Most people use AI for content like one tired assistant.
They paste in a half-baked idea and ask for "10 post ideas."
Then they ask for hooks.
Then they ask for captions.
Then they ask for images.
Then they get annoyed because the style changed three times and the final post sounds like it was written by a hotel chatbot.
That is not an AI content system.
That is a long prompt with a wig on.
The better setup is simple:
Give your AI a tiny content team.
Not one magic agent.
Not a giant prompt.
A team.
Each role gets one job. Each role passes the work to the next role. You stay the editor in chief.
I saw @seb.ai share a clean version of this idea as an "AI Content Team OS." I liked the frame, but I would change two things before handing it to a real creator or small business owner:
Make the approval gates clearer.
Use the right MCP stack for creating, designing, and posting.
Here is the version I would build.
Here is the carousel version I am posting with it:
The basic idea
Your AI content team has seven roles:
Manager
Researcher
Hook writer
Script writer
Designer
Publisher
Analyst
The manager runs the workflow.
The researcher finds the raw material.
The hook writer makes the idea clickable.
The script writer turns one hook into the actual post.
The designer makes the visual direction.
The publisher packages the post and prepares it for the right channel.
The analyst reads the results and tells the team what to do more of next time.
You do not need a big company to do this.
You need a folder.
The folder setup
If you are using Claude Code, you can use the .claude/agents pattern. If you are using Codex, Cursor, or another agent setup, use the same structure with whatever agent folder your tool supports.
The folder can be this simple:
content-team/
CLAUDE.md
brand/
voice.md
offers.md
swipe-file.md
.claude/
agents/
manager.md
researcher.md
hook-writer.md
script-writer.md
designer.md
publisher.md
analyst.md
pipeline/
ideas.md
calendar.md
drafts/
published.md
performance.mdThat is enough.
The system does not get smarter because the folder looks fancy. It gets smarter because the work has somewhere to live.
Your ideas stop disappearing in chat history.
Your brand voice stops being something you explain from scratch every morning.
Your results stop living only in your memory.
What goes in the main file
Your main file is the boss file. In Claude Code, that is usually CLAUDE.md.
Put the simple rules there:
You are my AI content team.
Use the agents in this folder to run the content workflow.
The manager coordinates the work.
The researcher finds source-backed ideas.
The hook writer creates hooks.
The script writer creates the post.
The designer creates visual direction and image prompts.
The publisher prepares the final package.
The analyst reviews results and updates performance notes.
Do not publish anything unless I clearly approve the final post.
Do not make up numbers, quotes, sources, or results.
Use brand/voice.md before writing.
Use brand/offers.md before making a call to action.
Use pipeline/performance.md before recommending what to repeat.That last part matters.
AI is very good at sounding confident.
That does not mean it knows what happened.
If your researcher is allowed to invent trends, your whole content machine gets worse.
If your analyst is allowed to invent results, your next 20 posts are built on fake feedback.
The rule is boring and important:
No fake data.
The seven roles
Here is how I would define each role.
1. Manager
The manager is not the best writer.
The manager is the traffic controller.
Its job is to ask, "What stage are we in, who needs to work next, and what needs my approval?"
Give it this job:
You coordinate the content workflow.
Start by reading the brand files and the current pipeline files.
Then decide which agent should work next.
Keep the output moving from idea to research to hook to draft to visual package to publish plan to performance review.
Stop for human approval after the idea is chosen and again before anything is published.This keeps you from asking one AI to do every job in one giant blur.
2. Researcher
The researcher should not write posts.
It should bring back the raw material.
Give it this job:
Find content ideas for this brand.
Use real sources, real audience problems, real questions, and real platform signals.
For each idea, include:
- the core topic
- the audience pain
- the reason it matters now
- the source or proof
- a plain-English angle
Do not write hooks yet.
Do not write the full post.That last line saves you a lot of mess.
When the researcher starts writing posts, it gets too broad. Keep it in its lane.
3. Hook Writer
The hook writer gets one researched idea and turns it into options.
