How I built a content team of AI agents in Claude Code

My content team has one human. Me.

Everyone else is an AI agent. Here's a normal morning:

My Team Lead sends me my competitors' viral videos with fresh ideas pulled from them, news from the big AI players, viral posts from X, and the posts I saved on X.
From all of that, it proposes 10 topics. Each one comes with sources.
I pick one.
A Fact Checker verifies the claims. A Scriptwriter drafts the script in Armenian, in my voice.
I read it, fix what doesn't sound like me, and film.

Below is the whole system, and how to build your own version in Claude Code. You don't need to write code.

The point of all this: AI output is only as good as the choices you make. Agents gather, check and draft. A human decides what gets made, how it's said and when it ships. Filming and final calls stay with me, on purpose.

What's inside

Part 1Why separate agents
Part 2Meet the team
Part 3The daily workflow
Part 4What I keep for myself
Part 5Build it in Claude Code
Part 6Starter prompts

Part 1

Why separate agents beat one giant chat

Imagine hiring one person and telling them: you're the researcher, the fact checker, the writer and the editor. They'll burn out by lunch. Worse, they'll never catch their own mistakes, because they made them.

One giant chat is that employee. Every new request makes the conversation heavier. Yesterday's links bleed into today's ideas, and at some point the model starts confidently describing things that never happened.

Separate agents work like a good newsroom:

Clean heads. In Claude Code, every subagent runs in its own context window and hands back only the result. The 40 links your Researcher dug up never clutter your Scriptwriter's thinking.
One role, one set of rules. You decide which tools each agent gets. The Fact Checker can't edit files. The Scriptwriter can't browse the web, so it can only write from verified facts.
Outside review. The Fact Checker didn't write what it's checking. Same reason newspapers don't let reporters edit themselves.

Part 2

Meet the team

Team Lead

DoesRuns the morning sweep, proposes 10 topics, splits the work and collects everything for my review.

ReceivesCLAUDE.md, the source list, my instructions.

Hands off10 topics to me, my pick to the Fact Checker, verified facts to the Scriptwriter.

Researcher

DoesFinds the material and attaches a source link to every single item.

ReceivesThe Team Lead's brief and the source list.

Hands offA list of findings, each with a link and a publish date.

Fact Checker

DoesOpens every source and confirms it actually says what we're about to say.

ReceivesThe chosen topic and its sources.

Hands offA verdict table (verified, wrong, unclear) plus a clean list of safe facts.

Scriptwriter

DoesWrites the script in my language and my voice.

ReceivesCLAUDE.md, my example posts, the verified facts, my chosen angle.

Hands offHook, body, CTA, on screen text and two backup hooks.

Me

I doPick the topic, decide what stays and what changes, film, make the final call.

Example"This hook isn't how I talk. Rewrite it."

Part 3

The daily workflow: from idea to a script I can film

08:30The routine starts on its own

I don't work before 9. The agents do. The Team Lead kicks off the morning research without me.

Step 2The Researcher gathers material

Competitor videos, AI news, viral X posts, my saved posts. Every item has a link and a date. No link, no item.

Step 3The Team Lead proposes 10 topics

Each one with why it matters now, an angle for my audience, a hook idea and sources. Saved to the daily folder.

09:00I choose

I read the 10 and pick one. Some days I pick none, and that's a decision too.

Step 5The Fact Checker verifies

Opens every source. Numbers, names, dates. Returns a verdict table and a list of safe facts only.

Step 6The Scriptwriter drafts

In Armenian, in my voice, from verified facts only.

Step 7I edit and film

I read the script out loud. Anything that doesn't sound like me goes back. Then the camera goes on.

Part 4

What I keep for myself, and why

A film set has a camera operator, a gaffer and an editor. But only the director yells "Action." On my team, I'm the director.

Choosing the topic. An agent can find 10 good topics. It doesn't know what people asked me in DMs yesterday, or which question my audience actually needs answered today.
The voice. If the hook doesn't sound like me, it won't work, however polished it is.
Final sign off on facts. The Fact Checker flags the risk. I make the call. It's my face on the video.
Filming. People trust a person, not a pipeline. My face, my voice and my energy don't need automating.
Publishing. Agents never post anything. Ever.

That split is deliberate. The better you choose, the better the team performs. AI speeds up the road. You still pick the destination.

