audit-fix-loop

Announce: "Using audit-fix-loop to plan a scored iterative improvement loop."

What this skill is (and isn't)

/goal is the mechanism: it re-fires turns and uses a separate-model evaluator to gate exit on a condition it reads from the transcript. This skill is the doctrine that tells you how to set a loop up so /goal enforces the right thing — /goal will just as happily enforce a wrong condition.

It is also the generic fallback. The domain workflows below are specialized, pre-wired instances of this same doctrine — prefer them when one fits. Use this skill directly only for ad-hoc "improve " requests that don't map to any of them:

Domain workflow	Use it instead when…
visual-verify	improving slides/charts/rendered output (Gemini vision, zero-blocking-defects gate)
bluebook-audit	correcting Bluebook citations in a DOCX
source-verify	verifying citations exist + quotes match sources
writing-review / writing-revise	improving a prose draft
workflow-creator Mode 3	hardening a skill/workflow (wc-audit substrate gate)

## Iron Law 1 — The auditor must not be the fixer

This is the one thing /goal cannot supply. /goal decides when to exit; it does not enforce that the score came from someone other than the fixer. If the agent that wrote the fix also grades it, you get rubber-stamping — the fixer's opinion of its own work is worthless.

The audit must be structurally independent every iteration: a fresh subagent (no fixer context), a different model (e.g. Gemini), or a mechanical checker. If you ran an audit with the fixer's context, DELETE the result and re-run with a fresh auditor — tainted findings are worse than none.

## Iron Law 2 — Gate on the substrate, not on a bare score

Terminate on the deterministic/categorical substrate going clean + the 0-10 score going FLAT — never on a bare "composite ≥ 9.5".

The substrate is the convergent signal: zero CRITICAL and zero HIGH findings outstanding (and, where the domain has them, mechanical gates pass / citations resolve / zero blocking defects). It monotonically converges and is what "done" actually means.

The 0-10 score is a noisy LLM proxy: it re-rolls ±0.2 each run and regenerates fresh minor findings every pass, so it asymptotes (empirically ~9.0) and never stably crosses 9.5. Chasing it is a treadmill where every fix surfaces a new nit and the last half-point is only buyable by over-engineering the artifact — which makes it worse. (See project_wc_mode3_asymptote.) Treat the score as an advisory thermometer and a flatness check, not the summit.

So the exit condition is: substrate clean AND score flat (within ±0.2 of the prior turn, at/above your chosen floor) — or the turn budget elapses. The threshold you pick in Step 1 is a floor the substrate must clear, not a bar to grind toward. For a pure-judgment scorer with no hard substrate (e.g. prose rhythm), gate on convergence/flat + zero blocking alone — there is no threshold to chase.

## Iron Law 3 — No `/goal` without a plan

A naive condition like "fix all issues" gives the evaluator nothing concrete to check. Before setting the /goal, identify: (1) the artifact, (2) the scoring surfaces, (3) how the audit stays independent, (4) the substrate gate + score floor. The condition MUST pin completion to external state the auditor writes — .planning/SCORES.md (score trend + finding counts) — so the evaluator reads the substrate and the flatness from the transcript, not from your say-so.

If you set a /goal without planning, run /goal clear, plan via Step 1, and set a new one. No patching a naive condition mid-flight.

The loop

PLAN (Step 1) → /goal <substrate-gate condition pinned to .planning/SCORES.md>
  │
  └─► each turn the evaluator re-fires:
        AUDIT  fresh independent auditor scores the artifact
               → findings → .planning/AUDIT.md ; score + CRITICAL/HIGH counts → .planning/SCORES.md
        DECIDE substrate clean (0 CRITICAL/0 HIGH) AND score flat?  → end turn; evaluator marks done
               substrate dirty, OR score still climbing run-to-run? → FIX
        FIX    apply targeted fixes worst-severity-first (CRITICAL → HIGH → MEDIUM); do NOT self-assess
        → end turn immediately (the /goal refires for re-audit)

/goal owns the re-firing and the exit decision. You own AUDIT → DECIDE → FIX inside each turn. After fixing, do not pause to summarize or ask "should I continue?" — end the turn so the goal refires. The evaluator decides when to stop, not you.

