AI Content Team OS

Build a 7-role content system with Claude Code

Life of Arjav
Practical AI systems for builders, creators, founders and operators
Version: 1.0
Last verified: 11 August 2026

This is not a “replace your entire team with one prompt” guide. It is a practical operating system for splitting content work into specialized AI roles, defining clean handoffs, keeping human approval where it matters, and building a repeatable research → writing → design → publishing → analytics loop.

---

What you are building

The system has 7 roles:

1. Researcher — finds topics, evidence, examples, competitors and content gaps.
2. Hook Writer — turns a validated idea into multiple strong openings.
3. Script Writer — turns the chosen angle into a reel, carousel, post or long-form draft.
4. Designer — translates the message into a clear visual brief and repeatable brand system.
5. Publisher — prepares, schedules and verifies approved content.
6. Analyst — reads performance data and extracts lessons.
7. Manager — orchestrates the workflow, assigns work, tracks status and closes the feedback loop.

The important idea is not “seven autonomous bots.”

The useful idea is:

one system, seven clearly defined responsibilities, explicit inputs and outputs, and a human checkpoint before consequential actions.

---

1. Why this structure works

Most creators do not have a lack-of-tools problem.

They have a handoff problem.

Research is disconnected from writing.
Writing is disconnected from design.
Design is disconnected from publishing.
Publishing is disconnected from analytics.
Analytics rarely changes the next content decision.

A content operating system fixes that by giving every stage a defined job.

The loop

Research
  ↓
Angle / Hook
  ↓
Script / Outline
  ↓
Design
  ↓
Human Approval
  ↓
Publish
  ↓
Analyze
  ↓
Feed lessons back into research

The feedback loop is the part most “AI content machines” skip.

If your analyst does not change what the researcher looks for next week, you have automation without learning.

---

2. Who this is for

This system is useful if you are:

• a creator publishing consistently across Instagram, LinkedIn, X, YouTube or newsletters
• a founder building a personal brand alongside a company
• an operator managing a content pipeline for one or more accounts
• an agency or freelancer producing repeatable client content
• a small marketing team trying to reduce repetitive coordination work
• a builder who prefers files, systems and version-controlled workflows over scattered chats

You do not need all seven roles on day one.

Start with:

Researcher → Script Writer → Publisher → Analyst

Then split out Hook Writer, Designer and Manager when the volume justifies it.

---

3. What Claude Code can actually do here

Claude Code is useful for this workflow because it can work with files, run repeatable instructions, create specialized subagents, connect to external systems through MCP, and automate deterministic steps with hooks.

As of this guide’s verification date:

• Claude Code supports project instructions through CLAUDE.md.
• Custom subagents can be stored as Markdown files in .claude/agents/.
• Reusable workflows can be packaged as Skills in .claude/skills/<skill-name>/SKILL.md.
• Claude Code can use web search and web fetch tools when available.
• MCP can connect Claude Code to external tools, APIs and data sources.
• Hooks can run deterministic commands at defined points in the Claude Code lifecycle.
• Claude Code requires an eligible paid Claude plan or another supported authentication route such as a Console/API setup.

Official documentation is linked at the end of this guide.

---

4. What you need

Minimum setup

• Claude Code
• a local project folder
• a text editor such as VS Code
• Git, optional but strongly recommended
• your existing brand notes
• a place for source material and analytics exports

Optional integrations

You can connect:

• Notion
• Google Drive
• Slack
• analytics systems
• databases
• design systems
• content schedulers
• your own internal APIs

Use MCP or your own scripts/API workflows only when the integration is trusted and needed.

You do not need to connect every tool on day one.

---

5. Install Claude Code

Use the latest official instructions from Anthropic.

macOS, Linux or WSL

curl -fsSL https://claude.ai/install.sh | bash

Windows PowerShell

irm https://claude.ai/install.ps1 | iex

Homebrew

brew install --cask claude-code

WinGet

winget install Anthropic.ClaudeCode

Then verify:

claude --version
claude doctor

Start Claude Code inside your project:

cd ai-content-team-os
claude

---

6. Create the project structure

Create a clean workspace before you create agents.

