A modular framework for orchestrating AI coding agents with Claude Code.
Pick the pieces you need. Skip the ones you don't. Build your own workflow.
V11 is a protocol for managing how AI coding agents work on your projects. It answers questions like:
- How do I stop Claude from editing files without a plan?
- How do I run multiple agents on different parts of the codebase without conflicts?
- How do I track what work has been done and what's left?
- How do I let Claude be more autonomous on safe tasks while requiring approval on risky ones?
This starter pack teaches the concepts and gives you modular pieces to adopt individually.
| Level | What You'll Get |
|---|---|
| Beginner | Copy the Starter CLAUDE.md into your project. Done. |
| Intermediate | Read the Guide, pick 2-3 modules that solve your pain points. |
| Advanced | Build a full hook-enforced pipeline with formations and autonomy levels. |
Copy prompts/starter-claude-md.md to your project root as CLAUDE.md. It teaches Claude task discipline, risk awareness, and session flow. No hooks, no scripts, no infrastructure.
- Read
modules/02-hooks/to understand the hook architecture - Copy the example hooks into your project's
.claude/directory - Configure
settings.jsonto wire them up
Work through the Guide start to finish, then adopt modules in order:
Task System (01) -> Hooks (02) -> Sessions (05) -> Autonomy (04) -> Formations (03) -> Agents (06) -> Scaling (07)
Each module is self-contained. Adopt any combination.
| # | Module | What It Does | Difficulty |
|---|---|---|---|
| 01 | Task System | Tasks as source of truth, not markdown checklists | Beginner |
| 02 | Hooks | Shell scripts that enforce rules before/after tool use | Intermediate |
| 03 | Formations | Pre-built team patterns (feature dev, bug hunt, review) | Intermediate |
| 04 | Autonomy | Progressive trust levels (A0-A4) for agent independence | Intermediate |
| 05 | Sessions | Structured project directories with specs and state | Beginner |
| 06 | Agents | Specialized agent definitions (implementer, tester, reviewer) | Advanced |
| 07 | Scaling | How to structure 10, 50, or 500-task projects | Intermediate |
Ready-to-use prompts and CLAUDE.md templates:
| Prompt | Purpose |
|---|---|
| starter-claude-md.md | Drop-in CLAUDE.md for any project |
| task-workflow.md | How to prompt Claude for task-driven work |
| team-lead.md | How to prompt Claude as a team coordinator |
| hook-builder.md | How to prompt Claude to build custom hooks |
| Example | Scale | What's Configured |
|---|---|---|
| tiny-project/ | 1-5 tasks | CLAUDE.md only, flat task list |
| small-project/ | 10-20 tasks | CLAUDE.md + hooks + sessions |
| medium-project/ | 50+ tasks | Full framework: formations, gates, autonomy |
- Tasks are truth. Don't track progress in markdown files. Use
TaskCreate/TaskUpdate/TaskList. - Read before write. Never edit a file you haven't read. Never guess at contents.
- Risk-aware. Low-risk actions proceed. Medium-risk needs context. High-risk needs approval.
- Modular. No module depends on another. Adopt incrementally.
- Simple is better. If a flat task list works, don't add formations. If CLAUDE.md works, don't add hooks.
This is extracted from a production protocol running 1,800+ Claude Code sessions with 83 specialized agents. The core patterns are universal; the infrastructure is optional.
MIT -- see LICENSE.