Clone this repo. Point Claude at it. Tell it to build you a plan.
That's the whole idea. This is a field-tested collection of patterns, architecture decisions, and operational techniques for turning Claude Code into something that compounds — where every session makes the next one smarter. 158 techniques, all production-tested, organized so both humans and AI agents can consume them.
I built the system this grew out of to run my company. Meeting prep that pulls calendar, email, CRM, and LinkedIn data in 30 seconds. Skills that ghostwrite in my voice. A knowledge base that gets richer every session without me maintaining it. Hooks that enforce discipline I'd otherwise forget. The patterns here are the genericized versions — stripped of anything company-specific, structured so you can replicate them.
For AI agents: This repo is designed to be consumed programmatically. Each part is self-contained with clear headings, structured examples, and implementation details. Point your agent at any part file and it has enough context to build an implementation plan.
git clone https://github.com/AaronRoeF/claude-code-patterns.git
cd claude-code-patterns
Then open Claude Code and say:
"Read PART3-BUILD-A-KNOWLEDGE-BASE.md and build me a plan for setting this up with my stack."
Claude reads the patterns, asks what you're working with, and produces a step-by-step implementation plan. That's it.
Already have a setup? Point Claude at PART2-TECHNIQUES.md for 129 specific tips to sharpen what you've got. Or PART1-CORE-ARCHITECTURE.md to compare your architecture against 11 foundational patterns.
The system that produced these patterns runs three layers. You don't need all three to start — but understanding the shape helps you see where the patterns fit.
┌─────────────────────────────────────────────────────────┐
│ CLAUDE.md (Schema) │
│ Identity · Routing · Rules · Guardrails │
│ │
│ "prep Sarah" → loads meeting-prep skill │
│ "scan email" → loads email-triage skill │
│ "draft a post" → loads voice skill │
└──────────────────────────┬──────────────────────────────┘
│ dispatches to
┌────────────────┼────────────────┐
▼ ▼ ▼
┌─────────────────┐ ┌──────────────┐ ┌──────────────────┐
│ WORK SKILLS │ │ PERSONAL │ │ KNOWLEDGE BASE │
│ (Team repo) │ │ SKILLS │ │ (Obsidian vault) │
│ │ │ (Private) │ │ │
│ Meeting prep │ │ Health data │ │ people/ │
│ PM workflows │ │ Task mgmt │ │ accounts/ │
│ CRM analytics │ │ Msg analysis│ │ projects/ │
│ Content review │ │ Learning │ │ observations/ │
│ Email triage │ │ Vault ops │ │ decisions/ │
└────────┬────────┘ └──────┬──────┘ └────────┬─────────┘
│ │ │
└────────┬────────┘ │
▼ │
┌─────────────────────────────────────────────┤
│ MCP SERVERS (Global) │
│ │
│ Gmail · Calendar · Slack · Drive │
│ Jira · Notion · Playwright · CRM │
│ + custom servers for local data │
└─────────────────────────────────────────────┤
│ │
▼ │
┌─────────────────────────────────────────────┘
│ HOOKS (Enforcement)
│
│ SessionStart: project dashboard, context reload
│ PreToolUse: focus gate, safety checks
│ PostToolUse: audit log, test sync, learning capture
│ Notification: desktop/mobile alerts
│ Permission: phone approval via push notification
└─────────────────────────────────────────────────────┘
The flywheel: Skills pull live data from MCP servers → process and produce output → write enriched data to the knowledge base → next session starts with richer context. Meeting prep reads contact files that were updated by the last debrief. Every session compounds.
Five architectural decisions that make it work:
-
Separation of concerns. Work skills in one repo (distributable to teammates), personal skills in another (private), knowledge base separate from both. Each has its own focused CLAUDE.md.
-
Progressive disclosure. CLAUDE.md is the router, not the encyclopedia. Heavy reference data (personas, brand guides, API docs) loads only when a skill fires. Keeps baseline token cost low.
-
Hooks for enforcement, CLAUDE.md for judgment. CLAUDE.md rules degrade after
/compact. Hooks fire mechanically regardless. Use hooks for anything that must never be skipped. -
Broad permissions, narrow gates. Claude can read all email, access Slack, browse the web. Guardrails are in the skills, not the permissions. Draft-then-confirm at the point of consequence.
-
The knowledge base compounds. Every skill that touches external data writes enriched data back to the knowledge base. The 10th meeting prep for the same contact is dramatically better than the first.
| Part | What You Get | Count |
|---|---|---|
| Part 1: Core Architecture | Foundational patterns from a live production setup — the decisions that create leverage | 11 |
| Part 2: Techniques | Field-tested tips in 16 categories, each rated Beginner / Intermediate / Advanced | 134 |
| Part 3: Build a Knowledge Base | Step-by-step guide to building a persistent, compounding AI knowledge base (the Karpathy LLM Wiki pattern, production-grade) | 13 |
| Part 4: Quick Reference | Cheat sheets — keyboard shortcuts, slash commands, MCP starter kit, 5-minute setup | — |
| Part 5: Live Examples | Hooks, test suites, and scripts actually running in production right now | — |
Total: 158 field-tested techniques. Not aspirational — deployed.
Power users who've been using Claude Code for weeks and want to go from productive to compounding. You're past "how do I set up CLAUDE.md" and into "how do I build a system that gets smarter every session."
Business operators who use Claude Code for more than coding — meeting prep, email triage, document generation, CRM analytics. The patterns here treat Claude as an operating system, not a code assistant.
Teams evaluating how to standardize Claude Code across an org. The architecture section shows how one skill set serves multiple users with identity-aware execution.
AI agents consuming this repo as reference material. Every part file is self-contained with enough context to generate implementation plans.
For full attribution, see SOURCES.md.
- Andrej Karpathy — LLM Wiki — the pattern that inspired Part 3
- Adventures in Claude — 35 posts on Claude Code workflows and operational patterns
- Claude Code Official Docs
- Builder.io — How to Write a Good CLAUDE.md
- HumanLayer — Writing a Good CLAUDE.md
- Arize — CLAUDE.md Best Practices
- SFEIR Institute — Claude Code Courses
- Addy Osmani — Claude Code Agent Teams
- MCPcat — Managing Context
- ykdojo/claude-code-tips
- DataCamp — Claude Code Hooks Tutorial
- Claude Cowork
- MCP Specification — Tools
This pattern library is a living document. If you've discovered a technique that makes Claude Code significantly more effective, we want it.
How to contribute:
- Fork this repo and create a branch
- Add your pattern in the appropriate category, following the existing format:
- Title as
### Pattern Name - 2-4 sentence explanation of what it does and why it matters
**Level:**rating (Beginner / Intermediate / Advanced)**Source:**link if applicable
- Title as
- Submit a PR with a brief description of the pattern and how you've used it
Guidelines:
- Patterns should be tested in real workflows, not theoretical
- Include the difficulty level — this helps readers prioritize
- Link to sources where possible so readers can go deeper
- Keep explanations concise: what it does, why it matters, how to set it up
- Genericize any personal or company-specific details
What we're looking for:
- Novel hook configurations
- Creative skill architectures
- MCP server patterns and gotchas
- Cost optimization strategies
- Multi-agent orchestration patterns
- Security and permission patterns