How to Stop ChatGPT From Rewriting Your Text When You Only Want Grammar Fixes

The short answer

Do not ask ChatGPT to 'proofread' or 'fix the grammar'. Both are open-ended instructions that the model interprets as a licence to improve the writing. Instead, give it a constrained prompt that explicitly forbids rewriting, limits it to the fewest possible changes, names the categories it must not touch, and specifies an output format with no commentary.

The prompt below is the one we use. It is followed by the results of running it against three different kinds of text, including the edits it made that we rejected.

Why ChatGPT rewrites when you asked for corrections

It is not a bug and it is not ignoring you. Three things combine.

1. The word 'proofread' is genuinely ambiguous. In publishing it means a narrow mechanical pass. In ordinary usage it means 'make this better'. The model picks the broader reading because that is more often what people want.
2. Chat models are tuned to be helpful, and helpfulness is rewarded during training with outputs that look more polished than the input. Returning your text almost unchanged feels like a weak response, so the model is pulled toward making visible changes.
3. An instruction like 'fix the grammar' has no stopping condition. Once the grammar is fixed, the model has budget left and spends it on style, concision and tone, because nothing told it to stop.

The fix is to give it a stopping condition. That is what rule 2 in the prompt below does, and it is the single most important line.

The minimal-edit prompt

Copy this. Replace the placeholder at the bottom with your text.

Correct only the grammar, spelling and punctuation in the text below.

Rules:
1. Do not rewrite, rephrase, reorder, shorten or combine any sentence
   that is already grammatically correct.
2. Change as few words as possible. If a sentence contains no error,
   reproduce it exactly as written.
3. Never change: proper nouns, personal names, numbers, dates, times,
   units, URLs, filenames, code, job titles, or technical terms.
4. Never change the level of certainty. Keep words like "might",
   "could", "may", "should", "I think", "probably" and "maybe"
   exactly where they are.
5. Never change the level of formality. Keep contractions, casual
   openings, sentence fragments and informal wording if that is how
   the original is written.
6. Do not add information, examples, greetings, sign-offs or
   conclusions that are not in the original.
7. Output the corrected text only. No preamble, no summary, no list
   of changes, no offer of further help.

Text:
<paste your text here>

Three notes on why it is built this way. Rule 2 gives the model a stopping condition, which is the thing 'fix the grammar' lacks. Rules 3 to 5 are the categories where corrections most often turn into meaning changes, and naming them explicitly works better than a general instruction to preserve tone. Rule 7 matters because a model that explains its changes is a model that has used output budget on something other than the text, and the explanation is where unwanted edits get justified rather than reported.

Test 1: a casual message

hey sara, sorry for the late reply, i was in back to back meetings all day. i think we could push the launch by a week? the staging env still have that flaky test and id rather not ship on a friday. let me know what u think

Test 2: a professional email

Hi Daniel,

Thanks for sending the contract over on Tuesday. I've reviewed section 4 and 5 with our legal team and we have two concerns.

The payment terms says net 45 but we agreed net 30 during the call on 12 September. Could you confirm which one is correct? Also the liability cap in section 5.2 is set at 6 months of fees, were under the impression it would be 12.

I'm not trying to reopen the negotiation, we just need these two point clarified before I can get it signed. If its easier to jump on a call I'm free Thursday after 2pm.

Best,
Amira

The two unwanted edits are worth dwelling on, because they are the kind you accept without noticing. Expanding 'I'm' to 'I am' in the sentence where you are defusing tension makes the email colder at precisely the moment it needs to be warm. That is not a grammar improvement. It is a tone change wearing grammar's clothes.

Test 3: a technical paragraph

The retry logic in fetchUser() currently swallow errors from the upstream service, which means a 503 is treated the same as a 404. We should seperate the two: retry on 5xx with exponential backoff capped at 3 attempt, and fail fast on 4xx. The withRetry helper already support a predicate option so this should be a small change, but I haven't tested it against the staging config yet so the backoff values might need tuning.

How to audit the output in fifteen seconds

The unwanted edits in these tests were all the same shape. Once you know what to look for, checking is fast.

1. Diff by eye on length. If the output is noticeably longer than the input, the model added something. Rule 6 forbids that, and it is the easiest check.
2. Look for contractions that expanded. 'I'm' to 'I am', 'don't' to 'do not', 'we'll' to 'we will'. Each one raises formality by a step, and none of them is a grammar fix.
3. Look for hedges that moved or doubled. 'could' to 'could consider', 'might' to 'may', 'I think' deleted. These change what you are committing to.
4. Check every number, name and date against the original. Do not skim this. It is the one category where an error is expensive rather than merely annoying.
5. Read the output aloud. If it does not sound like something you would say, put the sentence back.

Step 2 and step 3 catch most of it. In the three tests above, every unwanted edit was either an expanded contraction or a shifted hedge.

When to stop using ChatGPT for this

A chat model is a reasonable proofreader for a long document you are going to read carefully anyway. It is a poor fit for short everyday text, for four reasons.

The setup row is the one people underestimate. Pasting a message out of a composer, into a chat, then back again takes longer than correcting the message did. For a two-sentence reply, the round trip is the whole cost.

Where this breaks down

Frequently asked questions