skill-eval

You can evaluate Claude skills by running a structured evaluation harness that tests benchmark prompts against a SKILL.md file, measuring trigger accuracy, stability, pass rates, timing, tool usage, and evidence.

Skill release gating enforces quality requirements by running structured evaluations that supply gate evidence, ensuring new plugins behave as intended and pass benchmark thresholds before merging.

To detect drift in Claude skills, re-evaluate the skill against reference benchmarks to identify regressions, compare metrics like pass rate and timing, and patch the training data or prompts iteratively.

Yes, you can use benchmark evaluation metrics such as pass rate, timing, and tool usage to iteratively improve skill triggering descriptions, prompts, and training data for better accuracy.

A skill evaluation harness exports metrics including pass rate, timing, tool usage, and evidence, providing variance analysis and auditability for benchmark prompt sets.