claude-skills

Personal library of Claude Code skills, one directory per skill, published via git. README.md is the human-facing index; this file is for whoever (Claude or Tomasz) is authoring or updating a skill here.

Layout convention

Each skill is a self-contained folder at the repo root:

<skill-name>/
├── SKILL.md              required — YAML frontmatter (name, description) + instructions
├── references/           optional — domain-specific docs loaded on demand, linked from SKILL.md
└── evals/evals.json       test prompts + assertions used to grade the skill

Mirror truth-mode/ and blind-spots/ for structure and tone when adding a new one — short, why-explained instructions rather than rigid ALWAYS/NEVER lists, references split out by domain rather than one giant SKILL.md.

Creating or iterating a skill

Use the example-skills:skill-creator skill for this — don't hand-roll the eval loop. It drives: draft SKILL.md → write realistic test prompts → spawn with-skill and without-skill subagent runs in parallel → grade against assertions → aggregate a benchmark → open a browser review viewer → collect feedback → revise.

Important: the eval workspace (per-iteration run outputs, grading.json, benchmark.json, feedback.json) is NOT kept in this repo. It lives on the local machine at ~/.agents/skills/<skill-name>-workspace/iteration-N/, sibling to ~/.agents/skills/<skill-name>/. Only the finished skill and its evals/evals.json (the prompts + assertions, reusable for the next person who iterates on it) get committed here.

Adding a skill to this repo

1. Copy the finished skill folder in (SKILL.md, references/, evals/evals.json — no workspace/run output).
2. Add a section to README.md under ## Skills, matching the existing entries: a linked heading, a 1-2 sentence description of what it does, and a sentence on what phrases/contexts trigger it.
3. No per-skill install section needed — the root ## Installing a skill section already generalizes (swap the folder name).

License

Single LICENSE (MIT) covers every skill in the repo — don't add per-skill licenses.