Auto-Skill Core 🧠

中文版本 (Chinese Version)

A cross-platform memory management system for all your agentic AIs. Whether you use Claude, ChatGPT, or Gemini, your hard-earned technical experiences roam with you, ensuring you never lose your progress when switching models.

Transform your AI from a stateless tool into a self-evolving Second Brain.

Auto-Skill is a framework designed to empower AI Coding Assistants (like Antigravity, Cursor, or Claude Code) with a recursive distillation loop. It ensures that every success, failure, and technical pitfall is captured, refined, and reused.

---

🔄 The Self-Evolution Loop

Auto-Skill operates on a continuous feedback loop. Every coding task becomes a source of new intelligence.

graph TD
    A[Start Task] -->|SKILL.md| B[Preflight & Context Preparation]
    B --> C[Execution]
    C -->|Task Done| D[SKILL_CLOSE.md]
    D -->|Create Card| E[.diary_queue.md]
    D -->|Distill Raw Entry| F[Skill Pending Zone]
    
    subgraph "Knowledge Crystallization"
    E -->|Write Diary| G[Daily Review]
    F -->|Distill KB| H[Permanent Knowledge]
    end
    
    H -->|Enhanced Retrieval| A

Loading

---

🦋 Experience Sifting (經驗分流)

The heart of the system is how it classifies and stores experiences. We don't just "save notes"; we sift them based on their utility.

graph LR
    Entry[Raw Experience Entry] --> Result{Result?}
    
    Result -- Success --> Best{Is it a Best Practice?}
    Best -- Yes --> SkillFile[skill-domain.md: SOP & Commands]
    Best -- No --> Legacy[legacy-domain.md: Long-term Archive]
    
    Result -- Failure --> Worst[skill-domain-worst-practice.md: Pitfalls Library]
    
    Legacy --> Pointer[Case Index in Skill File]
    Worst --> Pointer

Loading

Where does it go?

1. Skill Files (skill-*.md): High-density SOPs, commands, and refined rules. This is what the AI reads every time it starts a task.
2. Legacy Files (legacy-*.md): Full logs of successful projects. Used as a reference when the AI needs to see "how we did it before."
3. Worst-Practice Files (*-worst-practice.md): A "minefield map." Records exactly why certain approaches failed to prevent re-trial.

---

🧰 Integrated Module Guide

📔 1. Diary Skill (The Narrative Layer)

Trigger: "Write Diary", "Daily Review"

The Diary skill aggregates all task cards generated during the day into a cohesive daily log.

• Goal: Keep your human-self and AI-self on the same page about progress.
• Workflow: Reads .diary_queue.md → Fuses with existing daily notes → Clears the queue.

🧪 2. Distill-KB Skill (The Crystallization Layer)

Trigger: "Distill {domain}", "Crystallize Knowledge"

This is where the actual "learning" happens.

• Goal: Move raw notes from the "Pending Zone" into the "Refined Zone."
• Workflow: Scans skill files → Extracts SOPs → Routes cases to Legacy/Worst-Practice → Updates the Case Index.

---

🛠️ Quick Start

1. Installation

git clone https://github.com/Allen930311/auto-skill-core.git

2. Configuration

• Rename auto-skill.config.example.json to auto-skill.config.json.

Note — vault path (optional): Setting a vault path is the original developer's personal habit for syncing experiences to an Obsidian vault. If you do not use such a setup, simply tell your AI: "Remove the vault path step from SKILL.md — I do not need it." The AI will trim the startup flow accordingly.

3. Global Shielding

Run the reinforcement script to bake the protocol into your IDE:

python scripts/global_reinforce.py

---

📜 Credits

This project is a refactored and generalized version of the original framework created by Toolsai/auto-skill.

⚖️ License

MIT License.