Skill Forge Audit

Skill Forge Audit

Batch-run skill-forge-judge across every skill in the repo. One consolidated report — grades, top issues, priority order. Read-only: no fixes applied.

Requires skill-forge-judge to be installed.

No references/ — this skill is single-file by design; the whole workflow fits in the body, so the absent directory is intentional, not missing progressive disclosure.

Workflow

Phase 1 — Discover

Glob for all SKILL.md files, run from the repo root. Skills live one level deep under skills/:

pattern: skills/*/SKILL.md

Collect the directory name of each match — that's the skill name.

To detect skipped directories: also Glob skills/*/ to list every skill directory. Any directory in the skills/*/ result that is not in the skills/*/SKILL.md result has no SKILL.md — list it in the Skipped section of the report with reason "No SKILL.md found".

If zero SKILL.md files found: output "No skills found in this repo." and stop.

Phase 2 — Evaluate

For each discovered skill, in order:

1. Read the SKILL.md file at its absolute path.
2. Invoke /skill-forge-judge on it.
3. From the judge output, extract and record:
   • Skill name
   • Grade (A/B/C/D/F)
   • Score (X/Y, Z%)
   • Numbered Improvements: up to 5 items (all if fewer than 5 exist)

Emit a one-line status per skill as each evaluation completes — do not wait until all are done:

✓ skill-forge-create   B  (98/120, 82%)
✓ skill-forge-hitl     C  (90/120, 75%)

If /skill-forge-judge is not installed: output "skill-forge-judge is required — install it before running skill-forge-audit." and stop.

If N > 10: output "About to run N skill-forge-judge evaluations — this may take several minutes. Proceed? (y/n)" and stop if denied.

If a skill's judge run errors: record ERR and the error message for that skill; continue to the next skill. Do not abort the audit.

Phase 3 — Render Report

Before rendering, ask: does every layout decision put highest-priority work at the top? If the Needs Work section is empty, lead with a one-sentence "All skills passing" summary instead of an empty header.

Sort results by grade ascending: F → D → C → B → A. Worst grades appear first — the report is a priority queue, not an alphabetical list.

Split skills into two groups:

• Needs Work: grade below B (< 80%)
• Passing: grade B or above (≥ 80%)

For the <YYYY-MM-DD> report header, use the current date already present in the session context — do not shell out for it ($(date) and command substitution trigger a permission prompt under this repo's bash rules).

Output a single markdown report:

# Skill Forge Audit
_<YYYY-MM-DD> — <N> skills evaluated_

## Grade Summary

| Skill | Grade | Score |
|-------|-------|-------|
| skill-forge-X | F | 55/120 (46%) |
| skill-forge-Y | C | 78/120 (65%) |
| skill-forge-Z | B | 98/120 (82%) |

## Needs Work

### skill-forge-X — F (55/120, 46%)

1. <improvement 1 from judge>
2. <improvement 2>
3. <improvement 3>

### skill-forge-Y — C (78/120, 65%)

1. <improvement 1>
2. <improvement 2>

## Passing (B+)

| Skill | Grade | Score |
|-------|-------|-------|
| skill-forge-Z | B | 98/120 (82%) |

## Skipped

| Directory | Reason |
|-----------|--------|
| some-dir  | No SKILL.md found |

If all skills pass: omit the "Needs Work" section entirely. If no skills pass: omit the "Passing" section. If nothing was skipped: omit the "Skipped" section.

NEVER

• NEVER display full judge reports inline for each skill Instead: Extract only grade, score, and top 1–5 numbered improvements per skill. Why: Unfiltered judge output for 6+ skills floods the context window and buries the aggregate signal the report exists to surface.

• NEVER apply any fixes during an audit run Instead: Output the report only. Direct the user to /skill-forge-hitl or /skill-forge-update for remediation. Why: Mixing diagnosis with treatment makes it impossible to know what the baseline was — the audit's value is the before-state snapshot.

• NEVER sort by skill name or alphabetically Instead: Sort by grade ascending (F first, A last) within each section. Why: Alphabetical sort buries the worst skills; the report's job is to surface where effort is needed most.

• NEVER read SKILL.md files via ~/.claude/skills/ paths Instead: Use the absolute path to the skills project directory (the repo root where this audit runs). Why: ~/.claude/skills/ is a symlink back to the project; Claude Code's permission system fires on symlink traversal and interrupts the audit mid-run with a permission prompt.

• NEVER abort the audit when a single skill errors Instead: Record ERR for that skill and continue to the next one. Why: A failed judge run on one skill should not discard results already collected for the others — partial results are more useful than none.