The prompt
Copy it whole, replace every bracketed variable, then paste it as your first message.
This is my last pass before I release the piece. Audit it. Do not edit it. Everything you find, I will fix by hand. I direct, you propose, I judge, and I own what goes out. ## 1. The draft as it stands [THE FINAL DRAFT] ## 2. How I actually write (my own samples) [MY WRITING SAMPLES] ## 3. What I meant (my original notes) [MY ORIGINAL NOTES] ## 4. Already verified; skip these [ALREADY VERIFIED] ## 5. Where it is going [WHERE IT WILL APPEAR] ## 6. Report these five findings, in this order, one line per item - AI tells: sentences or moves that read as generated. Quote each. Look for the reflexive three-item list, the paragraph that opens by restating the previous one, the tidy closing summary, the "not this, but that" reversal, the hedge followed by a confident assertion, the adjective where I would have used a number, and any word that appears nowhere in my samples. - Voice drift: places where the draft sounds unlike my samples. Quote the draft line and name what my samples do differently: sentence length, bluntness, how I open paragraphs, what I never say. - Claims needing a source: anything a reader could ask "how do you know?" about that is not listed in section 4. Quote it. Do not guess whether it is true. - Meaning changed: places where the draft says something my notes did not say, says it more strongly or more weakly, or drops a qualification the notes had. Quote the draft and quote the note. - The one paragraph to rewrite by hand: name it and say in two lines why that one. Choose the paragraph where the most findings overlap, or the one the reader will remember. ## 7. Rules - Do not rewrite, suggest replacement wording, or offer to fix anything. If you catch yourself drafting a better sentence, delete it and describe the problem instead. - Do not report that the piece is good, ready, or close. Report findings. An empty category is reported as "None found", and that is a weak claim, not a clearance. - Never say you verified anything. You have not. - No summary after the five findings. Stop.
- Best for
- Any piece about to be published, sent, or submitted under your name
- Drafts that have been through several AI turns and have lost track of what you originally meant
- Writers with a disclosure obligation, who need to know exactly what the assistant contributed before they describe it
- The step after the fact-check pass, when the claims are settled and the remaining question is whether the piece is yours
- Not for
- A piece you have not read carefully yourself. The audit catches what you cannot see. It does not replace reading
- Drafts still changing in substance. Run it last, once, and then fix by hand
- Writers with no samples of their own writing to compare against. The voice finding needs them. Without them you get the assistant's taste
The last pass is the one people skip, because the piece reads well. Reading well is the problem. A smooth draft is the hardest kind to audit, because every sentence sounds like it belongs, including the four the assistant added, the two where your careful claim became a confident one, and the close that could end any article ever written.
This prompt turns the last pass into an audit with five fixed findings. It compares the draft against two things only you can supply: samples of your own writing, for voice drift, and your original notes, for meaning drift. It lists the AI tells by pattern and the claims still waiting for a source. Then it does one thing no checklist does: it names the single paragraph you should rewrite by hand, and says why that one.
Nothing gets fixed in the chat. That is the rule, and it is the point of the stage. The final edit is the moment you take the piece back. The assistant's findings are a list to work through by hand, and the piece leaves the chat as yours.
Variables
What to put in each bracket. Required ones are marked. Thin answers here are the usual cause of generic output.
[THE FINAL DRAFT]RequiredThe draft as it stands after the fact-check pass, pasted whole.
ExampleThe all-staff memo on the new parental leave policy, as it stands after the fact-check pass
[MY WRITING SAMPLES]RequiredTwo or three pieces you wrote without AI, in the same register if possible. These are the evidence for voice drift.
ExampleThree of my previous all-staff memos, pasted in full. They are blunt, short, and never use the word 'excited'
[MY ORIGINAL NOTES]RequiredThe notes, brief, or outline the piece was built from. These are the evidence for meaning drift.
