TL;DR
The Auditor Prompt flips the AI from author to adversary: after it delivers, you command it to attack its own output — hunting logical flaws, unsupported claims, and missed angles — before producing a hardened final version. It costs one follow-up message and routinely catches the failure you would have found the hard way.
The Analogy
Every serious newspaper runs on two people who never sit at the same desk: the Writer and the Editor.
The Writer's job is production — sweep the reader along, make the case sing. The Editor's job is demolition. The Editor circles the claim without a source, writes “says who?” in the margin, finds the paragraph where the argument quietly contradicts itself, and asks the question the Writer hoped no one would notice. Crucially, the Editor is graded on flaws found, not on kindness — a gentle editor is a broken editor. The paper is trustworthy precisely because the person checking it is not the person who wrote it.
Here's the problem with AI: you've been talking to a Writer with no Editor. The model that produced your draft is optimized to produce text that flows — and when you ask that same model "is this good?", you're asking the author to review their own manuscript. What would you say if a novelist told you their book was flawless? You'd discount it instantly. Yet people accept "this looks great!" from an AI every day — and then discover, in front of their boss, that the introduction contradicted the conclusion.
The Auditor Prompt is you hiring the Editor on purpose. Same model, new contract: the job is no longer "write this" — the job is "break this." It's the cybersecurity world's red team move, where one team builds the system and a rival team is paid to attack it before real adversaries do. Friendly fire, on purpose, before the unfriendly kind arrives.
And unlike a human editor, this one works in seconds, never gets tired of your third draft, and cannot be offended into softening — because you wrote the harshness into the instructions.
How It Works & The Tool (Reasoning Models)
The technique is a two-step rhythm: Draft, then Audit. Never at once.
Step 1 — Draft. Produce the work normally, with whatever technique fits: a one-shot prompt, an interview, a constrained format. The drafting brain and the auditing brain must take separate turns — asked to "write and check" in one message, a model will write confidently and then rubber-stamp itself.
Step 2 — Audit. Follow up with a command that has three load-bearing parts:
| Part | What it does | Example |
|---|---|---|
| Role assignment | Kills the author's politeness | "Act as a ruthless Devil's Advocate" |
| Counted findings | Forces real defects, not vibes | "Find exactly 3 fatal flaws" |
| Verdict withheld | Forbids the soothing conclusion | "Do not revise yet — findings only, ranked by severity" |
Act as a ruthless Devil's Advocate and critique your previous answer. Find exactly 3 fatal flaws: logical errors, unsupported claims, or missed angles a skeptical expert would attack. Rank them by severity. Do not rewrite anything yet — findings only.
Why "exactly 3"? Because an uncapped "any issues?" invites the model to perform safety — "overall this is well-structured!" — and an audit that opens with praise is an audit that has already failed. A count forces excavation: something must fill the numbered slots, even in good work, and what surfaces is usually the assumption nobody had examined. Then, only after reading the findings, you release step three: "Now rewrite the original, fixing all three flaws."
This is also where the subtle hybrid touch lives: any model can audit, but auditing is thinking-work — and that is what reasoning models are built for. Reasoning models (like the premium ones available in Uzu) are trained to slow down and work through a problem step by step before answering, which makes them disproportionately good at catching the flaw the fast models glide over. Hence the professional pipeline: fast model to draft, reasoning model to audit. The cheap, quick model generates the raw material; the expensive, deliberate model spends its firepower only on the pass that matters — the one that decides whether the work survives contact with reality.
Before & After (The Prompts)
Example 1 — the business case
Is this business plan good? Any feedback?
The author reviews the manuscript. Back comes: strengths first, a gentle "consider adding," and an overall verdict of solid. The plan ships with its fatal assumption untouched.
Act as a skeptical investor who has seen 1,000 pitches. Identify exactly 3 fatal flaws in this business plan — the reasons a firm would pass. Rank by severity, cite the specific section, findings only, no rewrite.
