Claude-Code-Workflow/README.md

🚀 Claude Code Workflow (CCW)

Languages: English | 中文

---

Claude Code Workflow (CCW) is a next-generation multi-agent automation framework that orchestrates complex software development tasks through intelligent workflow management and autonomous execution.

🎉 Latest: v3.5.0 - UI Design Workflow with Triple Vision Analysis. See CHANGELOG.md for details.

What's New in v3.5.0:

• 🎨 UI Design Workflow: Complete design refinement workflow from style extraction to prototype generation
• 👁️ Triple Vision Analysis: Combines Claude Code + Gemini + Codex vision capabilities for comprehensive style extraction
• ⏸️ Interactive Checkpoints: User selection points for style variants and prototype confirmation
• 🎯 Zero Agent Overhead: Direct bash execution for CLI tools, removing unnecessary agent wrappers
• 🎨 Style Customization: Runtime override support with --style-overrides parameter
• 📦 Batch Task Generation: Optional automatic task creation for selected prototypes with --batch-plan
• 🔄 Semi-Autonomous Workflow: User-driven continuation at critical design decision points

---

✨ Key Features

• 🎯 Context-First Architecture: Pre-defined context gathering eliminates execution uncertainty and error accumulation.
• 🤖 Multi-Agent System: Specialized agents (@code-developer, @test-fix-agent) with tech-stack awareness and automated test validation.
• 🔄 End-to-End Workflow Automation: From brainstorming to deployment with multi-phase orchestration.
• 📋 JSON-First Task Model: Structured task definitions with pre_analysis steps for deterministic execution.
• 🧪 TDD Workflow Support: Complete Test-Driven Development with Red-Green-Refactor cycle enforcement.
• 🧠 Multi-Model Orchestration: Leverages Gemini (analysis), Qwen (architecture), and Codex (implementation) strengths.
• ✅ Pre-execution Verification: Validates plans with both strategic (Gemini) and technical (Codex) analysis.
• 🔧 Unified CLI: A single, powerful /cli:* command set for interacting with various AI tools.
• 📦 Smart Context Package: context-package.json links tasks to relevant codebase files and external examples.

---

⚙️ Installation

🚀 Quick One-Line Installation

Windows (PowerShell):

Invoke-Expression (Invoke-WebRequest -Uri "https://raw.githubusercontent.com/catlog22/Claude-Code-Workflow/main/install-remote.ps1" -UseBasicParsing).Content

Linux/macOS (Bash/Zsh):

bash <(curl -fsSL https://raw.githubusercontent.com/catlog22/Claude-Code-Workflow/main/install-remote.sh)

📋 Interactive Version Selection

After running the installation command, you'll see an interactive menu with real-time version information:

Detecting latest release and commits...
Latest stable: v3.2.0 (2025-10-02 04:27 UTC)
Latest commit: cdea58f (2025-10-02 08:15 UTC)

====================================================
            Version Selection Menu
====================================================

1) Latest Stable Release (Recommended)
   |-- Version: v3.2.0
   |-- Released: 2025-10-02 04:27 UTC
   \-- Production-ready

2) Latest Development Version
   |-- Branch: main
   |-- Commit: cdea58f
   |-- Updated: 2025-10-02 08:15 UTC
   |-- Cutting-edge features
   \-- May contain experimental changes

3) Specific Release Version
   |-- Install a specific tagged release
   \-- Recent: v3.2.0, v3.1.0, v3.0.1

====================================================

Select version to install (1-3, default: 1):

Version Options:

• Option 1 (Recommended): Latest stable release with verified production quality
• Option 2: Latest development version from main branch with newest features
• Option 3: Specific version tag for controlled deployments

💡 Pro Tip: The installer automatically detects and displays the latest version numbers and release dates from GitHub. Just press Enter to select the recommended stable release.

