Dual AI Collaboration Framework Builder — Establish a clear division of labor between Claude + Codex for new projects
🎯 Core Concept
Establish a clear AI collaboration system:
- Claude (Designer): Requirements analysis, technical solution design, architecture decisions, code review
- Codex (Implementer): Coding according to established plans, writing tests, running verification scripts
- Human (Decision Maker): Final judgment, requirement confirmation, direction control
Workflow: Human → Claude (Plan) → Codex (Implement) → Claude (Review) → Human (Decision)
📦 What This Skill Does
Build a complete dual AI collaboration framework from scratch, including:
- PRD Co-creation: Conversational guidance to complete Product Requirements Document
- Collaboration Standards Generation: Auto-generate
CLAUDE.mdandAGENTS.mdbehavior specifications - Documentation System: Create
.ai/brief.md,plan.md,review.md,decision-log.md - Verification Script Framework: Generate executable frameworks for
scripts/check.sh,build.sh,test.sh - Handoff Mechanism: Built-in quick-switch phrases to solve collaboration breakpoints
🚀 Usage
1. Install Skill
Choose your language:
- English: Save
SKILL.mdto your Claude skills library - 中文: Save
SKILL.zh-CN.mdto your Claude skills library
(Specific path depends on your Claude configuration)
Note: The skill will generate documents in the language you use during conversation. You can mix languages freely (e.g., English conversation → English PRD, Chinese conversation → Chinese PRD).
2. Invoke Skill
In a Claude conversation, type:
/DualForge-Workflow
Or simply say:
I want to set up a dual AI collaboration framework
3. Follow the Process
The skill will guide you through 4 stages:
- Stage 0: Project information collection (project name, scale, one-line description)
- Stage 1: PRD co-creation (depth adjusted based on project scale)
- Stage 2: Collaboration mechanism configuration (tech stack, test requirements, task file mechanism)
- Stage 3: Generate complete framework (13 files)
📂 Generated File Structure
project/
├── PRD.md # Product Requirements Document
├── CLAUDE.md # Claude behavior specification
├── AGENTS.md # Codex behavior specification
├── GETTING_STARTED.md # Getting started guide
├── scripts/
│ ├── check.sh # Verification script framework
│ ├── build.sh # Build script framework
│ └── test.sh # Test script framework
├── src/ # Source code root directory
└── .ai/
├── brief.md # Entry point + collaboration contract
├── plan.md # Project plan
├── review.md # Review checklist
├── decision-log.md # Architecture Decision Records
└── tasks/
└── _TEMPLATE.md # Task template (optional)
📋 Live Example
See a complete example of generated framework in examples/small-project-todoapp/
TodoApp - Small Project Example
A task management tool for teams of 10-20 people. Generated in ~15 minutes of conversation.
File Structure:
todoapp/
├── PRD.md # 7 functional requirements with acceptance criteria
├── CLAUDE.md # Claude work specifications
├── AGENTS.md # Codex work specifications
├── GETTING_STARTED.md # Step-by-step guide
├── scripts/
│ ├── check.sh # ✓ Format → Lint → Types → Tests
│ ├── build.sh # ✓ Build with Vite
│ └── test.sh # ✓ Run Vitest + coverage report
├── src/
│ └── README.md # Architecture notes
└─�� .ai/
├── brief.md # Project context + quick-switch phrases
├── plan.md # 13 tasks across 4 stages
├── review.md # Review checklist
└── decision-log.md # 1 ADR (tech stack selection)
Key Highlights:
- Complete PRD: 7 FRs (task management, assignment, status flow, members, team view, filter/search, comments)
- Design Principles: 5 principles for conflict resolution (transparency first, simple over feature-rich, local first, equal collaboration, clear process)
- Out of Scope: 10 items explicitly excluded from MVP (file attachments, notifications, task dependencies, etc.)
- Project Plan: 13 tasks organized into 4 stages with clear dependencies
- Ready Scripts: All three scripts have executable commands for TypeScript + React + Vite stack
Sample from PRD.md:
### FR-3 Status Management
**Description**: Tasks have four statuses and transition in a fixed order.
**Status Definitions:**
1. **To Do**: Task created, not yet started
2. **In Progress**: Task is being executed
3. **In Review**: Task completed, awaiting team confirmation
4. **Done**: Task approved and completed
**Transition Rules:**
- Status transitions in order: To Do → In Progress → In Review → Done
- Cannot skip statuses
- Can roll back to previous status
- All team members have permission to modify task status
**Acceptance Criteria:**
- [ ] New tasks default to "To Do" status
- [ ] Can only advance in order
- [ ] Can roll back to previous status
- [ ] Status changes immediately reflected in interface
Sample from scripts/check.sh:
#!/bin/bash
# Stage 4: Lint
echo "🔍 Running ESLint..."
npm run lint
if [ $? -ne 0 ]; then
echo "❌ Lint failed"
exit 1
fi
# Stage 5: Type Check
echo "📘 Running TypeScript type check..."
npm run type-check
if [ $? -ne 0 ]; then
echo "❌ Type check failed"
exit 1
fi
→ Explore the complete example
✨ Key Features
1. Smart Project Scale Adaptation
- Small projects (< 10 features): Lightweight PRD, 5-10 rounds of conversation
- Medium projects (10-30 features): Standard PRD, 10-15 rounds of conversation
- Large projects (> 30 features): In-depth PRD, 20+ rounds of conversation
2. Optional Task File Mechanism
- Complex projects: Write task files first, then implement
- Simple projects: Give plans directly in conversation for rapid iteration
3. Solves Handoff Breakpoints
- Handoff instructions: CLAUDE.md requires explicit handoff instructions after producing plans, ensuring smooth transition to Codex
- Quick-switch phrases: brief.md provides phrase templates for seamless AI switching during conversation
4. Complete Verification System
check.sh: 7-stage checks (structure → boundaries → format → lint → types → tests)build.sh: Build script frameworktest.sh: Test script framework
🎨 Use Cases
✅ Suitable for:
- Projects with clear requirements needing long-term iteration
- Team collaboration requiring standardized processes
- High complexity projects requiring separation of design and implementation
❌ Not suitable for:
- One-off scripts, experimental projects
- Pure exploration with unclear requirements
- Quick prototypes, small personal projects without collaboration needs
📖 Workflow Example
You: I want to build a task management system
↓
Claude invokes DualForge-Workflow skill
↓ Guides through PRD completion
Claude generates complete collaboration framework (13 files)
↓
You: Implement user login feature
↓
Claude produces task file .ai/tasks/1.1-user-login.md
+ Outputs handoff instructions
↓ Copy handoff instructions to Codex
Codex implements according to plan + tests + check.sh
+ Updates status
+ Proactively requests Review
↓ Copy Review request to Claude
Claude reviews and gives conclusion (can merge / needs changes / wrong direction)
↓
You: Decide whether to merge
🔧 Configuration Options
Project Scale
- Small (personal project/quick MVP)
- Medium (team project/standard product)
- Large (complex system/platform-level product)
Tech Stack
Supports any tech stack, skill generates corresponding script examples based on your input:
- TypeScript / JavaScript
- Python
- Go
- Rust
- Java
- And more...
Test Coverage
- Standard (overall ≥60%, core ≥80%)
- Strict (overall ≥80%)
- Relaxed (overall ≥40%, rapid iteration priority)
📚 Documentation
SKILL.md- Complete skill definition and implementation detailsGETTING_STARTED.md- Getting started guide (automatically created by skill)
🤝 Contributing
Issues and Pull Requests are welcome!
📄 License
MIT
Make AI collaboration clearer, make development more efficient.
Comments