| name | parallel-skills-audit |
| description | Triage all skills then spawn parallel agents to deep-audit flagged ones. Auto-activates when batch auditing skills or optimizing multiple skills at once. Trigger keywords: parallel audit, batch audit, audit all skills, bulk optimize skills. |
| license | MIT |
| metadata | {"version":"1.2.2","hermes":{"tags":["Skills","Audit","Parallel-agents","Quality"],"related_skills":["skill-decomposition-methodology"]}} |
Parallel Skills Audit
Step 1: Triage
Triage every skill: record each one's line count and whether its name collides across scopes (project vs global). Collect the status table.
Step 2: Filter and Confirm
Collect skills flagged LARGE (>300 lines) or DUP (same name in 2+ scopes after symlink resolution). Apply scope filter if provided. Sort by line count descending.
Present: "Found N skills to audit in batches of M. Proceed?" Wait for confirmation.
Step 3: Batch Dispatch
For each batch (default 5 concurrent):
Spawn parallel background agents. Each agent prompt:
Deep audit the skill at {skill_path} ({line_count} lines).
Read SKILL.md and references/. Then:
1. Spec check: frontmatter (name, description ≤1024), body (<500 lines, <5000 tokens)
2. Five-step optimize: question, delete obvious/duplicate/hypothetical, tighten
3. Decompose: if >300 lines, extract to references/
4. Fix frontmatter, report results
IMPLEMENT spec, dedup, and tightening fixes. Do NOT implement behavioral
changes (removing a documented option, swapping a banned command) — report
those as findings instead, which the user confirms before anyone applies them.
Return: skill name, before/after lines, what was deleted, behavioral findings.
Collect behavioral findings across the batch and present them to the user for confirmation before anyone applies them.
Report batch results: Skill | Before | After | Reduction. Ask before next batch.
Step 4: Summary
Cumulative results table. Total: skills processed, lines removed.