ExampleThe policy brief from legal and my own notes: 12 weeks paid, starts January 1, applies to adoptive and foster parents, the part-time rule is still being worked out, do not promise a date for it
[ALREADY VERIFIED]Claims you have already checked against sources, so the audit does not list them again.
ExampleThe 12 weeks, the January 1 start, and the adoptive and foster coverage, all checked against the signed policy
[WHERE IT WILL APPEAR]Venue and audience, including whether it will be forwarded or quoted.
ExampleEmail to all 180 employees from me as HR director; it will be forwarded and quoted
Why it is built this way
Each block does one job. Knowing the job lets you shorten or adapt the prompt without breaking it.
'Audit it. Do not edit it.' in the first line
The stage is defined by whose hand makes the fix. The first line sets it. Block 7 repeats it with a self-check ('if you catch yourself drafting a better sentence') because the urge to fix is strongest at the end.
Blocks 2 and 3: your samples and your notes
The two comparison sets. Voice drift without samples is the assistant's taste. Meaning drift without notes is unmeasurable. Both are things only you can supply, which is why this audit cannot be automated away.
Block 4: already verified
Keeps the claims finding from repeating the fact-check pass, and leaves a written record of what you actually checked, which is useful if the piece is ever questioned.
Block 6, first finding: the named tells
A list of patterns rather than words, because the words change with every model and the patterns do not. 'Any word that appears nowhere in my samples' is the single most useful test on the list.
Block 6, fourth finding: meaning changed, with both quotes
Quoting the draft and the note side by side makes the finding checkable and shows the direction of drift: stronger, weaker, or a dropped qualification. That is how AI drift usually happens, one adverb at a time.
Block 6, fifth finding: the one paragraph
Converts a list into a first action. Rewriting one paragraph by hand, before the small fixes, reliably resets your ear for the rest of the piece.
Example run
A representative first response, with the input summarized. Your results will differ by model and material.
An HR director's all-staff memo announcing a parental leave policy at a 180-person company, audited against three of her earlier memos and the policy notes from legal. The 12 weeks, the start date, and the adoptive and foster coverage were already verified. The memo will be forwarded.
AI tells - "This isn't just a policy change. It's a commitment to the people who make this company what it is." The reversal, plus a closing flourish. None of your three memos end with a sentence about the company. - "We're excited to share some important news." Your samples never use "excited", and two of three open with the operative fact. - "Whether you're welcoming a newborn, adopting, or fostering, we've got you covered." Reflexive triplet, and "we've got you covered" appears in no sample. - The final paragraph restates paragraphs two and three in softer words. Voice drift - "We believe that supporting families is one of the most meaningful things we can do as an employer." Your memos state the policy, then the date, then who to ask. They do not explain why the company believes things. - Draft sentences average about 22 words; your samples average about 13. Paragraph four has one sentence of 41 words. - Your samples open with the operative fact. The draft opens with a greeting and a warm-up sentence. Claims needing a source - "most companies our size offer six weeks or less" (paragraph three). Not in section 4, not in the notes. - "the most common question we've heard" (paragraph five). Implies a tally that may not exist. Meaning changed - Draft: "Part-time employees will be covered under a separate schedule we'll announce in the spring." Note: "the part-time rule is still being worked out, do not promise a date for it." The draft promises a date. - Draft: "Leave can be taken any time in the first year." Note: "within 12 months of the birth or placement." Probably the same meaning, but "the first year" could be read as the calendar year. Flagging as a possible change, not a certain one. - The draft omits "foster" in paragraph two and includes it in paragraph four. The note lists both throughout. The one paragraph to rewrite by hand Paragraph five, the part-time paragraph. It carries the only promise you were told not to make, it is the paragraph every part-time employee will forward, and it is the one furthest from your samples in tone.
The meaning-changed finding earned the whole pass. 'We'll announce in the spring' was a promise that did not exist in the notes and would have been quoted back at her in March. The 'most companies our size' claim was the assistant's own addition from an early draft, and nobody had caught it through two rounds of edits. She rewrote paragraph five by hand in four sentences, cut the opening and the close, and sent a memo that read like her other memos. Note the 'possible change, not a certain one' in the second meaning item: the audit is allowed to be unsure, and should say so.