Part 5

How to build it in Claude Code

Now the practical part. You don't need to be a developer. The whole team is a handful of text files in one folder. Here's the full structure:

Folder structure

content_team/
├── CLAUDE.md          ← voice, audience, rules
├── sources.md         ← where to look
├── examples/          ← your best posts and scripts
├── daily/             ← each morning's 10 topics
├── scripts/           ← finished scripts
└── .claude/
    ├── settings.json
    └── agents/
        ├── teamlead.md
        ├── researcher.md
        ├── factchecker.md
        └── scriptwriter.md

1. A project folder for the team

Create a content_team folder and open it in Claude Code. Everything lives here: rules, sources, daily topics, scripts. In sources.md you list competitor accounts, X accounts and news sources. In examples you drop the posts and scripts that sound most like you.

2. CLAUDE.md: the team handbook

This is the file everyone reads. Claude Code loads it automatically, and your custom subagents receive it too when they start work. Put your voice, audience, language and banned words here. Write it once, and the whole team follows it.

⚡

Keep it short. One page is plenty. Long rules turn into noise for an agent, the same way a 40 page onboarding PDF does for a new hire.

3. One file per agent

Each agent is a single Markdown file in .claude/agents/. The top is a short ID card: name and description are required, and tools controls what the agent can use. Below that, you describe its role in plain English.

Write the files yourself or just ask Claude to create them. Take the description seriously: it's how Claude decides who to call and when.

4. How the Team Lead calls the others

The Team Lead runs as your main session. Add one line to .claude/settings.json and every time you open the project, you're talking to the Team Lead:

.claude/settings.json

{
  "agent": "teamlead"
}

Its tools field says Agent(researcher, factchecker, scriptwriter), which means it can call those three and nobody else. The agents never talk to each other directly. Each one reports back to the Team Lead, who decides what to pass along. Exactly what a good project manager does.

5. Run the morning routine on a schedule

In the Claude desktop app, open Routines, click New routine and pick a mode:

Local. Runs on your own computer, but only while the app is open and the machine is awake. If it was asleep, that day's run is skipped.
Cloud. Runs in Anthropic's cloud even with your laptop closed. The project needs to live in a GitHub repo, and it's available on Pro, Max, Team and Enterprise plans. From the terminal, /schedule creates the same thing.

Paste the daily kickoff prompt (below) as the routine's prompt, set it for 08:30, and save. Routines are still a research preview, so check during the first week that it really fires every morning.

⚠️

About sources. Instagram and X often won't open through a plain web search. For competitor videos and your saved X posts, either connect the right connector or drop the links into sources.md each morning. An agent should never "remember" a page it couldn't open. That's exactly why the Researcher's rule is no link, no item.

Part 6

Starter prompts

Copy, replace what's in square brackets, and go. The Scriptwriter writes in whatever language you set in CLAUDE.md.

1. Project setup + CLAUDE.md

Set up a content team project in this folder.

Create:
1. CLAUDE.md with my brand voice, audience and rules (details below)
2. sources.md with sections: Competitor accounts, X accounts, Saved X posts, News sources
3. Folders: examples/, daily/, scripts/
4. An empty .claude/agents/ folder
5. .claude/settings.json with {"agent": "teamlead"}

My details:
Niche: [YOUR NICHE]
Audience: [YOUR AUDIENCE]
Language of my videos: [LANGUAGE]
Platforms: [INSTAGRAM / TIKTOK / LINKEDIN]
3 posts that sound like me:
[POST 1]
[POST 2]
[POST 3]
Words I never use: [LIST]

In CLAUDE.md, write a Voice section learned from my 3 posts: tone, sentence length, words I use, words I avoid. Keep CLAUDE.md to one page. Show it to me before saving.

Day one, once. Creates the folder and the team handbook.

2. .claude/agents/teamlead.md

---
name: teamlead
description: Content team orchestrator. Runs the morning research, proposes 10 topics, routes work to researcher, factchecker and scriptwriter, and collects results for my review.
tools: Agent(researcher, factchecker, scriptwriter), Read, Write, WebSearch, WebFetch
---

You are the Team Lead of a content team. The only human on the team is me, the creative director. You prepare, route and collect. You never pick the final topic, approve a script or publish anything.

Inputs: CLAUDE.md, sources.md, my instructions for today.

Morning routine:
1. Send the researcher to collect today's material from sources.md and the web.
2. From the research, propose exactly 10 topics.
3. Save them to daily/[TODAY]/topics.md and stop. Wait for my pick.

When I pick a topic:
1. Send the topic and its sources to the factchecker.
2. Send only the verified facts and my angle to the scriptwriter.
3. Save the script to scripts/[TODAY]_[TOPIC].md and show me the fact check table and the script.

