Skill2

Skills for your skills.

An installable skill library that teaches agents to create, test, package, publish, audit, and visualize your skills.

中文

GitHub stars MIT license

Skill2 terminal workflow

Install

Claude Code

/plugin marketplace add MisterBrookT/skill2
/plugin install skill2@skill2-marketplace

Installs six self-contained Skills.

Codex

npx skills add MisterBrookT/skill2 -g -a codex -y

Copies the six self-contained Skills for Codex. A curated /plugins listing is under review.

Manual fallback

git clone https://github.com/MisterBrookT/skill2.git ~/.skill2 && ~/.skill2/install.sh

Copies Skills only (install.sh supports --dry-run and conflict-gated --force from a checkout). Requires Git. uv is needed only when a Skill runs its deterministic script. Skill scripts use uv run --script; first run may fetch declared dependencies into the uv cache; offline use requires a warm cache. Data stays local; no hosted service, telemetry, or PyPI install for users.

Skill Library

Skill	Agent uses it when
`skill2-create`	Creating, updating, or restructuring a skill.
`skill2-test`	Testing activation and outcome in isolation.
`skill2-package`	Producing an installable candidate without remote writes.
`skill2-publish`	Preparing README, release, and public install checks.
`skill2-audit`	Finding contract, safety, and behavior gaps.
`skill2-visualize`	Viewing inventory, direct calls, zero-use candidates, test status, and conservative lifecycle review candidates.

Ask the agent directly:

Create a project-level skill for this workflow.
Audit this skill library and show only evidence-backed findings.
Visualize which skills are called directly and which have zero direct calls.

Local Evidence

Skill2 scans local Agent session logs for exact SKILL.md reads. It separates direct activation signals from broad scans, maintenance, and worker reads. APFS does not retain historical file-open counts, so Skill2 does not claim complete usage history.

Deterministic inventory and usage views run through the installed Skill-owned script:

uv run --script <skill-dir>/scripts/run -- visualize --skills ~/workspace/my-skill-library/skills --codex ~/.codex

Use --json for structured agent or script input. Low frequency is evidence, not a deletion decision. Output contains no prompt or transcript content.

Design

Skills are the product; deterministic scripts support them. The repository dogfoods every rule it teaches. Package never publishes. Publish requires dry-run and explicit confirmation. Visualize never changes a library.

Area	Prior art	Adopted
Skill format	Agent Skills spec, Anthropic Skills	Portable `SKILL.md`, progressive disclosure, owned resources.
Authoring	Superpowers, writing-skills	Skills-first structure, trigger-first descriptions, dogfood.
Evaluation	Superpowers evals, Tripwire, Waza, skill-eval, agent-skills-eval, skillci, skill-distill	Isolated runs, positive/negative routing, baseline, deterministic assertions.
Packaging	agent-scripts, awesome-copilot, scientific-agent-skills, superpowers-marketplace, Caveman, OpenAI Plugins	Idempotent installs, conflict gates, marketplace manifests, CI.

Skill2 thanks these projects and their maintainers. Their work shaped this repository's architecture and design principles.

See design and prior art.

Development

Contributor and CI use the checkout CLI. This is not a user install step:

uv sync
PYTHONPATH=src uv run python -m unittest discover -s tests
uv run ruff check .
uv run skill2 lint skills
uv run skill2 package-check .

MIT