skill-judge

To evaluate if a SKILL.md file provides expert knowledge or redundant basics, audit its knowledge delta by scoring content as Expert, Activation, or Redundant to detect token waste. This structured evaluation reveals whether the skill delivers real value or just re-explains what the model already knows.

Knowledge delta scoring in prompt engineering rates content as Expert, Activation, or Redundant to measure the actual expertise injected into a model. This mechanism identifies whether your prompt engineering adds unique value or simply wastes tokens on basic knowledge the model already possesses.

You audit prompt engineering anti-patterns in agent workflows by applying a quality checklist that evaluates mindset, specific NEVER lists, and description-trigger compliance. This validates frontmatter and keywords so the skill activates reliably without triggering unwanted behaviors.

Skill evaluation does not require any specific dependencies or components to function. You can use it independently to review candidate SKILL.md files and decide whether they are worth keeping, tightening, or rewriting before publishing.

The best way to reduce token waste in agent skills is to audit their SKILL.md content for knowledge delta and anti-pattern coverage. By scoring content against expert knowledge and removing redundant basics, you produce actionable improvement guidance to tighten the skill.

You should rewrite a SKILL.md package instead of keeping it when evaluation reveals low knowledge delta, poor anti-pattern coverage, or description-trigger non-compliance. If the content mostly re-explains basics the model already knows, rewriting is necessary to improve activation effectiveness.