skill-creator

Start by describing what the skill should do and when it should trigger. The skill creator interviews you about edge cases and output formats, drafts a SKILL.md with proper frontmatter, then runs test prompts to validate behavior before iterating on feedback.

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 comparing the two configurations.

Generate 20 realistic trigger and no-trigger eval queries, review them in the HTML editor, then run the optimization loop. It rewrites the description iteratively and selects the best version based on held-out test query scores to avoid overfitting.

A good description states both what the skill does and specific contexts when to use it, including phrasings users might actually type. Slightly pushy descriptions combat undertriggering, and all when-to-use information belongs in the description, not the body.

Yes, but with adaptations. Without subagents you run test cases sequentially yourself, skip baseline comparisons and quantitative benchmarking, and present results inline for qualitative feedback instead of using the browser-based review viewer.

Variance comes from non-deterministic model behavior, which is why the benchmark runs each configuration multiple times and reports mean with standard deviation. High-variance evals may indicate flaky assertions or ambiguous skill instructions worth clarifying.