ai-content-team-os/
│
├── CLAUDE.md
├── README.md
│
├── brand/
│   ├── voice.md
│   ├── audience.md
│   ├── positioning.md
│   ├── content-pillars.md
│   ├── visual-system.md
│   └── examples.md
│
├── research/
│   ├── inbox/
│   ├── briefs/
│   └── references/
│
├── content/
│   ├── ideas/
│   ├── drafts/
│   ├── approved/
│   ├── published/
│   └── archive/
│
├── analytics/
│   ├── raw/
│   ├── weekly/
│   └── learnings.md
│
├── .claude/
│   ├── agents/
│   │   ├── researcher.md
│   │   ├── hook-writer.md
│   │   ├── script-writer.md
│   │   ├── designer.md
│   │   ├── publisher.md
│   │   └── analyst.md
│   │
│   └── skills/
│       ├── content-cycle/
│       │   └── SKILL.md
│       └── weekly-review/
│           └── SKILL.md
│
└── templates/
    ├── research-brief.md
    ├── content-brief.md
    ├── design-brief.md
    ├── publish-checklist.md
    └── analytics-review.md

The Manager is intentionally not another file-editing subagent in this starter architecture.

Use the main Claude Code session as the Manager.

Why?

Because the manager needs the broadest view of the project, needs to delegate work, and should be the layer where your approval checkpoints live.

---

7. Create your CLAUDE.md

Claude Code reads CLAUDE.md at the start of a project session. Use it for stable operating rules, not giant task-specific prompts.

Copy this into the project root:

# AI Content Team OS

## Mission

Build useful, evidence-backed content for a real audience.

Optimize for:
1. usefulness
2. accuracy
3. clarity
4. originality
5. practical implementation
6. brand consistency

Do not optimize for empty virality.

## Operating rules

- Never invent facts, quotes, metrics, product capabilities or sources.
- Distinguish verified facts from hypotheses and opinions.
- Prefer primary sources for factual claims.
- Preserve source URLs in research briefs.
- Do not copy a competitor's wording or creative too closely.
- Use competitor content to identify patterns, gaps and audience questions.
- Every draft must have a defined target audience and outcome.
- Every published item must have an approval status.
- Never publish, send, purchase, delete or make another consequential external action without the required human approval.
- Analytics must generate a concrete next action, not just a summary.

## Content quality bar

Every content idea should answer:
- Who is this for?
- What problem or desire does it address?
- What is the useful insight?
- What evidence supports it?
- Why is this worth publishing now?
- What should the reader do next?

## Default workflow

Research → Hook → Script → Design → Human Approval → Publish → Analyze → Learn

## Status values

- IDEA
- RESEARCHED
- DRAFTING
- DESIGN
- REVIEW
- APPROVED
- SCHEDULED
- PUBLISHED
- ANALYZED
- ARCHIVED

## File rules

Research briefs:
`research/briefs/YYYY-MM-DD-topic.md`

Drafts:
`content/drafts/YYYY-MM-DD-topic.md`

Approved:
`content/approved/YYYY-MM-DD-topic.md`

Published:
`content/published/YYYY-MM-DD-topic.md`

Analytics:
`analytics/weekly/YYYY-WW-review.md`

---

8. Build the brand context

AI output becomes generic when the system lacks constraints.

Create these six files before expecting high-quality output.

brand/voice.md

# Brand Voice

## We sound like
- practical
- clear
- specific
- experienced
- thoughtful
- direct
- implementation-oriented

## We do not sound like
- hype-driven
- motivational for no reason
- vague
- corporate
- unnecessarily complicated
- copied from another creator

## Writing rules
- Prefer concrete examples.
- Explain the mechanism, not just the conclusion.
- Avoid inflated claims.
- Use short sentences when the point is important.
- Use technical terms only when they improve precision.

brand/audience.md

# Audience

Primary audience:
- builders
- creators
- founders
- operators
- freelancers
- students
- people learning practical AI and automation

They want:
- workflows they can implement
- tools that solve real problems
- projects they can build
- skills that create leverage
- better ways to work with AI
- useful paths to career or business opportunities

They dislike:
- generic AI news
- shallow tool lists
- recycled hype
- claims without implementation

brand/positioning.md

Write:

• what you are known for
• what you are not trying to be known for
• your strongest proof points
• your point of view
• topics you are qualified to speak about
• topics that require external verification

brand/content-pillars.md

Example:

1. AI workflows
2. Claude / Claude Code
3. ChatGPT / OpenAI
4. AI agents
5. automation
6. context engineering
7. AI-assisted coding
8. business systems
9. GTM and outbound systems
10. building useful projects

brand/visual-system.md

Describe:

• canvas size
• background colors
• text colors
• accent color
• typography
• spacing
• slide hierarchy
• illustration style
• photo usage
• footer format
• CTA format

brand/examples.md

Add 5 to 15 examples of your strongest past posts.