📦 Local Installation (Install-Claude.ps1)

For local installation without network access, use the bundled PowerShell installer:

Installation Modes:

# Interactive mode with prompts (recommended)
.\Install-Claude.ps1

# Quick install with automatic backup
.\Install-Claude.ps1 -Force -BackupAll

# Non-interactive install
.\Install-Claude.ps1 -NonInteractive -Force

Installation Options:

Mode	Description	Installs To
Global	System-wide installation (default)	~/.claude/, ~/.codex/, ~/.gemini/
Path	Custom directory + global hybrid	Local: agents/, commands/ Global: workflows/, scripts/

Backup Behavior:

• Default: Automatic backup enabled (-BackupAll)
• Disable: Use -NoBackup flag (⚠️ overwrites without backup)
• Backup location: claude-backup-{timestamp}/ in installation directory

⚠️ Important Warnings:

• -Force -BackupAll: Silent file overwrite (with backup)
• -NoBackup -Force: Permanent file overwrite (no recovery)
• Global mode modifies user profile directories

✅ Verify Installation

After installation, run the following command to ensure CCW is working:

/workflow:session:list

📝 Installation Notes:

• The installer will automatically install/update .codex/ and .gemini/ directories
• Global mode: Installs to ~/.codex and ~/.gemini
• Path mode: Installs to your specified directory (e.g., project/.codex, project/.gemini)
• Backup: Existing files are backed up by default to claude-backup-{timestamp}/
• Safety: Use interactive mode for first-time installation to review changes

---

⚙️ Configuration

Prerequisites: Required Tools

Before using CCW, install the following command-line tools:

Core CLI Tools

Tool	Purpose	Installation
Gemini CLI	AI analysis & documentation	npm install -g @google/gemini-cli (GitHub)
Codex CLI	AI development & implementation	npm install -g @openai/codex (GitHub)
Qwen Code	AI architecture & code generation	npm install -g @qwen-code/qwen-code (Docs)

System Utilities

Tool	Purpose	Installation
ripgrep (rg)	Fast code search	Download or brew install ripgrep (macOS), apt install ripgrep (Ubuntu)
jq	JSON processing	Download or brew install jq (macOS), apt install jq (Ubuntu)

Quick Install (All Tools):

# macOS
brew install ripgrep jq
npm install -g @google/gemini-cli @openai/codex @qwen-code/qwen-code

# Ubuntu/Debian
sudo apt install ripgrep jq
npm install -g @google/gemini-cli @openai/codex @qwen-code/qwen-code

# Windows (Chocolatey)
choco install ripgrep jq
npm install -g @google/gemini-cli @openai/codex @qwen-code/qwen-code

Essential: Gemini CLI Setup

Configure Gemini CLI for optimal integration:

// ~/.gemini/settings.json
{
  "contextFileName": ["CLAUDE.md", "GEMINI.md"]
}

Recommended: .geminiignore

Optimize performance by excluding unnecessary files:

# .geminiignore (in project root)
/dist/
/build/
/node_modules/
/.next/
*.tmp
*.log
/temp/

# Include important docs
!README.md
!**/CLAUDE.md

Recommended: MCP Tools (Enhanced Analysis)

MCP (Model Context Protocol) tools provide advanced codebase analysis. Recommended installation - While CCW has fallback mechanisms, not installing MCP tools may lead to unexpected behavior or degraded performance in some workflows.

Available MCP Servers

MCP Server	Purpose	Installation Guide
Exa MCP	External API patterns & best practices	Install Guide
Code Index MCP	Advanced internal code search	Install Guide

Benefits When Enabled

• 📊 Faster Analysis: Direct codebase indexing vs manual searching
• 🌐 External Context: Real-world API patterns and examples
• 🔍 Advanced Search: Pattern matching and similarity detection
• ⚡ Better Reliability: Primary tools for certain workflows

⚠️ Note: Some workflows expect MCP tools to be available. Without them, you may experience:

• Slower code analysis and search operations
• Reduced context quality in some scenarios
• Fallback to less efficient traditional tools
• Potential unexpected behavior in advanced workflows

---

🚀 Getting Started

Complete Development Workflow

Phase 1: Brainstorming & Conceptual Planning

# Multi-perspective brainstorming with role-based agents
/workflow:brainstorm:auto-parallel "Build a user authentication system"

# Review and refine specific aspects (optional)
/workflow:brainstorm:ui-designer "authentication flows"
/workflow:brainstorm:synthesis  # Generate consolidated specification

Phase 2: UI Design Refinement (Optional for UI-heavy projects)

# Extract design styles from reference images and generate prototypes
/workflow:design:auto --session WFS-auth --images "design-refs/*.png" --pages "login,register" --batch-plan