Topic format (one block each):
Topic | Why now (1 line) | Angle for my audience | Hook idea | Source links

Rules:
1. No topic without at least one source link.
2. Never invent numbers, quotes or dates.
3. Give each agent only what it needs.
4. If something is missing, say so. Do not guess.

The brain of the team. The only agent allowed to call the others.

3. .claude/agents/researcher.md

---
name: researcher
description: Collects daily content material: competitor viral videos, AI industry news, viral X posts and my saved X posts. Always attaches a source link to every item.
tools: Read, Write, WebSearch, WebFetch
---

You are the Researcher. You find material. You do not pick topics or write scripts.

Inputs: CLAUDE.md, sources.md, the Team Lead's brief.

Collect from the last 48 hours:
1. Viral videos from competitor accounts in sources.md: what it says, why it worked, how we could do it better.
2. News from the major AI companies. Official blogs and announcements first.
3. Viral X posts about AI in my niche.
4. My saved X posts listed in sources.md.

Output (one block per item):
Category | Headline (1 line) | What happened (2 lines) | Why my audience cares | Source URL | Date published

Rules:
1. Every item has a source URL. No URL, no item.
2. Prefer original sources over reposts and aggregators.
3. Mark anything you could not confirm as UNCONFIRMED.
4. If a page will not open, say so. Never fill gaps from memory.

Gathers the material. Its main rule: every item has a link.

4. .claude/agents/factchecker.md

---
name: factchecker
description: Verifies every claim, number, name and date in a chosen topic against its sources before any script is written.
tools: Read, WebSearch, WebFetch
---

You are the Fact Checker. You protect the creator's credibility.

Input: one topic with its research notes and source links.

For every claim:
1. Open the source and confirm it says what the notes say.
2. Check the date. Flag anything older than [7] days as OLD.
3. For big claims (launches, numbers, quotes), find one independent confirmation.

Output:
Claim | Status (VERIFIED / WRONG / UNCLEAR) | Source URL | Correction
Then a list titled "Safe facts for the script" with verified facts only.

Rules:
1. Never mark a claim VERIFIED without opening the source.
2. If a source is unreachable, mark it UNCLEAR. Do not guess.
3. You check. You do not rewrite or soften.

Read only on purpose. It checks, it never edits.

5. .claude/agents/scriptwriter.md

---
name: scriptwriter
description: Writes short video scripts in my voice and language from verified facts only. Use after the factchecker has cleared a topic.
tools: Read, Write
---

You are the Scriptwriter. You write like me, not like an AI.

Inputs:
1. CLAUDE.md: my voice, audience, language, banned words
2. examples/: posts and scripts that sound like me
3. Verified facts from the factchecker
4. The angle I chose

Write in [YOUR LANGUAGE]. Think in that language. Do not translate from English.

Output:
HOOK (first 3 seconds, 1 line)
BODY (spoken lines, short sentences, one idea per line)
CTA (1 line)
ON SCREEN TEXT (3 to 5 short captions)
LENGTH (estimated seconds)
Then 2 alternative hooks.

Rules:
1. Use only verified facts. No new numbers, names or claims.
2. Match my voice from examples/. If a line sounds like a press release, rewrite it.
3. Keep it filmable: I say this to camera.

No web access on purpose, so it can only write from verified facts.

6. Daily kickoff

Run the morning research and give me 10 topics with sources.
Use sources.md and the last 48 hours. Save to daily/[TODAY]/topics.md.

Every morning. This is your routine's prompt.

7. Topic → script

I pick topic #[NUMBER]. My angle: [ONE LINE].
Send it to the factchecker first, then give only the verified facts to the scriptwriter.
Show me the fact check table and the script together.

When you've picked a topic and want a script.

8. Revision

This hook isn't how I talk. Rewrite it.
Here's how I'd say it: "[YOUR WORDS]"
Keep this wording style for the whole script, not just the hook.
Also cut: "[LINE YOU DON'T LIKE]". Don't touch the facts.
Then add one rule to the Voice section of CLAUDE.md so this doesn't happen again.

When the script doesn't sound like you. The last line teaches the team so the same mistake doesn't come back.

Quick recap

Instead of one giant chat: four agents, each with one job and its own tools.
The Team Lead is the main session and the only one that calls the others.
The Researcher's rule: no link, no item. The Fact Checker clears facts before any script.
CLAUDE.md is the handbook everyone reads. Routines start your morning for you.
Choosing, voice, filming and the final call stay with you.

Want to build this team with me?

In my course «AI Agent-ների քո թիմը» (Your AI Agent Team), we build this exact content team together: your CLAUDE.md, your agents, your morning routine. You leave with a team that works, not a page of notes.

Join the course →