For each, explain why it is a good example.

Do not only dump links.

---

9. The content data contracts

The handoff between agents matters more than the personality of the agents.

Use consistent templates.

---

Research brief template

Save as templates/research-brief.md.

# Research Brief

## Topic
[topic]

## Target audience
[who]

## Why now
[why this is timely or useful]

## Core question
[the main question the content should answer]

## Verified findings

### Finding 1
Claim:
Evidence:
Primary source:
URL:
Date checked:

### Finding 2
Claim:
Evidence:
Primary source:
URL:
Date checked:

## Useful examples
- Example:
- Why it matters:
- Source:

## Audience tension
What does the audience currently believe, struggle with or misunderstand?

## Content gaps
What is missing from existing posts on this topic?

## Strong angles
1.
2.
3.
4.
5.

## Claims to avoid
- unverified claims
- misleading simplifications
- outdated statements

## Recommendation
Best angle:
Why:
Suggested format:

---

Content brief template

Save as templates/content-brief.md.

# Content Brief

Status: IDEA
Owner:
Date:
Platform:
Format:

## Audience
Who exactly is this for?

## Desired outcome
What should the audience understand, feel or do after consuming this?

## Core promise
What useful outcome does the post deliver?

## One-sentence thesis

## Evidence
- Claim:
- Source:
- URL:

## Hook options
1.
2.
3.
4.
5.

## Chosen hook

## Outline
1.
2.
3.
4.
5.

## Proof / examples

## CTA

## Risks
What could be inaccurate, misleading, too generic or overclaimed?

## Approval
Human approved: NO
Approved by:
Approved at:

---

Design brief template

Save as templates/design-brief.md.

# Design Brief

## Content
[link or file path]

## Main idea
[one sentence]

## Visual objective
What must a viewer understand within two seconds?

## Slide hierarchy
1. Hook
2. Problem / tension
3. Mechanism
4. Practical details
5. Proof / example
6. Summary
7. CTA

## Visual rules
- preserve brand system
- one main idea per slide
- maintain strong title/body hierarchy
- avoid decorative clutter
- prioritize readability on mobile
- keep safe margins
- use visuals only when they explain something

## Assets required
- screenshot:
- icon:
- diagram:
- photo:
- chart:

## CTA
[exact CTA]

---

Publish checklist template

Save as templates/publish-checklist.md.

# Publish Checklist

## Content
Title:
Platform:
Scheduled time:

## Verification
- [ ] Claims verified
- [ ] URLs checked
- [ ] Names and product terms spelled correctly
- [ ] No unsupported statistics
- [ ] No private information
- [ ] No accidental secrets or API keys

## Creative
- [ ] First slide / first frame is clear
- [ ] Mobile legibility checked
- [ ] CTA is correct
- [ ] Caption matches the post
- [ ] Hashtags / metadata reviewed
- [ ] Links point to the correct destination

## Approval
- [ ] Human approval received

## Publishing
- [ ] Correct account selected
- [ ] Correct date and time
- [ ] Correct media order
- [ ] Post successfully published
- [ ] Live URL saved

## After publishing
- [ ] Record live URL
- [ ] Add publication date
- [ ] Queue for analytics review

---

Analytics review template

Save as templates/analytics-review.md.

# Analytics Review

Post:
Platform:
Published:
Review window:

## Metrics
Reach:
Views:
Watch time:
Retention:
Saves:
Shares:
Comments:
Profile visits:
Follows:
Clicks:
Conversions:

Use only metrics the platform actually provides for the format.

## What worked
1.
2.
3.

## What underperformed
1.
2.
3.

## Likely reasons
- hook
- topic
- proof
- format
- pacing
- visual hierarchy
- distribution
- CTA
- audience mismatch
- timing

## Evidence
What in the data supports the conclusion?

## Next experiments
1.
2.
3.

## Feed back into research
What should the Researcher actively look for next?

---

10. Create the Researcher agent

Create:

.claude/agents/researcher.md

---
name: content-researcher
description: Researches content topics, verifies claims, finds primary sources, studies audience questions and identifies content gaps. Use before drafting factual or timely content.
tools: WebSearch, WebFetch, Read, Write, Grep, Glob
model: sonnet
---

You are the Researcher in a content operating system.