# Or run individual design phases
/workflow:design:style-extract --session WFS-auth --images "refs/*.png"
/workflow:design:style-consolidate --session WFS-auth --variants "variant-1,variant-3"
/workflow:design:ui-generate --session WFS-auth --pages "login,register" --variants 2
/workflow:design:design-update --session WFS-auth --selected-prototypes "login-variant-1,register-variant-2"

Phase 3: Action Planning

# Create executable implementation plan
/workflow:plan "Implement JWT-based authentication system"

# OR for TDD approach
/workflow:tdd-plan "Implement authentication with test-first development"

Phase 4: Execution

# Execute tasks with AI agents
/workflow:execute

# Monitor progress
/workflow:status

Phase 5: Testing & Quality Assurance

# Generate independent test-fix workflow (v3.2.2+)
/workflow:test-gen WFS-auth  # Creates WFS-test-auth session
/workflow:execute            # Runs test validation

# OR verify TDD compliance (TDD workflow)
/workflow:tdd-verify

Quick Start for Simple Tasks

Feature Development:

/workflow:session:start "Add password reset feature"
/workflow:plan "Email-based password reset with token expiry"
/workflow:execute

Bug Fixing:

# Interactive analysis with CLI tools
/cli:mode:bug-index --tool gemini "Login timeout on mobile devices"

# Execute the suggested fix
/workflow:execute

💡 When to Use Which Approach?

Use /workflow:plan + /workflow:execute for:

• Complex features requiring multiple modules (>3 modules)
• Tasks with multiple subtasks (>5 subtasks)
• Cross-cutting changes affecting architecture
• Features requiring coordination between components
• When you need structured planning and progress tracking

Use Claude Code directly for:

• Simple, focused changes (single file or module)
• Quick bug fixes with clear solutions
• Documentation updates
• Code refactoring within one component
• Straightforward feature additions

Code Analysis:

# Deep codebase analysis
/cli:mode:code-analysis --tool qwen "Analyze authentication module architecture"

---

🛠️ Command Reference

Unified CLI Commands (`/cli:*)

Use the --tool <gemini|qwen|codex> flag to select the desired tool. Defaults to gemini.

Command	Description
/cli:analyze	Deep codebase analysis.
/cli:chat	Direct, interactive chat with a tool.
/cli:execute	Execute a task with full permissions.
/cli:cli-init	Initialize CLI tool configurations for the workspace.
/cli:mode:bug-index	Analyze bugs and suggest fixes.
/cli:mode:code-analysis	Perform deep code analysis and debugging.
/cli:mode:plan	Project planning and architecture analysis.

Workflow Commands (`/workflow:*)

Command	Description
/workflow:session:*	Manage development sessions (start, pause, resume, list, switch, complete).
/workflow:brainstorm:*	Use role-based agents for multi-perspective planning.
/workflow:design:auto	NEW Semi-autonomous UI design workflow with interactive checkpoints for style and prototype selection.
/workflow:design:style-extract	NEW Extract design styles from reference images using triple vision analysis (Claude + Gemini + Codex).
/workflow:design:style-consolidate	NEW Consolidate selected style variants into validated design tokens and style guide.
/workflow:design:ui-generate	NEW Generate token-driven HTML/CSS prototypes with optional style overrides.
/workflow:design:design-update	NEW Integrate finalized design system into brainstorming artifacts.
/workflow:plan	Create a detailed, executable plan from a description.
/workflow:tdd-plan	Create TDD workflow (6 phases) with test coverage analysis and Red-Green-Refactor cycles.
/workflow:execute	Execute the current workflow plan autonomously.
/workflow:status	Display the current status of the workflow.
/workflow:test-gen [--use-codex] <session>	Create test generation workflow with auto-diagnosis and fix cycle for completed implementations.
/workflow:tdd-verify	Verify TDD compliance and generate quality report.
/workflow:review	Optional manual review (only use when explicitly needed - passing tests = approved code).
/workflow:tools:test-context-gather	Analyze test coverage and identify missing test files.
/workflow:tools:test-concept-enhanced	Generate test strategy and requirements analysis using Gemini.
/workflow:tools:test-task-generate	Generate test task JSON with test-fix-cycle specification.

Task & Memory Commands

Command	Description
/task:*	Manage individual tasks (create, breakdown, execute, replan).
/update-memory-full	Re-index the entire project documentation.
/update-memory-related	Update documentation related to recent changes.
/version	Display version information and check for updates from GitHub.

