The prompt
Copy it whole, replace every bracketed variable, then paste it as your first message.
The text below is mine. Edit it for one named problem, as an editor would, not as a ghostwriter. I direct, you propose, I judge. ## 1. The problem to fix [THE PROBLEM] That is the only problem you are fixing. Not taste, not tone, not anything else you notice along the way. ## 2. My text [MY TEXT] ## 3. What must survive - My stance. Do not soften claims I made plainly or harden claims I made carefully. - My rhythm. Keep the mix of long and short sentences, the fragments, and the places where I repeat a word on purpose. - My vocabulary. Do not swap my words for synonyms you prefer. If I wrote "cheap", do not write "affordable". - My specific details. Every number, name, example, and image stays. Do not generalize any of them. - These phrases, exactly as written: [MUST KEEP] How I write, for reference: [VOICE NOTES] ## 4. How much may change [LIMIT] ## 5. Output format Show the work as pairs. For each sentence or short run of sentences you change: OLD: the original, exactly NEW: your version Why: what this fixes, in ten words or fewer Do not show sentences you did not change. Do not return a clean copy; I will assemble it myself. After the pairs, one line: how many sentences changed out of how many, and the new word count. ## 6. Rules - Add no new claims, facts, examples, or sentences that were not there in some form. A transition that asserts something counts as a new claim. If the problem cannot be fixed without adding content, write [VERIFY: what is missing] and stop. - Do not add a topic sentence, a summary sentence, or a tidy close. - If a sentence is awkward on purpose, leave it and list it as "No change: reads as deliberate" so I know you saw it. - Never say you have checked a fact. You have not. - No preamble, no praise, no commentary beyond the pairs and the count.
- Best for
- A rough paragraph of yours that is right in substance and wrong in one way: too long, muddled, choppy, lead buried in sentence four
- Writers whose drafts come back from AI 'improvement' sounding like a press release
- Editing an AI draft you have already rewritten by hand, where the voice is now yours and must stay that way
- Writers working in a second language who want grammar and clarity fixed without their phrasing replaced
- Not for
- Text you have not written and have no stance on. Preserving a voice requires one to preserve
- Changing what the paragraph says. If the claim is wrong, rewrite it by hand. The prompt forbids new claims
- Whole-piece restructuring. Flow across sections is an outline problem, not a paragraph problem
Ask an assistant to 'improve' a paragraph you wrote and you get back a paragraph somebody else wrote. The grammar is fixed. So is the rhythm you chose, the word you repeated on purpose, and the blunt sentence that was doing the work. It reads well, and it is no longer yours.
The answer is not to stop rewriting with AI. It is to hire the assistant as an editor with one job and a strict contract: fix the problem I name, keep everything else, and show me every change so I can refuse it. An editor who works this way is useful. A ghostwriter who works any other way is a tax on your voice.
This prompt sets that contract. You name one problem. You list what must survive: stance, rhythm, vocabulary, specific details, and any phrases you want untouched. You cap how much may change. The assistant returns OLD/NEW pairs with a few words on each, never a clean copy, and it is forbidden from adding anything that was not already there. You assemble the result yourself, which is the point: your hand is the last one on it.
Variables
What to put in each bracket. Required ones are marked. Thin answers here are the usual cause of generic output.
[THE PROBLEM]RequiredOne problem, stated so that you could test whether it is fixed. Too long for the slot, lead buried, second half confusing. Not 'make it better'.
ExampleIt is 190 words and must be 120 or fewer. Both numbers and the ratio between them stay.
[MY TEXT]RequiredThe paragraph or short run of paragraphs, pasted exactly as you wrote them. Typos and all.
ExampleThe paragraph about salting pasta water, pasted exactly as I wrote it
[MUST KEEP]Phrases, sentences, or details that must appear exactly as written.
Example'like the sea' in quotation marks; 'ten grams per liter'; the last sentence exactly as written
[VOICE NOTES]RequiredHow you write, as behaviors. A Voice Sheet, or a few lines describing what your writing does.
ExampleShort sentences. I address the reader as 'you'. I say 'wrong' when I mean wrong. I repeat a key word rather than reach for a synonym. I never open with a question
[LIMIT]How much may change, as something you can count.
ExampleChange no more than half the sentences. Cutting a sentence counts as changing it.
Why it is built this way
Each block does one job. Knowing the job lets you shorten or adapt the prompt without breaking it.
'Mine' and 'editor, not ghostwriter' in the first line
Two words that change the job. An editor fixes what the writer asks and leaves the rest. A ghostwriter replaces. The assistant's default is the second, and it has to be told which one it is.
Block 1: one named problem, and 'the only problem'
Target one decision, from THREAD. The second sentence is what stops the fix from spreading. Without it the assistant fixes your problem and three it noticed, and the three are where your voice goes.
Block 3: the survival list
A positive description of what preserving a voice means, in terms an assistant can follow: stance, rhythm, vocabulary, details. The 'cheap' versus 'affordable' example does more work than the abstract rule, because synonym drift is the most common way a paragraph stops being yours.
Block 4: the cap
A bounded change you can count. If more sentences moved than the cap allowed, the turn failed regardless of how good the result reads.
Block 5: OLD/NEW/Why with no clean copy
Examine the difference without hunting for it. The Why line tells you what the assistant believes it fixed, so you can check whether it did. No clean copy keeps assembly in your hands and removes the place where silent extra changes hide.
Block 6: no new claims, and 'reads as deliberate'
The first stops the most common failure, a smoothing transition that asserts something you did not say. The second gives the assistant a way to leave a rough sentence alone and tell you it did, instead of quietly fixing it.