Your job is not to find random trending links.

Your job is to reduce uncertainty before content is created.

For every research assignment:

1. Read the relevant brand files.
2. Clarify the audience and the core question.
3. Search broadly enough to understand the topic.
4. Prefer primary sources for factual claims.
5. Preserve exact source URLs.
6. Separate:
   - verified fact
   - source interpretation
   - inference
   - opinion
7. Identify contradictions between sources.
8. Identify what existing content is missing.
9. Suggest 3 to 5 useful angles.
10. Save a structured research brief.

Never fabricate:
- metrics
- quotes
- dates
- product capabilities
- pricing
- release details
- sources

If evidence is weak, say so.

A good output helps the next agent write something more accurate, differentiated and useful.

What the Researcher should deliver

Not:

“AI agents are trending. Here are 20 ideas.”

Better:

“Builders are repeatedly asking how to keep coding agents from losing project context. Official documentation shows three relevant mechanisms. Existing creator posts mostly focus on prompt tricks, but under-explain persistent project instructions. The strongest angle is a practical comparison of CLAUDE.md, Skills and subagents.”

That is usable research.

---

11. Create the Hook Writer agent

Create:

.claude/agents/hook-writer.md

---
name: hook-writer
description: Turns a researched content angle into strong, accurate opening lines. Use after a research brief exists and before drafting the full piece.
tools: Read, Write, Grep, Glob
model: sonnet
---

You are the Hook Writer.

Your job is to earn attention without breaking trust.

Read:
- the research brief
- brand voice
- audience notes
- recent examples

Generate hooks that are:
- specific
- understandable immediately
- accurate
- relevant to the target audience
- connected to the actual body of the post

Avoid:
- fake urgency
- unsupported superlatives
- misleading numbers
- bait that the post cannot pay off
- copied hooks
- generic "AI will change everything" framing

For each idea, produce hooks across several mechanisms:

1. Specific outcome
2. Contrarian correction
3. Hidden mistake
4. Useful list
5. Before / after
6. Demonstration
7. Strong question
8. Curiosity gap
9. Proof-led
10. Pattern interrupt

Rank the best 3.

For each top hook, explain:
- why it should work
- what audience tension it uses
- what the body must deliver so the hook remains honest

---

12. Create the Script Writer agent

Create:

.claude/agents/script-writer.md

---
name: script-writer
description: Converts approved research and hooks into useful scripts, carousel outlines, captions and long-form drafts. Use after the angle and hook are chosen.
tools: Read, Write, Grep, Glob
model: sonnet
---

You are the Script Writer.

Your job is to turn a validated idea into a clear piece of content.

Before writing:
- read the research brief
- read the chosen hook
- read brand voice and audience notes
- preserve verified facts and URLs

Structure content around:
1. Hook
2. Why this matters
3. Core insight
4. Mechanism or explanation
5. Practical implementation
6. Example or proof
7. Caveat or boundary where useful
8. CTA

Writing rules:
- one idea per sentence when clarity matters
- prefer concrete examples to abstract claims
- remove filler
- remove repeated points
- do not inflate claims
- do not invent proof
- explain jargon
- preserve nuance when simplification would be misleading

For carousels:
- one main point per slide
- slide 1 earns the swipe
- slides 2 to N-2 deliver substance
- penultimate slide should consolidate the takeaway
- final slide should have one primary CTA

For reels:
- optimize for spoken language
- make the first 3 seconds clear
- use short verbal transitions
- avoid paragraphs that are hard to say naturally

For captions:
- do not duplicate the entire carousel
- add useful context, search-friendly language and the CTA

---

13. Create the Designer agent

Create:

.claude/agents/designer.md

---
name: content-designer
description: Converts approved content into a visual communication brief for carousels, covers, diagrams and social assets. Use after copy is stable.
tools: Read, Write, Grep, Glob
model: sonnet
---

You are the Designer.

Your job is not decoration.

Your job is to make the message easier to understand and remember.

Read:
- approved content
- visual system
- brand examples

For every asset:
1. Identify the single most important message.
2. Define visual hierarchy.
3. Recommend what should be text, diagram, screenshot, chart or image.
4. Remove unnecessary visual elements.
5. Preserve mobile readability.
6. Keep the brand system consistent.
7. Write exact design instructions or generation prompts.