---

🧩 How It Works: Design Philosophy

The Core Problem

Traditional AI coding workflows face a fundamental challenge: execution uncertainty leads to error accumulation.

Example:

# Prompt 1: "Develop XX feature"
# Prompt 2: "Review XX architecture in file Y, then develop XX feature"

While Prompt 1 might succeed for simple tasks, in complex workflows:

• The AI may examine different files each time
• Small deviations compound across multiple steps
• Final output drifts from the intended goal

CCW's Mission: Solve the "1-to-N" problem — building upon existing codebases with precision, not just "0-to-1" greenfield development.

---

The CCW Solution: Context-First Architecture

1. Pre-defined Context Gathering

Instead of letting agents randomly explore, CCW uses structured context packages:

context-package.json created during planning:

{
  "metadata": {
    "task_description": "...",
    "tech_stack": {"frontend": [...], "backend": [...]},
    "complexity": "high"
  },
  "assets": [
    {
      "path": "synthesis-specification.md",
      "priority": "critical",
      "sections": ["Backend Module Structure"]
    }
  ],
  "implementation_guidance": {
    "start_with": ["Step 1", "Step 2"],
    "critical_security_items": [...]
  }
}

2. JSON-First Task Model

Each task includes a flow_control.pre_analysis section:

{
  "id": "IMPL-1",
  "flow_control": {
    "pre_analysis": [
      {
        "step": "load_architecture",
        "commands": ["Read(architecture.md)", "grep 'auth' src/"],
        "output_to": "arch_context",
        "on_error": "fail"
      }
    ],
    "implementation_approach": {
      "modification_points": ["..."],
      "logic_flow": ["..."]
    },
    "target_files": ["src/auth/index.ts"]
  }
}

Key Innovation: The pre_analysis steps are executed before implementation, ensuring agents always have the correct context.

3. Multi-Phase Orchestration

CCW workflows are orchestrators that coordinate slash commands:

Planning Phase (/workflow:plan):

Phase 1: session:start       → Create session
Phase 2: context-gather      → Build context-package.json
Phase 3: concept-enhanced    → CLI analysis (Gemini/Qwen)
Phase 4: task-generate       → Generate task JSONs with pre_analysis

Execution Phase (/workflow:execute):

For each task:
  1. Execute pre_analysis steps → Load context
  2. Apply implementation_approach → Make changes
  3. Validate acceptance criteria → Verify success
  4. Generate summary → Track progress

4. Multi-Model Orchestration

Each AI model serves its strength:

Model	Role	Use Cases
Gemini	Analysis & Understanding	Long-context analysis, architecture review, bug investigation
Qwen	Architecture & Design	System design, code generation, architectural planning
Codex	Implementation	Feature development, testing, autonomous execution

Example:

# Gemini analyzes the problem space
/cli:mode:code-analysis --tool gemini "Analyze auth module"

# Qwen designs the solution
/cli:analyze --tool qwen "Design scalable auth architecture"

# Codex implements the code
/workflow:execute  # Uses @code-developer with Codex

---

From 0-to-1 vs 1-to-N Development

Scenario	Traditional Workflow	CCW Approach
Greenfield (0→1)	✅ Works well	✅ Adds structured planning
Feature Addition (1→2)	⚠️ Context uncertainty	✅ Context-package links to existing code
Bug Fixing (N→N+1)	⚠️ May miss related code	✅ Pre-analysis finds dependencies
Refactoring	⚠️ Unpredictable scope	✅ CLI analysis + structured tasks

---

Key Workflows

Complete Development (Brainstorm → Deploy)

Brainstorm (8 roles) → Synthesis → Plan (4 phases) → Execute → Test → Review

Quick Feature Development

session:start → plan → execute → test-gen → execute

TDD Workflow

tdd-plan (TEST→IMPL→REFACTOR chains) → execute → tdd-verify

Bug Fixing

cli:mode:bug-index (analyze) → execute (fix) → test-gen (verify)

---

🤝 Contributing & Support

• Repository: GitHub - Claude-Code-Workflow
• Issues: Report bugs or request features on GitHub Issues.
• Discussions: Join the Community Forum.

📄 License

This project is licensed under the MIT License. See the LICENSE file for details.