Step 1: Plan the loop

AskUserQuestion(questions=[
  {"question": "What artifact are you improving?", "header": "Artifact", "multiSelect": false,
   "options": [
     {"label": "Writing draft", "description": "Document, essay, paper, or prose"},
     {"label": "Skill or workflow", "description": "SKILL.md / workflow being hardened — prefer workflow-creator Mode 3"},
     {"label": "Visual output", "description": "Slides/charts/rendered docs — use visual-verify instead"},
     {"label": "Citations", "description": "Footnotes / quotes — use bluebook-audit or source-verify instead"}]},
  {"question": "Which scoring surfaces should the audit use?", "header": "Scorers", "multiSelect": true,
   "options": [
     {"label": "AI anti-patterns", "description": "12-category checklist for AI writing indicators"},
     {"label": "Style guide", "description": "Legal / econ / Strunk & White"},
     {"label": "Enforcement patterns", "description": "12 superpowers enforcement patterns (skills/workflows)"},
     {"label": "Source verification", "description": "Citations vs bib + quotes vs sources — use source-verify"}]}
])

Redirect to the domain workflow if the artifact is visual / citations / a skill — they already implement this pattern, calibrated.

Derive the parameters:

Parameter	How to derive
Audit method	Fresh subagent reads the scorer's rules (see table), then audits — independence per Iron Law 1
Fix method	Self-edit for small artifacts; parallel subagents for large ones
Turn budget	10 default, encoded as Stop after N turns
Substrate gate	0 CRITICAL / 0 HIGH (+ any domain mechanical/categorical gates)
Score floor	9.5/10 default — a floor the substrate must clear, NOT a bar to grind

Scorer	Audit method (independence mechanism)
AI anti-patterns	fresh subagent reads ../ai-anti-patterns/SKILL.md + refs, then audits
Style guide	fresh subagent reads the domain skill (writing-legal / -econ / -general), then audits
Enforcement patterns	fresh subagent reads references/enforcement-checklist.md, scores all 12
Source verification	invoke Skill(skill="workflows:source-verify") — mechanical bib grep + quote search

When multiple scorers are selected, every iteration runs ALL of them and the substrate is the union of their CRITICAL/HIGH findings.

Step 2: Initialize state files

mkdir -p .planning

.planning/AUDIT.md (overwritten each iteration) — findings table per scorer: # | Severity | Finding | Location | Suggestion (severities CRITICAL/HIGH/MEDIUM/LOW).

.planning/SCORES.md (append-only) — one row per iteration so the evaluator can read both the substrate and the trend:

# Score History
| Iteration | Score | CRITICAL | HIGH | Δ vs prior | Note |
|-----------|-------|----------|------|-----------|------|
| 1 | 6.5 | 3 | 2 | — | baseline |
| 2 | 8.9 | 0 | 0 | +2.4 | substrate clean; not yet flat |
| 3 | 9.0 | 0 | 0 | +0.1 | substrate clean + flat → done |

Step 3: Set the /goal

Pin the condition to the substrate (CRITICAL/HIGH counts) and score-flatness in .planning/SCORES.md:

/goal The artifact is substrate-clean — zero CRITICAL and zero HIGH findings outstanding in
.planning/SCORES.md across the selected scorers — AND its score has gone flat (within ±0.2 of the
prior turn, at or above the floor). Audit with a fresh independent auditor, then fix worst-first,
each turn. Stop after 10 turns. Do NOT keep iterating to lift a flat score once substrate-clean.

The fresh-auditor prompt (Phase A, every turn):

Agent(prompt="""
You are an independent auditor with NO knowledge of any prior fixes.
Read the scoring rules: [SCORER SKILL PATH]
Then audit this artifact: [ARTIFACT PATH]
Output findings EXACTLY as: | # | Severity | Finding | Location | Suggestion |
Severities: CRITICAL, HIGH, MEDIUM, LOW. Be thorough — a clean audit that misses issues is worse
than a harsh one. Do NOT soften findings. Do NOT say "overall good." Then give a 0-10 score.
""", subagent_type="general-purpose")

Compile into .planning/AUDIT.md, append the row to .planning/SCORES.md, then DECIDE → FIX → end turn.

Loop Facts

• A CRITICAL caused by a measurement artifact still keeps the substrate dirty. Verify it against the real files, then either fix it or neutralize the artifact — waiving it informally ("it's just an artifact") converts the gate back into the honor system, which is claiming the gate without checking it.
• A flat score below the floor with a clean substrate is the domain ceiling / judge noise, not a defect — record the number as the honest reading and stop. Grinding a flat score toward 9.5 over-engineers the artifact and makes it worse (Iron Law 2).