The Quality Audit Prompt — Get a Better Answer Out of the One You Already Have

This is the prompt we give away at the workshop, and the one alumni ask for most. It is below in full — free, no sign-up, nothing to buy.

Most people meet the limits of AI in the same place. The answer comes back and it is close. Structurally right, roughly what you asked for, and not quite usable — a section thinner than the rest, a requirement you mentioned in the third message quietly dropped, a number that does not survive checking.

The instinct is to start editing it yourself, or to ask for a fix and get a patch that breaks something else.

There is a better move, and it takes one paste. The reason the output is close but not right is almost never the model. It is that nothing ever asked it to check its own work. An AI answers the message in front of it. It does not, unprompted, go back through everything you said, hold the finished thing against all of it, and ask what is missing.

This prompt makes it do exactly that.

What it actually does

Four things, in order:

1. Re-reads the whole conversation — the original request plus every constraint, correction and afterthought you added in later messages. This is the part that matters most. Requirements you added at message six are the ones most likely to have been dropped by message ten.
2. Audits the output element by element against that accumulated set — looking for missing sections, requirements half-met, unstated assumptions, factual and arithmetic errors, code that would not run, internal contradictions, broken references, formatting defects and edge cases nobody covered.
3. Closes every gap it finds.
4. Regenerates the deliverable in full, with the fixes integrated and everything that was already right left alone.

That last constraint is what separates it from asking "can you improve this?" — which reliably returns something shorter, because the model paraphrases rather than rebuilds. This prompt forbids that explicitly: no placeholders, no summaries, no partial regeneration.

1. Paste it in the same chat — this is the one that catches people

The prompt reviews the most recent output you generated in this conversation. That phrase is doing a lot of work.

It means the prompt has to be your next message in the same conversation, after the output you want fixed. Not a new chat. Not tomorrow's chat. If you open a fresh window and paste it, there is nothing behind it to audit — and it will say so plainly rather than invent something, which is correct behaviour and still a wasted paste.

The conversation is not just where the output is. It is where the requirements are — your original brief and the six corrections you made along the way. Start a new chat and you have thrown away the very thing the audit is measured against.

If the work came from somewhere else — a document a colleague wrote, something you produced last week, output from a different tool — paste the document into the chat as one message first, then send this prompt as the message after it. It treats what you pasted as the output under review. Add a line of your own about what it was meant to achieve, and the audit gets considerably sharper.

2. Use the strongest model you have, with thinking turned up

Use Claude Fable 5.1 or GPT-5.6, not the default fast model and not a mini variant. Wherever your interface offers a thinking, reasoning or effort control, set it to the highest option.

This is a harder task than the one that produced the original output. The model has to hold your entire conversation in mind, compare a finished artefact against it point by point, notice what is absent — noticing absence is much harder than noticing error — and then rebuild the whole thing without losing anything that was already correct.

A light model will do something that looks like all that. It will tidy the headings, smooth a few sentences, and hand it back feeling improved. Nothing will have been audited.

3. Research mode: usually no, sometimes yes

Leave Deep Research or Research off for most runs. The audit is about internal consistency against your requirements, and research mode adds latency and burns your allowance without helping.

Turn it on in one case: when the output rests on external facts you need verified — market figures, regulations, prices, a competitor's claims, anything cited. Then the "factual errors" line in the prompt has something to actually check against.

The prompt

Copy all of it. There is nothing to fill in — no placeholders, no brackets, no editing. It works as written on a document, a spreadsheet formula, a proposal, a project plan, a customer email or code.

# Quality Audit & Full Regeneration Prompt

## Role

You are a rigorous quality auditor reviewing the most recent output you generated in this conversation — whatever its form: document, code, copy, analysis, plan, design, or any other deliverable — unless the user designates a different one.

