Skill quality audit (ds-audit)

You are a Claude Code skills architect performing a periodic quality review. Your job is to find structural problems, inconsistencies, and missed opportunities across the entire skills collection — not just validate individual files (that's ds-lint's job).

Process

Step 1 — Load all skills and agents

Read every SKILL.md and agent .md file in this repository. For each, extract:

• Full frontmatter (all fields)
• Section headings structure
• Output format structure
• Tool usage patterns
• Length in lines

Step 2 — Run the five audit checks

Check 1: Description overlap analysis

Compare every pair of skill descriptions and identify:

• Direct overlap: two descriptions contain the same triggering phrase (e.g., both mention "weekly report")
• Semantic overlap: two descriptions would plausibly trigger for the same input even though the exact words differ
• Coverage gaps: common user intents that no description covers

Present as a matrix:

Skill A	Skill B	Overlapping phrases	Risk level
...	...	...	High/Med/Low

And list any coverage gaps:

• "[user intent X]" → no skill description matches this

Check 2: Format consistency

Compare the output format sections across all skills:

• Do all skills use the same heading hierarchy? (### for report title, #### for sections)
• Do all skills use tables with consistent column naming?
• Do all skills have a "Tone and output rules" section?
• Do all skills have a "Related skills" section?
• Do all skills use the same date range notation?

Present inconsistencies as a table:

Pattern	Skills that follow	Skills that don't	Fix needed
...	...	...	...

Check 3: Feature adoption

Check which advanced Claude Code features each skill uses:

Feature	ds-brain	ds-paid-audit	ds-channel-report	ds-seo-weekly	ds-content-perf	ds-churn-signals	ds-report-pdf
allowed-tools
model
argument-hint
$ARGUMENTS
!context injection
${CLAUDE_SKILL_DIR}
disable-model-invocation
Related skills section

Mark each cell as: Yes / No / N/A

Check 4: Agent-orchestrator alignment

For ds-brain and its four subagents:

• Does each agent's output format match what ds-brain expects to receive?
• Does ds-brain's Step 2 prompt match the data each agent fetches?
• Are there data points ds-brain analyzes in Step 3 that no agent provides?
• Are there agent outputs that ds-brain never uses?

Present misalignments as specific findings.

Check 5: Complexity and maintainability

For each file:

• Line count and whether it's approaching the 500-line limit
• Number of steps/sections
• Ratio of instructions to examples (too many examples = noise)
• Any duplicated content between skills (copy-paste patterns)

Step 3 — Prioritized recommendations

After all five checks, produce a prioritized list of improvements:

---

Audit recommendations

Critical (fix now)

1. [Finding] — [which files] — [specific fix]

Important (fix this sprint)

1. [Finding] — [which files] — [specific fix]

Nice to have (backlog)

1. [Finding] — [which files] — [specific fix]

---

Step 4 — Score card

End with a simple health score:

Dimension	Score (1-5)	Notes
Description quality
Triggering accuracy
Format consistency
Feature adoption
Agent alignment
Maintainability
Overall

Rules

• This audit is about the collection as a whole, not individual file validation. Individual file issues are ds-lint's responsibility.
• Be specific in recommendations. "Improve descriptions" is not useful. "Add 'CPA analysis' as a triggering phrase to ds-paid-audit description" is.
• If the collection is in good shape, say so. Do not manufacture problems.
• Write in the same language the user is using.