven itself, I paste in the transcript. The Scriptwriter works out why it worked, I add my own substance on top, it maps the structure, and then it writes my version.

4. Teaching B-roll captions

On a B-roll there's no voice, so the text on screen and the caption do all the work. This skill writes four on-screen hooks for me to choose from, then a long caption that teaches the steps and ends on the keyword that sends people to the blog post. It counts the characters, because Instagram cuts a caption off at 2,200.

5. Every other short-form script

The fifth skill holds my voice rules and the persuasion structure I use for short-form scripts. The Scriptwriter reaches for it when no other skill fits the post, so there's always a skill behind the work, even for a one-off.

Next: LinkedIn

I haven't built the LinkedIn skills yet. When I do, I won't build a LinkedIn writer. I'll teach the Scriptwriter the LinkedIn formats and add them as rows in its index.

Why this works for me: the workflow

The reason one employee can do all of this is that my workflow is mapped in ClickUp, following the same framework.

Every Instagram post is a task, and every task has a Post Type field and an AI EMPLOYEE field. When I set AI EMPLOYEE to Scriptwriter and move the task to a "w/ agent" status, a small script called the dispatcher picks it up and starts the Scriptwriter. It reads the Post Type, opens its index, picks the skill and does that step. Then the task comes back to me at "w/ samantha for review".

The Post Type field is what makes the index work. The employee never has to guess what kind of post it's writing, because the task tells it.

What you need before you start

• The Claude desktop app on a paid plan, with the Claude Code tab.
• Your AI brain, with your org chart saved in it. If you haven't made one, start with How to map your AI org chart before you build any AI employees.
• An employee to teach. If you don't have one yet, start with your COO: The First AI Employee I'd Build as a Non-Techie Founder: An AI COO.
• A task manager with a custom field, if you want the index to read the type of job from the task. The steps say ClickUp, but any task manager with fields and statuses will work.

You don't need to code.

Step 1 - Find the jobs that should be one employee

Open the Claude desktop app, click the Claude Code tab and start a new chat. Click the + at the bottom of the chat and choose your AI brain folder.

If you don't have an AI brain set up yet, read How to set up your AI brain in Obsidian to get it set up first, then come back to this step.

Then paste this.

Read my AI org chart & every agent file in my AI brain.

Find the employees (or planned employees) that read the same things &
do the same kind of work in a different format, like three writers for
three types of content.

For each group, run them through my four-part test: own context,
different permissions or tools, a different model, or running at the
same time as the others. Tell me which ones should become one employee
with a skill for each format. Don't change anything yet.

What you'll see: a short list of groups, with a yes or no for each part of the test and a recommendation. If a group fails all four parts, it's one employee.

Step 2 - Write each format as a skill

Take one format at a time. In the same chat, paste this.

Let's write [FORMAT] as a skill for [EMPLOYEE NAME]. Interview me about
how I actually do it, one question at a time, like you're writing an SOP
for a new hire: what it reads first, the structure, the length, my voice
rules & what it should never do. Save it in my skill library. Show me
before you save.

What you'll see: one skill file in your skill library. Test it on a real job before you write the next one. The full walkthrough is in How to Build a Claude Skill.

Step 3 - Build the index

This is the skill that tells your employee which skill to use. Paste this.

Now write the index for [EMPLOYEE NAME] as its own skill. It's the
first thing the employee reads on every job.

It should tell the employee what to look at on the task to work out the
type of job, then give a table of each type of job & the skill it uses.
Add a rule that my own words in the brief win over the table, & that if
there's no skill for the job, it tells me & hands it back instead of
making one up. Add it to [EMPLOYEE NAME]'s agent file as the skill it
reads first. Show me before you save.

What you'll see: a table like the one at the top of this post, with your own formats in it, and one new line in the employee's agent file.

Step 4 - Connect it to your task manager

The index needs to know what kind of job it's looking at. The easiest way is a field on the task. In the same chat, paste this.

I want [EMPLOYEE NAME]'s index to read the type of job from my ClickUp
task. Tell me what dropdown field to add to my list & what the options
should be, one for each row in the index. Walk me through adding it one
click at a time, then update the index so it reads that field first.

Claude tells you when to do each of these, and exactly where to click.

1. Open the list in ClickUp.
2. Add a new custom field and choose Dropdown.
3. Name it, and add one option for each type of job in your index.

What you'll see: a new field on every task in that list. Set it on a real task, give the task to your employee and check it picks the right skill. How the task gets from ClickUp to your employee is in How to Build the Workflow for Your AI Employees.

Step 5 - Teach it the next skill

When you start something new, like LinkedIn posts, you don't build a new employee. Start a new chat, attach your AI brain folder and paste this.

[EMPLOYEE NAME] needs to learn [NEW FORMAT]. First check it against my
four-part test & tell me if it really belongs to this employee. If it
does, interview me for the skill one question at a time, save it in my
skill library, add a row to the index & add the option to the field in
my task manager. Show me before you save.

What you'll see: one new skill, one new row in the index and one new option on the field. Your employee has learned something new, and your team hasn't grown.

The catch, and the fix

The catch is that the index is only as good as the information on the task. If the type of job isn't set, or the brief says one thing and the field says another, the employee can pick the wrong skill and do a good job of the wrong thing.

The fix is two rules in the index. My words in the brief always win over the field, and if the employee can't tell which skill to use, it asks me once instead of guessing. Every time it picks wrong, the correction goes into the index as a rule, so it doesn't happen twice. If you find you're correcting the same skill again and again, the fix is in the skill file, not the chat: How to build a toolkit for your AI employees shows where those rules live.

What this looks like as a business

Every business owner you support has been told they need an agent for everything, and most of them have a pile of agents they don't use. Looking at what those agents do, merging them into a few employees and writing each process as a skill is the part they can't do for themselves. You're paid for knowing what to build and in what order.

Start with one

Pick the part of your business where you'd be tempted to build three agents, and run the Step 1 prompt on it before you build anything.

If you haven't built your COO yet, that's the foundation for your team: start here.