Skill Evaluator

A focused, low-cost skill for evaluating skills. Assumes the skill already exists — your job is to prove (or disprove) that it earns its keep, and feed the results back into the next iteration.

Why this exists separately from skill-creator

skill-creator is a full authoring + evaluation pipeline. That's expensive in tokens when all you want to do is run evals. This skill keeps just the evaluation parts:

• with-skill vs. without-skill (or vs. previous version) parallel runs
• Haiku-powered grader subagents
• The HTML viewer for qualitative review
• aggregate_benchmark.py for quantitative summary
• The iteration loop

It deliberately drops: skill drafting guidance, blind comparator/analyzer agents, description optimization, and the full schema reference. If you actually need to write a new skill, hand off to skill-creator. If you want to optimize triggering, see "Description optimization (deferred)" at the bottom.

The loop, at a glance