Not 50 random hooks.
Ten usable hooks in different styles.
Give it this job:
Turn one approved idea into 10 hooks.
Use the brand voice file.
Make the hooks clear, specific, and easy to understand fast.
Avoid fake urgency, fake numbers, and vague hype.
For each hook, explain the promise in one sentence.I like making the hook writer explain the promise because it catches weak hooks.
If the hook cannot explain what the reader gets, it is probably just a clever line.
Clever lines are not the goal.
The goal is getting the right person to stop and care.
4. Script Writer
The script writer turns one approved hook into the actual asset.
That could be:
a Substack section
an Instagram caption
a Reel script
a carousel outline
a thread
an email
Give it this job:
Turn the approved hook into one complete content draft.
Use the chosen channel format.
Keep the writing simple and direct.
Use the source notes from the researcher.
Use the brand voice file.
Include the offer only if it fits naturally.
End with one clear next step for the reader.This is where most people go wrong.
They ask AI to make "content."
That is too vague.
Ask for one channel, one format, one reader, and one next step.
5. Designer
The designer does not just "make it pretty."
The designer decides how the idea should look.
For a carousel, it creates slide direction.
For a Reel, it creates on-screen text and b-roll ideas.
For a Substack article, it creates the cover concept and any supporting visuals.
Give it this job:
Create the visual package for the approved draft.
Include:
- cover or thumbnail idea
- image prompt
- carousel slide direction if needed
- on-screen text if needed
- brand colors or visual notes
- alt text for accessibility
Keep visuals readable on a phone.
Do not use tiny text.
Do not create visuals that promise something the content does not deliver.This is also where I would bring in the right creation MCPs.
For content creation and image generation, I would start with Xenith Studio and PRISM.
Xenith Studio is the better fit when the task needs a studio-style creative flow: content, visuals, image direction, and review in one place.
PRISM is the better fit when you want to turn one source idea into multiple post-ready assets like carousels, captions, article visuals, and repurposed social posts.
You can still use a plain image generator for one-off images.
But if the goal is a repeatable content team, I want the visual system close to the content system.
6. Publisher
The publisher is the most dangerous role if you set it up wrong.
It should prepare the post.
It should not make the final judgment.
Give it this job:
Prepare the approved content for publishing.
Check:
- channel
- title or hook
- caption
- link
- image
- alt text
- tags
- schedule time
- final approval status
If approval is missing, stop and ask.
Do not publish without approval.For posting MCPs, I would test these three first:
TinyPoster for simple post handoff and scheduling.
Xenith Studio for content studio workflows where creation, approval, and posting need to stay together.
PRISM for source-to-publish workflows where one idea becomes many channel-ready assets.
That is the stack I would use before I started duct-taping a random scheduler to a random chat.
Posting is not just "send this."
Posting includes channel choice, format, link, tags, media, timing, and proof that the right thing went to the right place.
That is why the publisher needs a checklist.
7. Analyst
The analyst is how the team gets better.
Most creators skip this role, which is why they keep making decisions from vibes.
Give it this job:
Review performance after the content has been live long enough to learn something.
Look for:
- which topics got attention
- which hooks got clicks or saves
- which posts brought subscribers, leads, sales, or replies
- which formats were weak
- what to repeat next week
Update pipeline/performance.md with the lesson.
Do not invent results.
If data is missing, say what is missing.The analyst is not there to make you feel good.
It is there to stop you from repeating weak content because you liked the idea.
The MCP stack I would connect
MCPs are the connectors that let your AI use outside tools.
For this setup, I would keep the stack small at first.
Use basic research and file access so the team can read sources and update its own pipeline.
Then add the creator tools:
Xenith Studio for content creation, image generation, visual packaging, and review.
PRISM for repurposing, carousels, captions, article assets, and source-to-content workflows.
Then add the posting tools:
TinyPoster
Xenith Studio
PRISM
The point is not to connect every tool you own.
The point is to give each role the tools it actually needs.
The researcher needs sources.
The designer needs image and visual tools.