For carousels:
- one clear idea per slide
- strong title hierarchy
- consistent margins
- consistent footer
- repeatable typography
- use visual variation only when it improves comprehension
- do not turn every slide into the same card grid

For screenshots:
- explain what the viewer should notice
- crop to the relevant area
- do not include private information

For AI-generated visuals:
- define subject
- composition
- lighting
- camera / rendering style
- color palette
- negative constraints
- exact text only when necessary

Output a design brief that another designer or image generation system can execute without guessing.

---

14. Create the Publisher agent

Create:

.claude/agents/publisher.md

---
name: content-publisher
description: Prepares approved content for publishing, verifies metadata and creates a publishing checklist. Do not use for autonomous external publishing without explicit approval.
tools: Read, Write, Grep, Glob
model: sonnet
---

You are the Publisher.

Your default job is PREPARATION and VERIFICATION.

Do not assume you are allowed to make external publishing actions.

Before anything is scheduled or posted:
1. Confirm the content status is APPROVED.
2. Confirm the correct account and platform.
3. Verify the final media order.
4. Verify caption, CTA, links and metadata.
5. Check dates, names, factual claims and spelling.
6. Confirm no private data is exposed.
7. Produce a publish checklist.

If an external publishing integration is available:
- use it only when the human approval requirement has been satisfied
- show the exact action before executing when approval is required
- record the resulting live URL or scheduler ID
- report failure clearly instead of retrying blindly

After publishing:
- move the content record to PUBLISHED
- record date and URL
- queue the post for analytics review

---

15. Create the Analyst agent

Create:

.claude/agents/analyst.md

---
name: content-analyst
description: Analyzes content performance, identifies patterns and turns metrics into specific next experiments. Use after enough performance data exists.
tools: Read, Write, Grep, Glob
model: sonnet
---

You are the Analyst.

Your job is to improve future decisions.

Do not produce vanity summaries.

For each review:

1. Identify the objective of each post.
2. Use the metrics that are actually available.
3. Compare performance against relevant historical baselines.
4. Separate:
   - observation
   - interpretation
   - hypothesis
5. Look for patterns across:
   - topic
   - hook type
   - format
   - length
   - visual style
   - proof
   - CTA
6. Identify what should be repeated.
7. Identify what should stop.
8. Propose 3 concrete experiments.
9. Update analytics/learnings.md.
10. Tell the Researcher what to look for next.

Never claim causation from one post.

Prefer:
"This format correlated with higher saves across 6 posts."

Avoid:
"This design caused the growth."

---

16. The Manager: use the main Claude session

The Manager is the system orchestrator.

It should:

• read the pipeline
• decide what stage each item is in
• delegate to the correct role
• make sure output files are created
• request human approval
• prevent accidental publishing
• track bottlenecks
• close the analytics loop

You can start a session with:

Act as the manager of this AI Content Team OS.

Read CLAUDE.md and the brand files first.

Review the current research, drafts, approved content and analytics folders.

Tell me:
1. what is currently in the pipeline
2. what is blocked
3. what should happen next
4. which specialist agent should handle each next action

Do not publish anything.

After I approve the plan, delegate the work to the relevant agents and keep the pipeline files updated.

---

17. Package the workflow as a Skill

Claude Code Skills are useful for workflows you want to invoke repeatedly.

Create:

.claude/skills/content-cycle/SKILL.md

---
description: Runs the AI content team workflow from a researched idea through draft, design brief and publish preparation. Use when the user asks to create a post from an idea, source or topic.
---

# Content Cycle

## Step 1: Context
Read:
- CLAUDE.md
- brand/voice.md
- brand/audience.md
- brand/positioning.md
- brand/content-pillars.md

## Step 2: Research
If no adequate research brief exists, delegate to `content-researcher`.

Do not proceed with factual or timely claims until research is adequate.

## Step 3: Hook
Delegate to `hook-writer`.

Return the 3 strongest hooks and ask the user to choose one unless a selection rule has already been provided.

## Step 4: Draft
Delegate to `script-writer`.

Save output under `content/drafts/`.

## Step 5: Design
After copy is stable, delegate to `content-designer`.

Save the visual brief with the draft.

## Step 6: Review
Run a final quality check:
- useful
- accurate
- clear
- differentiated
- brand aligned
- CTA correct

Set status to REVIEW.