Follow-up prompts
The second turn is where vibe writing happens. These are the short messages that keep the session moving.
When you have rewritten the named paragraph
Paragraph five, rewritten by hand: [PASTE]. Against my samples and my notes only: does it still drift, or change meaning? One line each. No wording.
When you want only the tells list on the edited version
Run only the AI-tells finding again on the full draft after my edits. Quote each, nothing else.
When you want a mechanical check
List every sentence over 25 words with its word count. No commentary.
When one note matters more than the rest
Compare paragraph two against the note on who is covered, word by word. Report differences only.
When you want to know what a regular reader would catch
Which of the voice drift items would a colleague who has read my last ten memos notice first? One line.
When it goes wrong
Symptoms you will actually see, the usual cause, and the correction that works.
It rewrites: 'Here is a tighter version of paragraph five.'
- Cause
- The urge to fix is strongest at the end, when the piece is almost done.
- Fix
- Reply: 'Section 7. Delete the wording. Describe the problem.' If it happens twice, run the prompt alone in a fresh session.
'The piece is in good shape overall' opens or closes the report.
- Cause
- The default shape of feedback.
- Fix
- Reply: 'No verdicts on readiness. Findings only.' Then ignore the sentence. It is not evidence.
The AI-tells list is generic: 'the tone is somewhat formal'.
- Cause
- It stopped quoting, and a finding without a quote is an impression.
- Fix
- Reply: 'Quote each sentence. If you cannot quote it, it is not a finding.'
Meaning drift is missed because the assistant read the draft closely and skimmed the notes.
- Cause
- The draft is smooth and the notes are messy, so the assistant anchors on the smooth one.
- Fix
- Reply: 'Go note by note. For each line in my notes: where is it in the draft, and is it stated the same, stronger, weaker, or missing?'
The audit of a draft the assistant wrote itself finds almost nothing.
- Cause
- Self-review by the same session is weak evidence. It is grading its own work.
- Fix
- New session, the prompt alone, and paste the draft as 'the draft', not 'your draft'. Compare the two reports. The second is usually longer.
Before you publish
The prompt cannot do these. You can.
- Fix everything by hand. If you find yourself pasting the assistant's description back and asking for wording, you have left the stage.
- Rewrite the named paragraph first, before the smaller fixes. It resets your ear for the rest.
- 'None found' in any category is weak evidence, not clearance. Read that category yourself with the pattern list in hand.
- Everything under 'claims needing a source' goes back to the fact-check pass or gets cut. Nothing unsourced ships as fact.
- Decide on disclosure now, while you can still say what the assistant did. The ethics lesson covers what to say and where.
- Read the final version aloud once, start to end, after every fix. Then release it, and keep the audit with your notes in case of corrections.
Questions
Why no rewriting at the very end, when it would be fastest?
Because the final edit is where ownership becomes literal, and because the assistant's fixes reintroduce the same tells the audit just found. Rewriting one paragraph by hand takes ten minutes and changes how you read the rest. Letting the assistant do it takes one minute and changes nothing.
What if I have no writing samples?
Use anything you wrote without AI that you would stand behind: emails, old posts, a report. Three short ones are enough. With none at all, skip the voice finding and run the other four, then build a Voice Sheet before the next piece.
Is this the same as the 'final check' in the mega prompt?
Related, but not the same. The final check is a quick in-session audit from the assistant's own view of its draft. This is a standalone audit in a fresh session, with your samples and notes as evidence, and it adds the paragraph to rewrite. Use this one when the piece matters.
How do I decide what to disclose?
The meaning-changed and AI-tells findings tell you what the assistant actually contributed, which is the fact you need before you can describe it. The ethics lesson covers the policy questions: who requires disclosure, in what words, and where. Keep the audit with your notes either way.