The publisher needs posting tools.
The analyst needs results.
That is it.
The 30-minute build
Here is how I would set this up fast.
Minutes 0 to 5: Make the folder
Create the folder and the files.
Do not overthink the structure.
Your first version can be ugly. It just needs a place for the team to work.
Minutes 5 to 10: Write the brand file
In brand/voice.md, write:
who you help
what you teach
words you use
words you avoid
examples of posts that sound like you
examples of posts that do not sound like you
This file is more important than most of the agent prompts.
Bad brand input makes every role worse.
Minutes 10 to 18: Write the seven agent files
Do not make them long.
Each file should say:
the role
the inputs it uses
the output it creates
what it is not allowed to do
when it must stop for approval
Short beats clever here.
Minutes 18 to 25: Connect the MCPs
Add only what you need for the first content cycle.
For a starter setup:
research access
filesystem access
Xenith Studio
PRISM
TinyPoster
If you cannot connect everything yet, start manually.
The workflow matters more than automation on day one.
Minutes 25 to 30: Run one post through it
Ask the manager to run the full workflow for one post.
Not a week of content.
Not 30 ideas.
One post.
Use this prompt:
Use the content-team folder.
Run one full content cycle for my brand.
Start with three researched ideas.
Ask me to pick one.
Then create hooks for that idea.
Ask me to pick one.
Then draft the post.
Then create the visual package.
Then prepare the publish checklist.
Stop before publishing until I clearly approve the final package.That prompt shows you whether the system works.
If the manager skips a step, tighten the manager file.
If the writing sounds wrong, fix the brand voice file.
If the ideas are weak, fix the researcher file.
If the visuals are confusing, fix the designer file.
Do not rebuild the whole thing. Fix the role that failed.
The approval gates
There are two places where the AI should stop.
First, after research.
You should approve the idea before the team spends time writing the wrong post.
Second, before publishing.
You should approve the final package before anything goes live.
That does not make the system slow.
It keeps the human judgment where it belongs.
The AI does the handoff work.
You make the taste call.
What this replaces
This replaces the messy way most people make content with AI:
one chat for ideas
one chat for hooks
one chat for captions
one random image prompt
one forgotten spreadsheet
one social scheduler
no memory
no feedback loop
That system gets worse over time because nothing learns.
The folder version gets better because every cycle leaves a trace.
The idea goes into ideas.md.
The post goes into published.md.
The results go into performance.md.
The next round uses that history.
That is the part people miss.
AI content gets better when your system remembers what worked.
The beginner version
If all of this feels too technical, start with three roles:
Researcher
Writer
Publisher
That alone will clean up your process.
The researcher finds ideas.
The writer makes the post.
The publisher turns it into a final checklist and stops for your approval.
Add designer when you start making carousels or images.
Add analyst when you have results to review.
Add manager when the handoffs get annoying.
You do not need the whole machine on day one.
You need one repeatable path from idea to post.
The real win
The win is not "AI made me more content."
More content is not automatically useful.
The win is that your content stops depending on your mood.
You do not need to wake up inspired.
You do not need to remember your own brand rules.
You do not need to rebuild the process from scratch every time.
You open the folder. You ask the manager to run the workflow. You approve the important parts.
That is the whole point.
Not replacing you.
Removing the parts that make you avoid posting.
Source note
This article was inspired by an AI content team outline from @seb.ai. I rewrote the workflow for a beginner creator setup, with clearer approval gates and the MCP stack I would use now: TinyPoster, Xenith Studio, and PRISM for posting, plus Xenith Studio and PRISM for creation and image work.
Keep reading
B*tchwork my AI Did for Me - Part 10: Told Me EXACTLY How to Go Viral on Instagram
B*tchwork my AI Did for Me - Part 12: I Built an AI That Posts Breaking AI News
B*tchwork my AI Did for Me - Part 13: I Had My AI Build a Pinterest Traffic Engine
Tell me in the comments which role you want your AI to take over first: research, hooks, writing, visuals, publishing, or analysis.