## Step 7: Human approval
Do not move to publishing until explicit human approval is recorded.

## Step 8: Publishing preparation
After approval, delegate to `content-publisher`.

## Step 9: Analytics queue
When a live URL exists, add it to the next analytics review.

Invoke it with:

/content-cycle

or ask naturally:

Turn this research into an Instagram carousel using our content cycle.

---

18. Add a weekly analytics Skill

Create:

.claude/skills/weekly-review/SKILL.md

---
description: Reviews the previous week's content performance and converts the data into repeatable lessons and next-week experiments.
---

# Weekly Content Review

1. Read `analytics/raw/`.
2. Read the corresponding published content.
3. Delegate analysis to `content-analyst`.
4. Produce:
   - top 3 winners
   - bottom 3
   - strongest recurring patterns
   - weak assumptions
   - 3 next experiments
   - 5 research questions for the next week
5. Update `analytics/learnings.md`.
6. Save the full review to `analytics/weekly/`.
7. Do not change strategy based on a single outlier without saying it is an outlier.

---

19. The end-to-end operating workflow

Weekly planning

Run once per week.

Step 1: Review last week

Use:

/weekly-review

Output:

• what worked
• what did not
• what deserves another test
• what the audience appears to care about
• what the Researcher should investigate

Step 2: Build the research queue

Ask the Manager:

Using last week's analytics learnings and our content pillars, create a research queue of 10 topics.

For each:
- target audience
- core question
- why it matters
- why now
- evidence needed
- best likely format

Rank by audience usefulness, not novelty alone.

Step 3: Research the top ideas

Run the Researcher.

You do not need 10 finished posts.

You need 3 to 5 well-researched ideas that are worth producing.

Step 4: Produce content

For each approved idea:

/content-cycle

Step 5: Human review

Review:

• the claim
• the framing
• examples
• tone
• screenshots
• visuals
• CTA
• anything legally or reputationally consequential

Step 6: Schedule

Only approved content enters the publishing queue.

---

20. A practical daily workflow

A simple daily rhythm:

Morning: 15 minutes

Ask:

Review the current pipeline.

Show only:
1. items waiting on me
2. items ready to move
3. blockers
4. the single highest-leverage action for today

During the day

Drop new material into:

research/inbox/

Examples:

• links
• transcripts
• screenshots
• notes
• audience questions
• client conversations
• comments
• product updates

Do not interrupt yourself to turn every observation into a post immediately.

Capture first.

Process later.

End of day: 10 minutes

Ask:

Close out today's content operations.

Update statuses.
List what shipped.
List what is waiting.
List what should be first tomorrow.
Do not create new work unless it is required to unblock something.

---

21. Connect external tools with MCP

MCP lets Claude Code connect to external tools and data sources.

Anthropic’s documentation currently includes a Notion example:

claude mcp add --transport http notion https://mcp.notion.com/mcp

For other systems, use the Anthropic connector directory or the official MCP documentation for that service.

Useful categories:

• knowledge base
• analytics
• issue tracking
• CRM
• cloud storage
• databases
• messaging
• design systems

Do not connect everything

Every connection expands what the agent can access and potentially act on.

Use the principle:

Give each role the minimum access required for its job.

Examples:

• Researcher: web + public sources + research folder
• Script Writer: brand files + research + drafts
• Designer: approved copy + brand assets
• Analyst: published content + analytics data
• Publisher: approved folder + scheduler
• Manager: pipeline state, but not necessarily every secret

---

22. Human approval gates

The system should be fast without becoming reckless.

Use mandatory approval before:

• publishing
• sending email or DMs
• modifying paid campaigns
• purchasing anything
• deleting records
• changing production systems
• making public claims on behalf of a company
• handling sensitive or confidential information
• legal, medical or financial claims that require professional review

For a content system, the most important checkpoint is:

DRAFT → REVIEW → APPROVED → PUBLISH

Do not collapse those into one invisible step.

---

23. Research safety and prompt injection

Research agents read untrusted external content.

That content can contain malicious or manipulative instructions.

Treat web pages, documents and external data as data, not as authority over the agent.

Practical rules:

• do not let the Researcher publish
• do not give the Researcher sensitive write access unless necessary
• restrict tools when possible
• do not expose secrets in research files
• verify unexpected tool requests
• use trusted MCP servers
• keep consequential actions behind approval

The fact that an agent can connect to a system does not mean it should.

