Agent Skill System

Structured, self-evolving skill memory for AI agents — with built-in regression tests.

中文文档

AI agents forget. Teach one to review contracts, and tomorrow it makes the same mistakes — no memory across sessions, no way to catch degradation.

agent-skill-system turns corrected behavior into versioned, testable skill packages. Each skill bundles instructions, accumulated experience, and regression tests into a standalone directory. Next time, the engine matches the task type, loads context, and avoids repeating past errors.

vs Addy Osmani's agent-skills

agent-skills (63K stars) is a catalog of human-written prompt templates for one-shot tasks.

This solves a different problem: persistent, self-evolving skill memory.

• Self-evolving — .memory.md auto-logs successes/failures so skills improve with use instead of going stale.
• Regression tests — 8 assertion patterns verify no degradation.
• Auto-repair — refiner.py diagnoses failures, patches SKILL.md, reruns tests (up to 3 rounds).
• Portable — Each skill is a directory. cp -r to any agent.

Think agent-skills = recipe book. This = chef's notebook that learns.

Quick start

pip install agent-skill-system

agent-skill list                   # what skills are available?
agent-skill search "contract"      # find the right skill
agent-skill load contract-review   # print skill + memory context

With LLM backend (for creating/refining skills):

export OPENAI_API_KEY="sk-..."
export LLM_MODEL="gpt-4o"

agent-skill health contract-review  # run regression tests
agent-skill register my-new-skill   # register a new skill
agent-skill scan                    # auto-register skills added to skills/

Lifecycle: Create → Evaluate → Refine → Register → Use → Remember

Skill structure

skills/contract-review/
├── SKILL.md       # How to do the task correctly
├── .memory.md     # What went right/wrong — auto-accumulated
├── config.json    # Trigger keywords, version, metadata
└── tests/         # Regression tests (8 assertion types)

Each skill is a standalone directory — no framework lock-in.

vs other approaches

License

MIT