skill-creator

Fires when

• ▸user says "make a skill for X" or "turn this into a skill"
• ▸user has a draft SKILL.md and wants to iterate on it with evals + benchmarks
• ▸user wants to optimize a skill's description for better triggering
• ▸user wants to package a folder into a .skill bundle for distribution
• ▸skill author is auditing their own skill before publishing

Skip when

• ✕user wants a one-shot generator with no human-in-the-loop iteration (skill-creator's spine is "draft → run → review → improve")
• ✕user is on Claude.ai with no filesystem (the bundled scripts assume a writeable cwd and Python)
• ✕user wants to evaluate a skill that depends on paid third-party APIs the session lacks
• ✕user just wants a SKILL.md template and not the whole eval/benchmark/packaging apparatus (overkill for that)

Takes

• prompt:natural-language a description of what the new skill should do, when it should trigger, and what it should output; skill-creator's "Capture Intent" step pulls answers from chat history when possible
• file:existing skill folder for the improve / package / description-optimize paths, pass the existing skill directory

Returns

• file:skill folder (SKILL.md + scripts/ + references/ + assets/) SKILL.md passes the bundled quick_validate.py; structure follows the three-level progressive-disclosure pattern documented in the skill
• file:.skill bundle (zip archive) produced by `python -m scripts.package_skill` — validates the folder first, then zips; emits to the skill-creator parent dir by default (not next to the input folder)
• structured-data:evals.json + benchmark.json + grading.json only produced if you run the full eval loop; on Claude.ai or "vibe with me" runs these are explicitly skipped per the skill's own guidance

Install

pip install pyyaml

• notes: Pure Python; no system deps for the create/validate/package path. The eval-viewer path additionally wants a browser; in Cowork/headless use --static.

No requirements.txt ships with skill-creator. quick_validate.py needs pyyaml (you'll get ModuleNotFoundError on first run otherwise). Every bundled script must be invoked as `python -m scripts.X` from inside the skill-creator directory — `python scripts/X.py` fails with `ModuleNotFoundError: No module named 'scripts'` because of the relative imports in utils.py. SKILL.md uses the correct form, but a casual reader running `python scripts/package_skill.py ./my-skill` will hit it.

Caveats

• quick_validate.py imports pyyaml but skill-creator ships no requirements.txt — first run dies with ModuleNotFoundError until you `pip install pyyaml` yourself
• every bundled script must be invoked as `python -m scripts.X` from inside the skill-creator dir; `python scripts/X.py` fails on relative imports (utils.py is `from scripts.utils import ...`). SKILL.md uses the right form but a casual user copying the path will hit this
• package_skill.py writes the .skill bundle into the *current* working directory (the skill-creator folder), not next to the input skill. Easy to lose track of in a multi-skill workspace
• the eval/benchmark/grading loop is the meat of the skill but requires subagents or the `claude` CLI in $PATH; on Claude.ai and Cowork the skill's own guidance is "skip benchmarking, do qualitative only" — so for ~50% of users the rigorous half is non-functional
• description optimization (`run_loop.py`) shells out to `claude -p` and is silently a no-op without Claude Code installed — there is no preflight check
• the skill encourages "pushy" descriptions to combat undertriggering; for a curator this is a double-edged sword (we got a 700-char description that does its job but reads like SEO copy)
• no schema validation on test-case JSON beyond the inline example — malformed evals.json fails at run_eval.py time, not at lint time