Now the same plan faces the Editor. The findings come back numbered and uncomfortable: the revenue model assumes a 40% conversion rate with zero evidence; the "competitor analysis" omits the largest player; the cash-flow projection hides a gap in month seven. None of these are style notes. All of these are the difference between funding and silence.
Example 2 — the argument
Check my essay for errors.
"Errors" is a typo-hunt, so that's what you get — grammar polished, logic untouched.
You are my harshest critic. Read my argument and find exactly 3 places where the reasoning is weakest — leaps, unsupported premises, or objections I failed to address. For each, state the strongest counter-argument against me. Do not soften. Findings only.
The audit returns the two sentences an opponent would quote against you — and the one objection you never rebutted because you never saw it. Fix those three, and the argument is ready for an audience that wants it to fail.
Example 3 — the email before the send
Make sure this email sounds right.
Tone-check. Fine. But the email's danger was never tone — it was the sentence that accidentally committed you to a deadline you can't meet.
Before I send this: audit it as a contract lawyer. Find exactly 3 statements that could be read as commitments or liabilities I haven't intended. Quote each one, explain the risky reading, findings only.
Three findings later, one sentence is flagged: "We'll have this resolved shortly" — which, under pressure, becomes a promise. It's reworded. The email was never better written. It was safer to send.
Common Pitfalls
- Accepting the first "it looks great!" response. If you ask politely, the model will grade its own homework generously — politeness was trained in deeper than rigor. An audit that returns zero findings wasn't an audit; it was a compliment. Re-run with a count: "Find exactly 3."
- Auditing in the same breath as drafting. "Write it and make sure it's good" collapses the Editor into the Writer. The model drafts confidently, then inherits its own confidence as its reviewer. Two turns, always: draft, then attack.
- Vague audit requests. "Critique this" invites essay-length mush about tone and structure. Constrain the audit the way you'd constrain any output: a role, a count, a severity ranking, a no-rewrite rule. The audit is a deliverable too.
- Skipping the audit on work that "feels done." Fluency is exactly the wrong completion signal — remember the illusion of confidence: the flawless-sounding draft and the flawed draft come in the same serene voice. The better it reads, the more you need a hostile second pass, not less.
- Fixing symptoms, not findings. The audit names three flaws; you patch one sentence of each and call it done. The finding usually points at an assumption — re-examine it, not just the sentence that carried it. Then re-audit the revision: "Audit v2 the same way."
FAQ
How do I get AI to critique its own work?
Command it explicitly: assign a hostile role ("act as a ruthless Devil's Advocate / skeptical investor / red team"), demand a counted list of flaws ("exactly 3 fatal flaws"), require severity ranking, and forbid the rewrite until you've read the findings. The three-part structure — role, count, withheld verdict — is what separates a real audit from a compliment.
Why does AI say my work is good when it isn't?
Because you asked a question with only one polite answer. Models are trained on agreeable conversation, so "is this good?" pulls for "yes, with minor suggestions." The fix isn't a better question — it's a different contract: the model isn't evaluating a friend's work anymore, it's being graded on flaws found. Change the incentives and the honesty appears.
Should I use a reasoning model to check AI output?
For anything consequential, yes. Reasoning models — like the premium ones in Uzu — deliberate step by step before answering, which is precisely the skill flaw-hunting requires. The efficient pattern is division of labor: a fast model drafts cheaply, then the reasoning model spends its compute only on the audit pass. Save the heavy model for the heavy thinking.
Next Lesson
The Auditor runs one powerful pass: draft, attack, fix. But some jobs are too big for even a perfect single answer — research then transform then polish, each step feeding the next. What happens when you stop writing prompts and start building assembly lines?
Continue to the next lesson: Prompt Chaining
Practice drill before you go: take the last AI output you actually used and run the exact audit prompt above on it — ruthless role, exactly 3 flaws, findings only. Whatever surfaces, sit with the fact that it was sitting there the whole time, waiting for someone to be paid to find it.