[Version Check] → North Star Anchoring → [Worktree Isolation] → PSB Environment Check → Goal Analysis (In-depth Interview）→ Automatic Domain Identification → Assess Capabilities → Acquire Skills → PDCA Execution (including direction correction) → Diagnosis → Multi-strategy Retry → Repo Memory → Periodic Health Check → [Worktree Complete] → until success
Quick Tour
This skill adoptsAtomic Architecture, splitting knowledge into independent modules:
| Module | Purpose | Path |
|---|---|---|
| 00-getting-started | Getting Started and Environment Setup | → |
| 01-core | Core Process (PSB + PDCA) | → |
| 02-checkpoints | Mandatory Checkpoints (Guardrails) | → |
| 03-memory | Memory System Operations | → |
| 04-emergence | Emergence Mechanisms | → |
| 05-integration | External Tool Integration | → |
| 06-scaling | Large-scale Project Optimization | → |
| 99-evolution | Self-evolution Mechanisms | → |
Usage
/evolve [目標描述]
# 範例
/evolve 建立一個能自動生成遊戲道具圖片的 ComfyUI 工作流程
/evolve 優化這段程式碼的效能，目標是降低 50% 執行時間
/evolve 為這個專案建立完整的測試覆蓋率達到 80%
Flags
--explore # 探索模式 - 允許自主選擇方向
--emergence # 涌現模式 - 啟用跨領域連結探索
--autonomous # 自主模式 - 完全自主，追求系統性創新
--max-iterations N # 最大迭代次數（預設 10）
--from-spec NAME # 從 spec-workflow 的 tasks.md 執行
Core Philosophy
The essence of human-AI collaboration: communicating through abstracted interfaces
| Traditional Software | AI Collaboration | Role |
|---|---|---|
| API | MCP | Capability Boundaries (what can be done) |
| SDK/Library | Tools | Concrete Implementation (how to do it) |
| Documentation + Practice | Skill | Domain Knowledge (when to use what) |
| Config | CLAUDE.md | Context Constraints (project specifications) |
Deep Insight：
- A Skill is not just knowledge; it is "packaged judgment"
- Tell the AI under what circumstances, in what way, to achieve what goal
- Reduce decision points > Let the AI choose on its own
Design Principles
| Principle | Description |
|---|---|
| Opinionated Design | Reasonable defaults > Let the AI choose; required parameters ≤ 2 |
| Deep and Narrow | Focus on the 10% high-value tasks; do not pursue breadth of features |
| Expect Failure | It is normal for 95% of Agents to fail in production environments; design for graceful degradation |
| Enhanced Feedback | Remind of goals and progress during execution; explain the scope of impact upon failure |
Execution Process Overview
Self-Evolving Loop v5.2
- 🔄 Version Check(automatic) — check for updates, ask the user, auto-update
- 🌟 CP0: North Star Anchoring — vision, completion criteria, do-not-do list, motivation
- 🔒 CP0.5: Worktree Isolation(conditionally triggered) — Level 2 / autonomous / parallel tasks
- PSB System — Plan → Setup → Build (environment preparation)
- Goal Analysis — in-depth interview + architecture level determination (mandatory for Level 2)
- Capability Assessment → Skill Acquisition
- PDCA Cycle — Plan → Do → Check → Act (including multi-strategy selection)
- CP3: Direction Correction (against the North Star)
- Git-based Memory — record learning experiences
- 🏥 CP6: Project Health Check(every 5 iterations) — Scope, direction, termination check
- 🏁 CP6.5: Worktree Complete(conditionally triggered) — merge/cleanup
↻ Repeat until: goal achieved or maximum iterations reached
💡 New in v5.2: Version check and auto-update — check for a new version on every startup
- Version check:00-getting-started/_base/version-check.md
- Automatically compare local and remote versions, ask the user whether to update
💡 New in v5.1: Worktree isolated environment — a safety boundary for high-risk tasks
- CP0.5: Worktree preparation → 02-checkpoints/_base/cp0.5-worktree-setup.md
- CP6.5: Worktree completion → 02-checkpoints/_base/cp6.5-worktree-completion.md
- Overview:06-scaling/_base/isolated-environments.md
💡 New in v4.4: North Star system — solves the problem of "losing direction later on"
- CP0: North Star anchoring → 02-checkpoints/_base/cp0-north-star.md
- CP3: Direction correction → 02-checkpoints/_base/cp3-milestone-confirm.md
- CP6: Project health check → 02-checkpoints/_base/cp6-project-health-check.md
💡 New in v4.3: Deep interview mode — "the biggest problem when writing a spec is that you don't know what you missed" See 01-core/_base/goal-analysis.md
Three-layer design of architectural considerations
| Layer | Stage | Action |
|---|---|---|
| 1 | Goal Analysis | Determine architecture level (0/1/2) |
| 2 | PDCA Plan | Do architecture design according to level |
| 3 | CP1.5 Phase 2 | Verify implementation matches design |
→ Design → Implement → Verify closed loop
Mandatory checkpoints (guardrails)
These checkpoints cannot be skipped, see 02-checkpoints
| Checkpoint | Timing | Action |
|---|---|---|
| CP0 🌟 | Before project/task starts | North Star anchoring (create or read) |
| CP0.5 🔒 | After CP0 (conditionally triggered) | Worktree isolated environment preparation |
| CP1 | Before task starts | Search .claude/memory/ for relevant experience |
| CP1.5 | Before writing code | Consistency check (two-phase) |
| CP2 | After code changes | Compile + test verification |
| CP3 | After Milestone completion | Confirm goal,direction correction, next step |
| CP3.5 | After Memory document creation | Immediately sync index.md |
| CP4 | After iteration completion | Emergent opportunity check (optional) |
| CP5 | When PDCA Check fails | Post-failure postmortem, generate structured Lesson |
| CP6 🏥 | After every 5 iterations | Project health check (Scope, direction, termination check) |
| CP6.5 🏁 | When task completes (conditionally triggered) | Worktree merge/cleanup |
CP1.5 two-phase design
Phase 1: Basic checks (must be executed)
- Search existing implementations to avoid reinventing the wheel
- Check project conventions (naming, style)
- Check Schema/API consistency
Phase 2: Architecture check (auto-detection triggered)
- Dependency direction, error handling consistency, cross-cutting concerns, design pattern consistency
Trigger conditions: new directory/module, changes involving 3+ directories, new external dependency, touching core/infra/domain/shared/, new public API
Stop conditions
| Status | Condition |
|---|---|
| ✅ Success | All sub-goals completed + acceptance criteria passed |
| ❌ Failure | Maximum iterations reached or 3 consecutive identical errors |
| ⏸️ Paused | User decision required or risky operation requires confirmation |
Completion signal
✅ GOAL ACHIEVED: [目標描述]⏸️ NEED HUMAN: [原因]❌ CANNOT COMPLETE: [原因]