## Task

- First, re-read the original request plus every constraint, correction, or requirement the user added in later turns; audit against this full accumulated set
- Re-examine the entire output, element by element, for gaps of any size — however minor
- Check specifically for:
  - Missing steps, sections, or components
  - Requirements left unmet or partially met
  - Unstated assumptions and logical inconsistencies
  - Factual, technical, or computational errors — including code that would not run as written
  - Ambiguities in sequencing, instructions, or ownership
  - Internal contradictions, redundancies, and broken references
  - Formatting defects and unaddressed edge cases
- Plug every gap identified
- Regenerate the output in full with all fixes integrated

## Constraints

- Preserve everything already correct — regeneration must lose no existing detail
- No placeholders, summaries, or partial regeneration; deliver the complete final version only
- If length limits force a split, continue seamlessly in the next message without omission
- Where a gap cannot be closed from available context, state the assumption inline (code comment, bracketed note) rather than inventing content
- If no prior output exists in the conversation, state that plainly instead of fabricating one to audit
- Output only the regenerated deliverable — no audit notes, gap lists, or commentary

## Tone

Match the original output's style, format, and conventions exactly.

What comes back

The finished thing, and only the finished thing. No audit notes, no list of what it found, no commentary. That is deliberate — the output is meant to be usable the moment it lands, not a report you then have to act on.

If you want to see the reasoning, ask for it afterwards: "what did you change, and why?" Worth doing the first few times, because it teaches you where your own briefs tend to be thin. If the same category of gap keeps appearing — missing success criteria, undefined ownership, no edge cases — that is a pattern in how you write requests, and it is fixable at the source.

If it stops mid-way through a long piece, the prompt instructs it to continue seamlessly in the next message. Reply continue and it picks up exactly where it left off.

Getting more out of it

• Point it at something other than the last message. The prompt audits the most recent output unless you designate a different one. So you can write: "Audit the proposal from earlier in this conversation, not the summary you just wrote."
• Run it twice. The second pass usually catches formatting and edge cases the first was too busy with substance to reach. A third almost never earns its keep.
• Use it on code. "Code that would not run as written" is one of the things it checks for specifically, and it is one of the more reliable catches — a function called with the wrong arity, a variable used before it exists, a dependency never imported.
• Give it the standard to audit against. If the work has to satisfy something external — a tender format, a client's template, a compliance checklist — paste that into the conversation before you run the audit. It becomes part of the accumulated requirements.
• Read the assumptions it states. Where a gap cannot be closed from what it has, the prompt requires it to state the assumption inline rather than invent content. Those inline notes are the most valuable part of the output, because each one marks a place where you need to supply something.

If something looks wrong

"There is no prior output to audit." You are in a new conversation. See section 1 — paste the work in first, then the prompt.

It came back shorter. It has summarised instead of regenerating, which the prompt forbids. Say: "You have shortened it. Regenerate in full, preserving every section of the original, with only the gaps closed." If it happens twice, you are on too light a model.

It changed things that were already fine. Tell it which parts were correct and ask it to redo the regeneration leaving those untouched. The instruction to preserve what is right is in the prompt, but a long document gives it plenty of room to get enthusiastic.

It invented a fact to fill a hole. The prompt explicitly requires an inline assumption instead. Challenge it directly — "where did that figure come from?" — and it will usually retract. If a model does this repeatedly on your work, stop giving it gaps to fill: the fix is more input, not more prompting.

It gave me an audit report, not the deliverable. Reply: "Output only the regenerated deliverable, no commentary."

One caution before you paste

Whatever is in that conversation goes to a third party — and this prompt is designed to be used on your real work, which is exactly when that matters.

Keep customer names, personal data and anything under a client confidentiality clause out of it. Under India's DPDPA 2023 the responsibility sits with you as the business holding the data, not with the AI vendor. A proposal with the client's name replaced by "the client" audits just as well. A spreadsheet of your customers does not belong in a chat window at all.

Prompts like this one are what we build with owners at our two-day workshop — on your own work, on your own numbers, with the output good enough to send before you leave the room. Details: Accelerate Your Profits with AI.