Skill Judge

Skill Judge

Review a Skill as an executable knowledge package, not as an essay. Judge whether the package triggers at the right time, adds knowledge the AI model needs, routes to the right resources, and helps an agent produce better work.

Applicability

Use this as an opinionated diagnostic framework. It is informed by public specifications, examples, and practical use, but it is not an official compliance suite. Do not turn a numeric score into a claim of universal quality.

Choose the review depth

• Quick review: identify blockers, the highest-value weakness, and the next change. Do not score unless a score helps the decision.
• Full audit: inspect the entire package, score all dimensions, and produce prioritized recommendations. Read rubric.md completely before scoring.
• Comparison: evaluate each Skill independently first, then compare evidence and trade-offs. Do not force a winner when the Skills serve different use cases.
• Revision review: compare the current package with the previous version and judge whether the change improves real behavior, not only presentation.

Read the complete package

1. Read SKILL.md completely.
2. List every bundled reference, script, asset, and agent metadata file.
3. Follow every resource link and loading instruction that affects behavior.
4. Check whether referenced files exist and whether shipped scripts or examples are plausible and tested.
5. Distinguish observed facts from assumptions about how an agent might behave.

Do not score a package from its frontmatter or first screen alone.

Apply the hard gates first

Report these before subjective scoring:

• missing or invalid SKILL.md frontmatter;
• folder name and name mismatch;
• vague description that does not communicate function and trigger context;
• broken resource paths or routing instructions;
• advertised capabilities that the package does not implement;
• unsafe scripts or instructions without proportionate guardrails;
• contradictory requirements that make correct execution impossible.

A polished package with a hard-gate failure is not ready.

Evaluate actual knowledge value

Classify substantive sections using three labels:

• Expert: non-obvious decisions, trade-offs, failure modes, domain procedures, or local knowledge the AI model is unlikely to supply reliably.
• Activation: knowledge the AI model may know but benefits from seeing at the right moment.
• Redundant: generic explanation, obvious advice, or repeated content that consumes context without changing behavior.

Use this classification diagnostically. Do not invent precise paragraph percentages when the boundary is ambiguous.

Ask:

• What would the agent do better after loading this Skill?
• Which content changes a decision rather than merely describing the domain?
• Which instructions encode experience that is difficult to reconstruct on demand?
• Which sections could disappear without affecting the result?

Review trigger quality

The description is the primary activation surface. It should make clear:

• what the Skill does;
• when it should be used;
• which concrete tasks, artifacts, file types, tools, or user phrases should trigger it;
• when a neighboring Skill is a better choice, if overlap is likely.

Do not reward keyword stuffing. A long description that triggers everywhere creates routing conflicts instead of discoverability.