skill-judge

Run an evaluation that scores the skill across eight dimensions totaling 120 points, including knowledge delta, specification compliance, and progressive disclosure. The output is a graded report (A through F) with critical issues and prioritized improvements.

A good description answers three questions: what the skill does, when it should be used, and which keywords trigger it. Agents only see descriptions before deciding to load a skill, so vague descriptions mean the skill never activates.

Knowledge delta is the gap between what a skill provides and what the model already knows. Content is classified as Expert (keep), Activation (keep if brief), or Redundant (delete), and good skills contain over 70 percent Expert content.

SKILL.md should stay under 500 lines, ideally under 300. Heavy content belongs in references files loaded on demand through explicit triggers embedded in the workflow, with Do NOT Load notes to prevent over-loading.

Nine patterns are detected, including The Tutorial (explaining basics Claude knows), The Dump (800+ line files), The Orphan References (files never loaded), and The Invisible Skill (great content but weak description).

No external dependencies are required. The evaluation is self-contained and only needs read and write capabilities to inspect skill files and generate the structured evaluation report.