SkillLens

Use this skill to critique, rate, or compare AI agent skills. Keep the work lightweight, evidence-based, and practical.

Operating Rules

1. Respond in the user's language unless the user asks for another language.
2. Keep all generated skill content, prompts, templates, and bundled resources in English unless the user explicitly asks otherwise.
3. Read the target skill artifacts before judging: SKILL.md, visible frontmatter, referenced resources, scripts/assets when relevant, and UI metadata such as agents/openai.yaml when present.
4. Do not edit the target skill unless the user explicitly asks for changes.
5. Do not add extra mode labels. If the user asks for a broad assessment, treat it as critique unless they request a numeric score or comparison.
6. End every critique, rating, or comparison with one concise, light humorous summary sentence.

Mode Selection

• Use critique when the user wants qualitative judgment, issues, risks, or improvement priorities for one skill.
• Use rate when the user asks for a score, grade, percentage, ranking number, or "how good is this skill?"
• Use compare when the user provides two or more skills or asks which skill is better.

If the user request is ambiguous, choose the mode that requires the least extra assumption: compare for multiple skills, rate for explicit scoring language, otherwise critique.

Workflow

1. Identify the mode and target skill artifacts.
2. Inspect the artifact set before forming conclusions. Note missing artifacts as evidence, not as automatic failure.
3. For rate or compare, load references/rubric.md and use the 100-point rubric.
4. For output shape, load references/output-formats.md.
5. For score calibration, load references/calibration-examples.md when rating or comparing.
6. Produce a direct verdict first, then evidence and improvements. Keep the answer compact unless the user asks for depth.

Judgment Principles

• Judge whether another agent could use the skill successfully with only the skill package and the user's request.
• Prefer specific evidence over abstract taste. Cite file names, sections, or line numbers when available.
• Separate "where information lives" from "whether it is useful": progressive disclosure and resource usefulness are different criteria.
• Penalize bloated instructions, vague triggers, unanchored scores, unnecessary scripts, hidden side effects, and outputs that leave the user unsure what to do next.
• Reward concise triggers, executable workflows, clear mode behavior, useful resources, calibration anchors, and safe defaults.

Resources

• Use references/rubric.md for the full 100-point scoring rubric.
• Use references/output-formats.md for critique, rate, and compare templates.
• Use references/calibration-examples.md for score anchors.