Built for AI-powered founders, operators, and builders
This guide explains how to structure a Claude Code style agent system using a .claude folder architecture for marketing, content, conversion, and execution workflows.
The goal is simple:
- Move from random AI usage
- Build a structured multi-agent system
- Create repeatable workflows that improve over time
Note: This is a custom operating structure inspired by Claude Code conventions. Adjust folder names and files to match your actual setup.
What this system is
Modern AI tools work best when they are used as structured agents with defined roles, inputs, and outputs.
The .claude system turns AI into a modular execution team:
- Research agents
- Strategy agents
- Content agents
- Conversion agents
- Analytics agents
Each agent has one job, clear boundaries, and a specific output.
Core architecture
1. Root system file
CLAUDE.mdThis is the instruction layer for the entire system.
It defines:
- Brand voice
- Business context
- Rules for agent behavior
- Output expectations
- Do’s and don’ts
This file should be loaded first.
2. System folder
.claude/This is the operational folder for agents, workflows, rules, and memory.
It contains:
- Agent instructions
- Workflow files
- Brand and writing rules
- Reusable business context
- Past learnings and campaign data
The 8-agent system
The system is divided into four execution layers.
Layer 1: Research
Agent 1: ICP Finder
Purpose: Identify who to target.
Inputs:
- Market
- Niche
- Offer
Outputs:
- Ideal customer profiles
- Exclusion criteria
- Targeting logic
Agent 2: Market Research
Purpose: Extract real demand signals.
Sources:
- Forums
- Reviews
- Communities
- Competitor feedback
Outputs:
- Pain points
- Unmet needs
- Buying triggers
- Objections
Layer 2: Strategy
Agent 3: Competitor Analysis
Purpose: Find positioning gaps and differentiation angles.
Inputs:
- Competitors
- Your offer
- Market context
Outputs:
- Positioning gaps
- Differentiation angles
- Winning angles
- Messaging opportunities
Agent 4: Content Strategy
Purpose: Turn research into a clear content direction.
Outputs:
- Content pillars
- Weekly posting plan
- Messaging angles
- Hook ideas
- Campaign themes
Layer 3: Creation
Agent 5: Content Writer
Purpose: Generate usable content from strategy.
Outputs:
- Posts
- Hooks
- Scripts
- Carousel drafts
- Short-form content variations
Focus:
- Consistency over perfection
- Speed of iteration
- Clear, reusable output
Agent 6: Ad Copy Agent
Purpose: Create performance-focused copy for testing.
Outputs:
- Ad angles
- Headlines
- Primary text variations
- CTA variations
Frameworks:
- Curiosity
- Problem
- Proof
- Outcome
Layer 4: Convert and Learn
Agent 7: Landing Page Agent
Purpose: Convert attention into action.
Outputs:
- Landing page structure
- Offer explanation
- CTA alignment
- Conversion-focused copy
- Objection handling
Agent 8: Analytics Agent
Purpose: Close the feedback loop.
Outputs:
- What is working
- What is not working
- What to scale
- What to stop
- What to test next
This agent feeds insights back into research, strategy, and content.
System flow
This system is a loop, not a one-time workflow.
Research -> Strategy -> Creation -> Conversion -> Analytics -> ResearchThe loop compounds when outputs are saved, reused, and improved.
Recommended folder structure
.claude/
│
├── agents/
│ ├── icp-finder.md
│ ├── market-research.md
│ ├── competitor-analysis.md
│ ├── content-strategy.md
│ ├── content-writer.md
│ ├── ad-copy.md
│ ├── landing-page.md
│ └── analytics.md
│
├── workflows/
│ ├── weekly-content-plan.md
│ ├── campaign-build.md
│ └── offer-validation.md
│
├── rules/
│ ├── brand-voice.md
│ ├── writing-style.md
│ ├── positioning.md
│ └── constraints.md
│
└── memory/
├── offers.md
├── audience-insights.md
└── past-campaigns.mdKey principles
1. One agent equals one job
Do not mix responsibilities.
Bad:
- One agent for content, strategy, ads, research, and analytics
Good:
- One agent per function
2. Output matters more than explanation
Every agent should return:
- Usable outputs
- Clear decisions
- Next actions
Avoid:
- Theory dumps
- Long explanations
- Generic advice
3. Systems compound
The value is not the agents by themselves.
The value comes from:
- Feedback loops
- Reusable outputs
- Structured iteration
- Better inputs over time
4. Simplicity wins
Most agent systems fail because they are overbuilt too early.
Start with the minimum useful setup:
- ICP Finder
- Content Strategy
- Content Writer
Expand only when a real bottleneck appears.
How to implement this
Step 1: Create the .claude/ folder
Set up the folder that will hold your agents, workflows, rules, and memory files.
Step 2: Add 3 core agents
Start with:
- ICP Finder
- Content Strategy
- Content Writer
Step 3: Run 5 to 10 real workflows
Use the system before adding complexity.
Step 4: Add the analytics loop
Track what performs, what fails, and what should change.
Step 5: Expand into the full 8-agent system
Add the remaining agents only when they solve a clear workflow problem.
Common mistakes
Avoid these:
- Overbuilding agents before using them
- Mixing strategy and execution roles
- Creating agents with vague responsibilities
- Skipping the feedback loop
- Writing long instruction docs that do not produce usable outputs
- Adding complexity before a bottleneck exists
Why this works
Most AI setups fail because they are:
- Unstructured
- Reactive
- Single-threaded
- Hard to repeat
This system works because it creates:
- Clear roles
- Modular workflows
- Reusable outputs
- Compounding feedback loops
Final takeaway
If your AI setup does not have structure, roles, and feedback loops, it is not a system.
It is just prompts.