audit-content-quality

name	audit-content-quality
author	JM Labs (Javier Montaño)
description	Audits all SKILL.md files in a plugin against a 6-criteria excellence rubric, producing per-skill scores, plugin averages, and actionable improvement recommendations. Trigger: audit content quality, score skills, check skill quality, content rubric.
allowed-tools	["Read","Glob","Grep","Bash"]

Audit Content Quality

"Quality is never an accident; it is always the result of intelligent effort." -- John Ruskin

Scores every SKILL.md in a plugin across 6 quality dimensions (Completeness, Description Quality, Procedure Clarity, Quality Criteria, Anti-Patterns, Edge Cases). Produces per-skill scorecards, identifies the bottom 3 weakest skills, and flags systematic gaps across the plugin.

---

Procedure

1. Discover all SKILL.md files -- glob for **/SKILL.md within the target plugin root. Build an inventory with skill name (from directory name) and file path. If zero skills are found, report an error and halt.

2. Parse each SKILL.md -- for every skill file, extract:

   • YAML frontmatter block (between --- markers)
   • Procedure section (numbered steps)
   • Quality Criteria section
   • Anti-Patterns section
   • Edge Cases section
   • Guiding principle quote

3. Score each skill on 6 criteria (1-10 scale):

   • Completeness (1-10): Has frontmatter with name, description, allowed-tools. Has procedure, quality criteria, anti-patterns sections. 10 = all present with edge cases and quote; 1 = missing most sections.
   • Description Quality (1-10): Trigger phrases present in description, clear what/when explanation, description length >30 characters. 10 = multiple trigger phrases + precise explanation; 1 = empty or single-word description.
   • Procedure Clarity (1-10): Steps are numbered, each step has a specific action (not vague), references actual tools or commands, logically ordered. 10 = every step is concrete and tool-aware; 1 = vague bullet points.
   • Quality Criteria (1-10): At least 4 verifiable criteria present. Each criterion is testable (not subjective). 10 = 5+ measurable criteria; 5 = exactly 4 criteria; 1 = fewer than 2.
   • Anti-Patterns (1-10): At least 3 anti-patterns with explanations of why they are harmful. 10 = 5+ well-explained anti-patterns; 5 = exactly 3; 1 = none or unlabeled.
   • Edge Cases (1-10): At least 2 edge cases with explicit response strategies. 10 = 4+ edge cases with handling; 5 = exactly 2; 1 = none.

4. Calculate per-skill scores -- sum the 6 criteria for each skill (max 60). Compute a percentage. Assign a letter grade: A (90-100%), B (80-89%), C (70-79%), D (60-69%), F (<60%).

5. Calculate plugin average -- average all per-skill scores. Identify the bottom 3 weakest skills by total score.

6. Identify systematic gaps -- for each of the 6 criteria, compute the average across all skills. Flag any criterion whose average falls below 6.0 as a systematic weakness.

7. Generate content quality report with:

   • Per-skill scorecard table (skill name, 6 scores, total, grade)
   • Plugin average score and grade
   • Bottom 3 weakest skills with specific improvement suggestions
   • Systematic gaps section with remediation priorities
   • Summary statistics (total skills, average score, score distribution)

Quality Criteria

• Every SKILL.md is scored; none are silently skipped or missed by the glob.
• Scores are justified -- each score includes a one-line rationale explaining why that number was assigned.
• Systematic gaps are identified when any criterion averages below 6.0 across the plugin.
• The weakest 3 skills are correctly identified by total score, with specific (not generic) improvement advice.
• The final report includes a machine-parseable scorecard table.

Assumptions & Limits

• Read-only. This skill never modifies SKILL.md files.
• Scoring is structural, not semantic -- a well-formatted but factually wrong procedure still scores high on Procedure Clarity.
• The 6-criteria rubric is designed for PQA-style skills. Skills with unconventional structures (e.g., pure reference skills with no procedure) may score low on Procedure Clarity despite being fit for purpose.
• Cannot assess whether anti-patterns are actually harmful or whether edge cases are realistic -- only their presence, count, and explanation quality.
• Systematic gap detection requires ≥3 skills to be statistically meaningful. For 1-2 skill plugins, report per-skill scores only.

Good vs Bad

Bad finding:

validate-hooks: 7/10

Missing: no per-criterion breakdown, no rationale, no improvement advice.

Good finding:

| validate-hooks | 9 | 8 | 9 | 7 | 8 | 9 | 50/60 (83%) | B |
Quality Criteria scored 7: only 4 criteria present, needs 5+ for score ≥8. Add a 5th verifiable criterion.

Includes: per-criterion scores, total with percentage, grade, specific rationale for lowest score, actionable improvement.

Anti-Patterns

1. Binary scoring -- giving only 1 or 10, never intermediate values. Scores must reflect gradations.
2. Counting without reading -- scoring quality criteria count as 10 just because 4 items exist, without checking if they are actually verifiable.
3. Ignoring frontmatter parsing errors -- if YAML is malformed, the skill should score 1 on Completeness with an explicit note, not be silently skipped.
4. Generic improvement advice -- telling every weak skill to "add more content" instead of specifying which section needs what.
5. Averaging away outliers -- a plugin average of 7.5 can hide a skill scoring 2.0. Always surface the bottom performers.

Edge Cases

1. Single-skill plugin -- the "bottom 3" analysis should adapt to report only the one skill. Systematic gap analysis still applies across the 6 criteria of that single skill.
2. SKILL.md with no frontmatter -- score Completeness as 1, Description Quality as 1. Still attempt to parse and score remaining sections.
3. Extremely long SKILL.md (500+ lines) -- do not truncate. Read the full file to ensure accurate scoring. Note the unusual length as an INFO observation.
4. Non-English SKILL.md -- score based on structural criteria (sections present, counts) rather than prose quality. Note the language in the report.