Skill Evaluation

Test both triggering and behavior. A skill that always loads is not good even if its happy-path output improves.

Measure:

• routing recall on intended prompts;
• negative-guard specificity;
• task success with and without the candidate;
• input/total tokens and duration when telemetry exists;
• variance/flakiness across repeated live trials for risky changes.

Prefer forward tests on realistic fixtures. Promote only when correctness is preserved and the claimed efficiency/quality improvement is actually measured.