Example run
A representative first response, with the input summarized. Your results will differ by model and material.
A food writer's 190-word paragraph on why she stopped telling readers to salt pasta water 'like the sea', for a 120-word slot in a printed cookbook. Must keep 'like the sea', 'ten grams per liter', and the final sentence. Limit: no more than half the sentences.
OLD: For years I told people to salt their pasta water "like the sea," because that is what I was told, and because it sounds like the kind of thing a person who knows what they are doing would say. NEW: For years I told people to salt their pasta water "like the sea," because that is what I was told, and because it sounds like something a person who knows what they are doing would say. Why: cut four words, same shape. OLD: The sea is about thirty-five grams of salt per liter. Nobody wants pasta cooked in that. I tried it once, on purpose, to see, and it was inedible, and I had to make the whole thing again from the start with a new pot of water. NEW: The sea is about thirty-five grams of salt per liter. Nobody wants pasta cooked in that. I tried it once, on purpose, and had to start over with a new pot of water. Why: cut the retelling; kept the cost. No change: "Wrong. Not a little wrong." Reads as deliberate. OLD: What you actually want is somewhere around ten grams per liter, which is roughly a tablespoon for a big pot, and even that is more than most people are used to putting in. NEW: What you actually want is ten grams per liter, roughly a tablespoon for a big pot, and even that is more than most people are used to putting in. Why: removed hedges around a number you gave exactly. 3 of 9 sentences changed. 121 words.
The pairs made the judgment fast. She accepted the first two. She rejected the third: 'somewhere around' was deliberate, because she does not weigh it and did not want to sound as if she did. The Why line made the change easy to spot, which is what it is for. With her original third sentence restored she was at 131 words and cut one clause herself. The 'No change' line told her the assistant had seen the two fragments and left them, which is the behavior the prompt is buying.
Follow-up prompts
The second turn is where vibe writing happens. These are the short messages that keep the session moving.
When you have judged the pairs
Pair 3 rejected: 'somewhere around' is deliberate. Pairs 1 and 2 accepted. Give me the word count with my original third sentence, nothing else.
When a pair changed your meaning
Pair 2 changed the meaning: 'start over' on its own hides the cost. Same sentence, same cut, keep the cost. OLD and NEW.
When vocabulary crept in
Pair 4 swapped 'putting in' for 'adding'. That was not the problem. Revert that word, keep the rest of the NEW.
When you have the next paragraph
Same problem, same rules, next paragraph: [PASTE PARAGRAPH].
When you want a second opinion on rhythm, not a fix
Read my original and your pairs together. Which NEW line breaks the rhythm of the paragraph? One line. No fix.
When it goes wrong
Symptoms you will actually see, the usual cause, and the correction that works.
It returns a clean rewritten paragraph with no pairs.
- Cause
- The word 'rewrite' triggers the whole-text habit, and the format rule gets dropped.
- Fix
- Reply: 'Section 5. Pairs only. No clean copy.' Paste the original again so the pairs are built from your text, not from its rewrite.
Synonyms crept in: 'affordable', 'utilize', 'commence'.
- Cause
- The assistant's register pulls toward the formal, one word at a time.
- Fix
- Reply: 'List every word in NEW that is not in OLD. Revert any that is a synonym rather than part of the fix.'
It fixed problems you did not name, and the cap is exceeded.
- Cause
- An editor's reflex, and 'while I'm here'.
- Fix
- Reply: 'Only [THE PROBLEM]. Six of nine sentences is over the limit. Redo with four or fewer.' Count again.
A new transition asserts something ('Which is why most cooks get this wrong').
- Cause
- Smoothing over a cut. The assistant needed a bridge and built one out of a claim.
- Fix
- Delete it and reply: 'No new claims. The transition in pair 2 asserts something I did not say.' Bridges can be punctuation.
Every sentence is now medium length.
- Cause
- The default rhythm. Fragments and long sentences both read as errors to an assistant unless it is told they are not.
- Fix
- Reply: 'My paragraph had two fragments and one long sentence. Your NEW has neither. Restore the fragments and show me which sentence is the long one.'
Before you publish
The prompt cannot do these. You can.
- Read each pair for meaning before style. The Why line says what the assistant thinks it did. Check what it actually did.
- Assemble the paragraph yourself. Pasting NEW over OLD by hand is where you catch the pair you should have rejected.
- Count. If more sentences changed than your limit allowed, the turn failed, however good the result sounds.
- Read the assembled paragraph aloud next to the original. If the music changed, find the pair that changed it and reject it.
- Keep a master copy outside the chat. The assistant's memory of your text drifts across turns, but yours should not.
Questions
What counts as 'one problem'?
Something you could state in one sentence with a test attached: too long for the slot, the lead is in sentence four, the second half is confusing. 'Make it better' is not one problem. It is an invitation to replace everything. If you have two problems, run the prompt twice.
Why not just ask for the clean version and read it?
Because the clean version hides what changed, and the changes you do not notice are the ones that take your voice. Pairs make each change a decision. Assembly takes a minute, and that minute is when you notice the synonym swap.
Does this work on a draft the assistant wrote?
Only once you have rewritten it enough that it carries your voice. At that point it is yours. If it is still the assistant's draft, there is no voice to preserve, and you want the Write First Draft prompt with better voice notes instead.
What if the problem really does need new content?
Then the prompt stops with a [VERIFY: what is missing], which is the right outcome. You add the content by hand, from your own knowledge, and run the prompt again on the result.