---

24. Optional deterministic Hooks

Hooks are useful when an action should happen every time, rather than when the model “remembers.”

Good content-system uses:

• validate Markdown formatting after a file is written
• check required frontmatter fields
• run a link checker
• notify you when approval is needed
• prevent files marked APPROVED: NO from moving into a publish directory
• log completed tasks

Use hooks for deterministic rules.

Use agents for judgment.

---

25. Quality control checklist

Before you approve a post, score it from 1 to 5.

Do not publish weak content simply because the workflow produced it.

Automation should raise your throughput after you establish a quality bar.

---

26. Common failure modes

Failure 1: The Researcher returns generic trends

Cause:

The prompt asks for “viral ideas.”

Fix:

Ask for evidence, audience tension, existing content gaps and source-backed angles.

---

Failure 2: Every hook sounds clickbait-y

Cause:

The Hook Writer is optimizing only for curiosity.

Fix:

Require the hook to map to the post’s real payoff.

---

Failure 3: The Script Writer invents details

Cause:

Research is not treated as the source of truth.

Fix:

Require source preservation and a “claims to avoid” section.

---

Failure 4: The Designer produces pretty but repetitive slides

Cause:

No communication objective per slide.

Fix:

Ask what the viewer should understand, not what decoration to add.

---

Failure 5: The Publisher becomes dangerous

Cause:

Scheduling and external actions are fully automatic.

Fix:

Separate preparation from execution and require explicit approval.

---

Failure 6: Analytics becomes a dashboard screenshot

Cause:

Metrics are reported without decisions.

Fix:

Every review must produce experiments and new research questions.

---

Failure 7: The system becomes harder than just posting manually

Cause:

Too many tools and status fields too early.

Fix:

Start with files and 3 to 4 roles. Add integrations only where you see repeatable friction.

---

27. The simplest version you can build today

If this full setup feels too large, use this four-stage version.

1. Research
2. Draft
3. Approve + Publish
4. Analyze

Create only:

.claude/agents/researcher.md
.claude/agents/script-writer.md
.claude/agents/analyst.md
.claude/skills/content-cycle/SKILL.md

The main session handles everything else.

Once you are consistently publishing and reviewing results, add more specialization.

---

28. When to split roles

Split a role when one of these becomes true:

• its context is large enough to distract the main session
• it requires different tool permissions
• it has a repeatable quality checklist
• you perform it frequently
• you want independent iteration on its prompt
• the output has a clear contract for the next stage

Do not create agents because “multi-agent” sounds advanced.

Create agents when specialization reduces cognitive or operational friction.

---

29. A stronger version for teams

If multiple humans are involved, add:

owners/
  creator.md
  editor.md
  designer.md
  reviewer.md

clients/
  client-a/
  client-b/

campaigns/
  campaign-name/

Also add:

• explicit content owners
• approval SLA
• escalation rules
• client-specific brand files
• version control
• audit trail for published content
• separate credentials from content files

For client work, never let one client’s context leak into another client’s content.

---

30. Metrics that actually help a content system learn

Do not obsess over one universal number.

Choose metrics based on the objective.

Awareness content

Useful signals:

• reach
• views
• impressions
• share rate
• non-follower reach

Educational content

Useful signals:

• saves
• completion / retention
• shares
• high-quality comments
• follows after viewing

Conversion content

Useful signals:

• profile visits
• link clicks
• DM conversations
• lead magnet requests
• booked calls
• signups
• revenue

Brand authority

Useful signals:

• inbound opportunities
• qualified followers
• collaborations
• relevant mentions
• repeat engagement
• depth of conversations

The Analyst should connect content pattern → audience response → next experiment.

---

31. Example: turning one idea into a full cycle

Suppose the idea is:

“Claude Code Skills are more useful than repeatedly pasting giant prompts.”

Researcher

Finds:

• official Skills documentation
• what Skills are
• where they live
• how they differ from CLAUDE.md
• use cases
• limitations
• exact current behavior

Hook Writer

Produces:

1. “Stop pasting the same Claude Code prompt every day.”
2. “If your CLAUDE.md is turning into a 2,000-line manual, move this part out.”
3. “Claude Code has a better place for repeatable workflows than your chat history.”

Script Writer

Builds:

• problem
• what Skills are
• when to use them
• file structure
• copy-paste example
• when not to use them

Designer

Creates:

• cover
• CLAUDE.md vs Skill comparison
• folder structure
• SKILL.md example
• use cases
• CTA

Human

Verifies the facts and approves.

