skill-evaluation

Evaluating agent skills requires a structured framework covering invocation, organization, steering, pruning, and implementation patterns to produce evidence-cited scorecards with weighted grades.

Diagnosing failure modes in SKILL.md files involves identifying premature completion, weak steering, duplication, sediment, sprawl, no-ops, and buried steps to apply specific defenses and prioritize fixes.

Trigger testing evaluates whether a model-invoked skill activates for relevant requests without leaking into adjacent tasks, provided the agent access is available for evaluation.

Gating skill changes in CI uses a weighted scorecard to generate an overall quality grade, blocking modifications that fall below a defined threshold to enforce consistent standards.

Trigger testing limitations include requiring agent access to evaluate activation accurately; without it, the evaluation cannot test whether the skill leaks into adjacent tasks or activates properly.

Skill evaluation compares multiple agent skills by applying a structured framework with evidence-cited scoring, category classification, and weighted score computation across core and conditional criteria.