Agent Skill System
Structured, self-evolving skill memory for AI agents — with built-in regression tests.
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.mdauto-logs successes/failures so skills improve with use instead of going stale. - Regression tests — 8 assertion patterns verify no degradation.
- Auto-repair —
refiner.pydiagnoses failures, patches SKILL.md, reruns tests (up to 3 rounds). - Portable — Each skill is a directory.
cp -rto 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
| Stage | Engine | What |
|---|---|---|
| Create | creator.py |
Conversation trace → SKILL.md + test cases |
| Evaluate | test_runner.py |
8 assertion patterns verify the skill |
| Refine | refiner.py |
Diagnose failures → patch → retry (max 3×) |
| Remember | memory.py |
.memory.md accumulates successes/failures |
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
| Prompt Eng | RAG | Cursor Rules | This | |
|---|---|---|---|---|
| Creates from experience | ❌ | ❌ | ❌ | ✅ |
| Independent memory | ❌ | ❌ | ❌ | ✅ |
| Automated tests | ❌ | ❌ | ❌ | ✅ |
| Self-healing on failure | ❌ | ❌ | ❌ | ✅ |
| Cross-agent portable | Manual | Tied to DB | Tied to editor | ✅ cp |
| Training needed | None | None | None | None |
License
MIT
Metadata
Release files for agent-skill-system 1.0.6
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| agent_skill_system-1.0.6.tar.gz | 45.0 kB | Without provenance | Details for agent_skill_system-1.0.6.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| agent_skill_system-1.0.6-py3-none-any.whl | Python 3 | none | any | Details for agent_skill_system-1.0.6-py3-none-any.whl |
Total release size: 101.1 kB
Release files / agent_skill_system-1.0.6.tar.gz
| Download URL | agent_skill_system-1.0.6.tar.gz |
|---|---|
| Size | 45.0 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
2ae359d0eaf4f37bf69d913f73408b24b22febe70eeef58a6b3008241d0d9aeb
|
|
BLAKE2b-256 checksum How to use checksums |
8bf321aa1a688f9a05e4144db06483d139afcc3336d090f9bc1ce22886c70522
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.2.0 CPython/3.12.8
|
Release files / agent_skill_system-1.0.6-py3-none-any.whl
| Download URL | agent_skill_system-1.0.6-py3-none-any.whl |
|---|---|
| Size | 56.1 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
9bf8708934d5634c72aa10de4c071c22d7a4558b0a51583743536e96d4cde154
|
|
BLAKE2b-256 checksum How to use checksums |
ccd70139247c3f6acdf9a4c31032aa02ec4133857391b2c60f4e2f8c1802408f
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.2.0 CPython/3.12.8
|
Security for agent-skill-system 1.0.6
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
No provenance found
None of the files in this release were uploaded with provenance metadata.
A note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.