오토스킬 – AI 에이전트를 위한 자율 학습 기술

Self-improving skills for AI agents. No commands. No setup. Just continuous improvement. curl -fsSL https://raw.githubusercontent.com/Samurai412/autoskill/main/install-remote.sh | bash curl -o ~/.agents/skills/autoskill/SKILL.md \ https://raw.githubusercontent.com/Samurai412/autoskill/main/SKILL.md git clone https://github.com/Samurai412/autoskill.git cd autoskill ./install.sh mkdir -p ~/.agents/skills/autoskill curl -fsSL https://raw.githubusercontent.com/Samurai412/autoskill/main/SKILL.md \ -o ~/.agents/skills/autoskill/SKILL.md A system where: - Skills adapt to YOUR project — learns your patterns, utilities, conventions - Projects improve from skills — skill knowledge finds opportunities in your code - Learning is continuous — happens while you work, not as a separate step Traditional skills are static — they give the same advice regardless of project. autoskill makes skills adaptive — they learn from your codebase and give project-specific guidance. Before: "Use GetComponent caching" After: "Use GetComponent caching — this project caches in Awake with expression bodies, see EnemyController.cs:15" autoskill/ ├── SKILL.md # The main skill ├── SKILL-FORMAT.md # Instrumented skill format spec └── examples/ └── unity-game-dev/ └── SKILL.md # Example instrumented skill Mode 0: Auto-Detection (Session Start) Agent starts working → Scans project for file types (*.cs, *.py, *.tsx) → Matches file types to skill registry → Auto-loads matching skills → Continuous learning now active No user action. Just start working. Mode 1: Explicit Adaptation User: "adapt unity-game-dev.md" → Full scan of project → Proposes code improvements → Bulk learns project patterns For deep one-time passes. Mode 2: Continuous Learning Agent reads file during work → Notices patterns in skill's domain → Queues learnings → Writes to skill on task complete Happens automatically while you work. ┌─────────────────────────────────────────────┐ │ │ │ SKILL ──────────────────────────┐ │ │ │ │ │ │ │ proposes │ learns │ │ │ improvements │ from │ │ ▼ │ │ │ PROJECT ─────────────────────────┘ │ │ │ │ Each cycle: skill gets smarter │ │ project gets better │ │ │ └─────────────────────────────────────────────┘ Add to frontmatter: domain: file_patterns: ["*.cs"] # What files to watch watch_for: - pattern: "GetComponent in Awake" category: component-caching type: exemplar - pattern: "GetComponent in Update" category: component-caching type: anti-pattern principles_locked: true Add sections: ## Principles 1. Cache expensive operations 2. Avoid allocations in hot paths ## Project Learnings _Updated as agent works._ > adapt my-skill.md Scanning project... Found 5 opportunities, 8 exemplars Proposals: 1. PlayerController.cs:45 - Cache GetComponent Apply? [All / Select / Learn only] After initial adaptation, just work. The skill learns continuously: Agent reads code → notices patterns → updates skill You never invoke anything. Skill just gets better. What gets learned: - Patterns that appear 2+ times (not one-offs) - Patterns that don't violate any principle - Project-specific implementations of general techniques What gets rejected: - Anti-patterns (even if frequently used) - Patterns that violate locked principles - One-off implementations Before working on a project: ## Project Learnings _Updated automatically._ After working on "Horse & Knight Arena": ## Project Learnings ### Horse & Knight Arena (2026-03-21) - [Learned]: Expression body caching: `_rb = GetComponent();` - [Learned]: ObjectPool utility at Scripts/Utils/ObjectPool.cs - [Learned]: [RequireComponent] attribute on all dependencies - [Learned]: ScriptableObject event channels at /Events/ | Traditional Skills | autoskill | |---|---| | Static advice | Project-specific guidance | | Same for all projects | Learns your patterns | | Manual updates | Self-improving | | Generic examples | References YOUR code | | One-size-fits-all | Fits YOUR project | Skills that get smarter every time the agent reads your code. MIT