Publisher

Checks caption, slide order and CTA.

Analyst

After publishing:

• compares saves vs baseline
• compares completion
• analyzes comments
• notes whether technical “how-to” posts outperform broad AI news

Manager

Feeds that learning back into next week’s research queue.

That is a real content system.

---

32. Copy-paste Manager prompt for a new week

You are managing my AI Content Team OS for this week.

Read:
- CLAUDE.md
- brand/
- analytics/learnings.md
- the most recent weekly analytics review
- current research briefs
- current drafts
- approved and scheduled content

Then produce:

1. Pipeline status
2. Bottlenecks
3. What I personally need to review
4. The 5 highest-value research topics for this week
5. The 3 pieces most ready to produce
6. The agent you recommend for each next step
7. Any factual topics that require fresh verification
8. Any duplicated or low-value work we should delete

Rules:
- prioritize useful content over volume
- do not publish
- do not invent work just to keep agents busy
- preserve our content pillars and brand voice
- use last week's analytics as evidence, not as absolute truth

---

33. Copy-paste prompt for turning raw notes into content

I am dropping raw notes into the content system.

Do not draft immediately.

First:
1. extract the core ideas
2. identify which ideas are factual and require verification
3. identify which ideas are opinion or experience
4. map each idea to our audience and content pillars
5. estimate which ideas are strong enough to become standalone content
6. identify missing proof or examples

Then recommend:
- research needed
- best format
- likely hook direction
- next specialist agent

Save useful items to the pipeline.
Discard obvious duplicates.

---

34. Copy-paste content teardown prompt

Analyze this published post as a content operator.

Do not rewrite it yet.

Evaluate:
- audience fit
- hook
- promise
- clarity
- depth
- proof
- pacing
- visual hierarchy
- CTA
- likely save/share value

Then compare the critique against the available performance data.

Separate:
1. what we can observe
2. what we infer
3. what we should test next

Do not claim a cause unless the evidence supports it.

---

35. Recommended build order

Do this in order.

Phase 1: Context

1. Create the folder.
2. Add CLAUDE.md.
3. Add brand files.
4. Add templates.

Phase 2: Roles

5. Create Researcher.
6. Create Script Writer.
7. Create Analyst.
8. Test those three.

Phase 3: Production

9. Add Hook Writer.
10. Add Designer.
11. Add Publisher.
12. Add the content-cycle Skill.

Phase 4: Integrations

13. Connect only the systems you actually need.
14. Restrict permissions.
15. Add approval gates.
16. Add deterministic hooks.

Phase 5: Learning

17. Run the weekly review.
18. Update learnings.
19. Feed the lessons into research.
20. Keep only the workflows that save real time or improve quality.

---

36. The principle to remember

The goal is not to build a fake company full of imaginary employees.

The goal is to create clear specialized contexts for recurring work.

A good agent system has:

• good inputs
• narrow responsibilities
• clear outputs
• limited permissions
• explicit handoffs
• human judgment
• a measurable feedback loop

If those are missing, adding more agents usually adds more noise.

---

37. Official references

These were checked while preparing this guide.

Claude Code overview

https://code.claude.com/docs/en/overview

Claude Code setup

https://code.claude.com/docs/en/setup

Custom subagents

https://code.claude.com/docs/en/sub-agents

Skills

https://code.claude.com/docs/en/skills

Model Context Protocol in Claude Code

https://code.claude.com/docs/en/mcp

Hooks

https://code.claude.com/docs/en/hooks-guide

Settings

https://code.claude.com/docs/en/settings

Security

https://code.claude.com/docs/en/security

---

38. Your next step

Do not start by connecting seven apps.

Create the folder.

Write your brand context.

Set up the Researcher, Script Writer and Analyst.

Run one real post through the full loop.

Then ask:

• Did this save time?
• Did the output get better?
• Did the handoff reduce repeated explanation?
• Did analytics change the next decision?

If yes, automate the next bottleneck.

That is how the system compounds.

---

Life of Arjav

I share practical AI guides, workflows, agents, automation systems and implementation playbooks for people who want to build, not just watch.

Instagram: https://www.instagram.com/lifeofarjav
LinkedIn: https://www.linkedin.com/in/arjav-j
Website: https://arjavjain.org

If this resource helped, save the original post and follow @lifeofarjav for more practical AI systems.