Skill Creator

Problem

When building an Agent Skill, the hard part is not writing a description, but confirming whether the skill triggers on realistic prompts, produces stable output, and performs better than a no-skill baseline. Editing SKILL.md by intuition can hide failures until later use.

How It Works

skill-creator turns creation into an executable loop:
- Capture intent: confirm the skill goal, triggering contexts, output format, and whether tests are needed.
- Draft the skill: write name, description, compatibility notes, and body content, with the description covering concrete use cases.
- Create tests: generate 2-3 realistic user prompts and store them in evals/evals.json.
- Run comparisons: launch with-skill and baseline runs in parallel, then collect timing.json, grading results, and benchmark data.
- Iterate: rewrite the skill based on qualitative feedback, quantitative assertions, and edge cases, then expand the test set.

Notes

This skill fits verifiable workflows such as file transformation, data extraction, code generation, or fixed procedures. For subjective outputs like writing style or design quality, avoid forcing quantitative assertions; keep human review and treat test results as iteration signals rather than final proof.