efficiency-audit

Phase 0: Check Intent

Ask: "Standard audit, or specific areas to focus on?" Elevate ad-hoc requirements to High Impact in Phase 3.

Phase 1: Analyze

PLUGIN_ROOT=$(ls -dt ~/.claude/plugins/cache/litianningdatadog-marketplace/efficiency-audit/*/ 2>/dev/null | head -1)
python3 "${PLUGIN_ROOT}/scripts/analyze_conversations.py" --days 30 --project "$(basename "$PWD")" --output json 2>/dev/null \
| python3 "${PLUGIN_ROOT}/scripts/synthesize_findings.py" 2>/dev/null

Synthesis succeeds → Phase 3. Fails → re-run without pipe, go to Phase 2.

Score separately (independent exit code — never combine with synthesis above):

MEMORY_MD=$(python3 "${PLUGIN_ROOT}/scripts/resolve_memory_path.py" 2>/dev/null)
python3 "${PLUGIN_ROOT}/scripts/score_efficiency.py" .claude/CLAUDE.md ~/.claude/CLAUDE.md "$MEMORY_MD" 2>/dev/null

Score < 0.5 → warning; 0.0 → run references/recipe-book.md first. Check: references/terminal-title-check.md.

Phase 2: Synthesize

Read references/category-guide.md.

Phase 3: Report

Detect CLAUDE.md locations:

for f in ~/.claude/CLAUDE.md .claude/CLAUDE.md CLAUDE.md; do [ -f "$f" ] && echo "exists: $f"; done

One → route silently. Multiple → read references/claude-md-routing.md, wait for choice. None → ask.

Always use this exact format — never a table:

---
High Impact
Rule N — <title> (~X tokens/month)
▎ <evidence>
▎ Routing: A) <path> [← recommended] | B) <path>
---
Medium Impact  [omit if empty]
Rule N — ...
---
Hook Issues  [omit if hook_errors empty; wording below]
---
Total: ~X tokens/month. Reply: numbers (1 3), all, skip N. Dry-run before writing.

Hook Issues — if hook_errors non-empty, never write "skipping" or "already handled":

ls ~/.claude/plugins/cache/litianningdatadog-marketplace/hook-doctor/*/skills/hook-doctor/SKILL.md 2>/dev/null | grep -q . && echo installed || echo not_installed

• not_installed → Hook errors found. Install: claude plugin install hook-doctor@litianningdatadog-marketplace
• installed, not yet diagnosed → Run /hook-doctor to scan. (Evidence: X failures, Y sessions.)
• installed, deferred this session → ⚠ Run /hook-doctor when ready. (Evidence: X failures, Y sessions.)
• installed, confirmed clean → omit section.

Only cite transcript numbers (failures, sessions). Never use hook-doctor output counts.

Phase 4: Apply

Read references/governance.md. Order: memory → CLAUDE.md → settings.json. Hook fixes → hook-doctor.

python3 "${PLUGIN_ROOT}/scripts/apply_rules.py" --dry-run <path> '["r1", "r2"]'
python3 "${PLUGIN_ROOT}/scripts/apply_rules.py" <path> '["r1", "r2"]'

One rule at a time: run --dry-run, print output, STOP — ask "Looks right? yes / edit / skip." Do not write until they reply in this turn. On yes: write, verify, report. Repeat.

Phase 5: Karpathy Guardrails (opt-in)

Read references/karpathy-guardrails.md; offer once if threshold met, skip if declined. Governs Phase 4 apply, not analysis.

Utilities

• Noise false positives: references/noise-filters.md
• CLAUDE.md > 200 lines: references/recipe-book.md first
• Re-run every 2–4 weeks