skill-creator

Start by defining what the skill should do, when it should trigger, and its expected output format. Then write a SKILL.md with name and description frontmatter, draft 2-3 realistic test prompts, and iterate based on evaluation results.

Run each test prompt twice in parallel: once with the skill and once without as a baseline. Grade both outputs against objective assertions, then aggregate pass rates, timing, and token usage into a benchmark comparison.

Generate about 20 realistic should-trigger and should-not-trigger queries, then run the automated optimization loop. It evaluates description variants on a train split and selects the best performer on held-out test queries.

A good description states both what the skill does and specific contexts when to use it, including phrases users actually type. Slightly assertive wording helps because Claude tends to undertrigger skills by default.

Add scripts for deterministic repetitive tasks that test runs keep reimplementing, references for detailed documentation loaded on demand, and assets for templates or files used in outputs. Keep SKILL.md itself under roughly 500 lines.

Yes, but with adaptations: run test prompts sequentially yourself instead of spawning parallel subagents, skip baseline comparisons and quantitative benchmarking, and collect feedback inline in the conversation rather than through the browser viewer.