## 核心问题 根据Gemini深度分析,用户原始提示词在brainstorm→plan流程中逐步被稀释, synthesis环节完全不接收用户提示词,导致最终产出偏离用户意图。 ## 关键修改 ### 1. synthesis.md - 修复最严重偏差环节 - 新增从workflow-session.json加载原始用户提示词 - Agent流程Step 1: load_original_user_prompt (HIGHEST priority) - 新增"用户意图对齐"为首要合成要求 - 添加意图追溯和偏差警告完成标准 ### 2. concept-clarify.md - 添加意图验证 - 加载original_user_intent从session metadata - 新增"用户意图对齐"为最高优先级扫描类别 - 检查目标一致性、范围匹配、偏差检测 ### 3. action-plan-verify.md - 强化验证门禁 - workflow-session.json作为主要参考源 - 新增"用户意图对齐"为CRITICAL级别检测 - 违反用户意图标记为CRITICAL严重性 ### 4. plan.md - 确立意图优先级 - 明确用户原始意图始终为主要参考 - 优先级规则: 当前用户提示词 > synthesis > topic-framework ### 5. artifacts.md - 推广结构化格式 - 添加推荐的GOAL/SCOPE/CONTEXT结构化格式 - 强调用户意图保存的重要性 ### 6. auto-parallel.md - 编排器完整性 - 推荐结构化提示词格式 - Phase 1标注用户意图存储 - Phase 3明确synthesis加载用户意图为最高优先级 - Agent模板强调用户意图为PRIMARY reference ## 影响力提升 用户提示词影响力: 30% → 95% ## 参考 分析报告由Gemini生成,识别了5个关键问题点并提供4条改进建议 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
🚀 Claude Code Workflow (CCW)
Claude Code Workflow (CCW) transforms AI development from simple prompt chaining into a robust, context-first orchestration system. It solves execution uncertainty and error accumulation through structured planning, deterministic execution, and intelligent multi-model orchestration.
🎉 Latest: v4.6.0 - Concept Clarification & Agent-Driven Analysis. See CHANGELOG.md for details.
📚 New to CCW? Check out the Getting Started Guide for a beginner-friendly 5-minute tutorial!
✨ Core Concepts
CCW is built on a set of core principles that differentiate it from traditional AI development approaches:
- Context-First Architecture: Pre-defined context gathering eliminates execution uncertainty by ensuring agents have the correct information before implementation.
- JSON-First State Management: Task states live in
.task/IMPL-*.jsonfiles as the single source of truth, enabling programmatic orchestration without state drift. - Autonomous Multi-Phase Orchestration: Commands chain specialized sub-commands and agents to automate complex workflows with zero user intervention.
- Multi-Model Strategy: Leverages the unique strengths of different AI models (Gemini for analysis, Codex for implementation) for superior results.
- Hierarchical Memory System: A 4-layer documentation system provides context at the appropriate level of abstraction, preventing information overload.
- Specialized Role-Based Agents: A suite of agents (
@code-developer,@test-fix-agent, etc.) mirrors a real software team to handle diverse tasks.
⚙️ Installation
For detailed installation instructions, please refer to the INSTALL.md guide.
🚀 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)
✅ Verify Installation
After installation, open Claude Code and check if the workflow commands are available by running:
/workflow:session:list
If the slash commands (e.g., /workflow:*) are recognized, the installation was successful.
🛠️ Command Reference
CCW provides a rich set of commands for managing workflows, tasks, and interacting with AI tools. For a complete list and detailed descriptions of all available commands, please see the COMMAND_REFERENCE.md file.
For a detailed technical specification of every command, see the COMMAND_SPEC.md.
🚀 Getting Started
The best way to get started is to follow the 5-minute tutorial in the Getting Started Guide.
Here is a quick example of a common development workflow:
- Start a Session:
/workflow:session:start "Implement user login feature" - Create a Plan:
/workflow:plan "Implement JWT-based user login and registration" - Execute the Plan:
/workflow:execute
🤝 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.