# Critique prompt for ChatGPT (copy-paste below the bundle)

---

You're reviewing a 9-pass engineering audit of my trade-accountability system's todo list. The audit was run by another AI assistant; I want an independent second opinion **before** I act on its recommendations.

**Don't validate. Find what's wrong, missed, or weakly reasoned.**

The audit ranks 110 pending tasks across two dimensions: leverage (compounding infrastructure, direct unlocks) and reversibility (safety for autonomous execution). It produces two queues — an autonomous Saturday queue and a supervised queue — and names a single headline pick.

## What I specifically want from you

1. **Challenge the headline.** The audit names `catalog-refactor-shared-json` as the supervised headline and `p0-freshness-dashboard` as the autonomous top pick. Read the evidence in compounding_infrastructure.md and weekend_picks_*.md. Is there a higher-leverage task the audit deprioritized or missed? Name it specifically.

2. **Challenge the bucket assignments.** Look at the 14 EXECUTION_READY tasks in classifications.md. Which one looks misclassified to you? Either bumped up to ready when it shouldn't be, or held back when it could ship.

3. **Challenge the reversibility calls.** Look at the LOW reversibility classifications (catalog-refactor + add-user-role-approval). Are the "verification" and "rollback" entries actually sufficient, or are they hand-waving past real risk?

4. **Challenge the compounding-infra rejections.** The audit qualified 7 compounding tasks and explicitly rejected 6 (ux-cross-nav, mgp-multi-tf-chart-viewer, mgp-sip-db, pm-nightly-task, freshness-db-source-side, freshness-smb-cross-host). Was any rejection wrong? Should anything qualified be demoted?

5. **Find the inconsistency.** Pick the strongest internal contradiction in the audit's reasoning — somewhere it argues X in one pass and ~X in another. There almost certainly is at least one.

6. **One thing the audit got right.** Just one. Be specific. (This keeps you honest about not over-criticizing.)

## Constraints on your response

- **Be concrete.** Reference task IDs and file names. Don't say "this analysis could be deeper" — say "the reversibility call for `add-user-role-approval` doesn't address what happens to the role column when row-level security policies depend on it."
- **No bullet-point soup.** Write in paragraphs. 600-1000 words total.
- **No restatement of the audit's findings.** Assume I've read them. Engage, don't summarize.
- **If you genuinely think the audit is solid on a question, say so and move to the next one** — don't manufacture critiques to fill space.

The audit follows this prompt. Read it, then answer the 6 questions above in order.

---

[paste contents of bundle.md below this line]
