Audit

Audit Skill

Pre-flight

1. Check .claude/autoimprove/config.md exists. If not, stop: "Run /autoimprove:setup first."
2. Check git is available and working tree is clean.

Step 1: Run Measurement Suite

Run the measure skill to get the composite score and per-metric breakdown. Capture both the scores AND the raw command output for each metric.

For each metric defined in the config:

• Run the command (e.g., tsc --noEmit 2>&1, pnpm test 2>&1, pnpm lint 2>&1)
• Record the score (using the measure skill's scoring logic)
• Also capture the raw output for deficiency counting

Step 2: Count Individual Deficiencies

Parse the raw output from each command to count specific issues:

• Type errors: Count lines matching error patterns (e.g., error TS for TypeScript, error: for Rust). Group by file.
• Build: Pass/fail only — no granular count. If build fails, it becomes top priority.
• Tests: Count passing vs total from test runner output. Identify failing test names if any.
• Lint: Count warning/error lines from linter output. Group by rule if possible.

Step 3: Calculate Efficiency

For each metric with a gap (score

If the user picks a number, use that area's estimate. If they press Enter, use the most efficient area.

### Starting the improve loop

Count all `**Branch:** autoimprove/experiment-` entries in `.claude/autoimprove/log.md` (if it exists) to determine the experiment offset.

Invoke the improve-loop skill directly with:
- **iterations:** The estimated count for the selected area
- **focus:** Generated focus string for that area:
  - Type safety → "Fix type errors"
  - Build → "Fix build errors"
  - Tests → "Add unit tests for untested functions"
  - Lint → "Fix lint warnings"
- **session_mode:** `new`
- **baseline_score:** The score just measured
- **start_iteration:** 1
- **total_iterations:** Same as iterations
- **experiment_offset:** Count of existing experiments

### After area completes

Re-run the measurement suite (Step 1-3) to get updated scores. Then show:

━━━ Area Complete: [area name] ━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 📊 Score: XX → YY/100 (+delta) [area]: WW/WW ██████████ ✓ maxed

Remaining work: # Area Gap Est. iterations Efficiency 1 [next] Xpts M iterations X.X pts/iter ...

Remaining: ~N iterations (~XXXK tokens)

Continue to next area? (Y/n)

**If yes:** Repeat the selection prompt (or auto-select if only one area left). Start a new improve-loop session for that area.

**If no:** Stop. Print final score.

**If all areas maxed:** "🎉 Codebase score: 100/100. All areas maxed!"

### If area doesn't fully improve

If the improve loop finishes its estimated iterations but the area isn't maxed yet, show actual results:

Area [name] partially improved: XX → YY (expected ZZ) Run more iterations on this area? (Y/n)

If yes, run more iterations (re-estimate based on remaining gap). If no, move to the next area.

## Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [benmarte](https://github.com/benmarte)
- **Source:** [benmarte/autoimprove](https://github.com/benmarte/autoimprove)
- **License:** MIT

Install and usage instructions live in the source repository linked above.