skill-output-scoring

Score agent skill outputs with a rubric. Use when Codex is asked to evaluate a skill output, compare outputs, apply a quality rubric, summarize reviewer scoring, or decide whether a skill output is go

SKILL.md

GitHub

Skill Output Scoring

Core Workflow

  1. Identify the output, scenario, target artifact, and scoring rubric.
  2. Score each criterion with brief evidence, not just a number.
  3. Separate output quality from prompt quality and missing context.
  4. Call out blockers before averages.
  5. Compare baseline and skill-enabled scores when both exist.
  6. Recommend keep, revise, split, merge, defer, or retire based on the score and risks.

Safety Rules

  • Do not average away high-risk failures.
  • Do not treat polished writing as correctness.
  • Do not score claims as correct without evidence or source material.
  • Escalate when the output affects production, security, legal, finance, employment, medical, compliance, or customer commitments.

Deliverable Shape

For scoring reports, provide:

  • Scenario and artifact
  • Rubric used
  • Score table
  • Blockers
  • Improvements and regressions
  • Confidence level
  • Decision recommendation

References

  • Read references/skill-output-scoring-checklist.md when scoring or comparing agent skill outputs.