AI can fix a sentence and still make it feel like it belongs to someone else. The prompt has to be clear about the job: clean up the text, keep the point, and leave the writer's voice intact.
This matters most in small everyday writing. An email, a Slack reply, a LinkedIn post, a support note. You usually do not need a new draft. You need your draft with fewer rough edges.
The prompts below work in ChatGPT, Claude, Gemini, or any similar writing assistant. They are not magic. They are just specific about what can change and what cannot.
Start with one rule
Before you paste a draft into AI, decide whether you want an edit or a rewrite.
If you ask for a rewrite, the model is allowed to replace your structure, soften your opinion, change your rhythm, and add words you never would have used. Sometimes that is useful. Most of the time, it is too much.
If you ask for an edit, you can keep control. A good voice-preserving prompt has four parts:
- the exact text to fix
- what should change
- what must stay the same
- what the model should return
That is the difference between "make this better" and a useful editing request.
1. The light cleanup prompt
Use this when the text is mostly right, but needs grammar, punctuation, or clarity fixes.
Fix grammar, punctuation, and clarity in the text below.
Keep my wording and sentence structure as much as possible.
Do not add new ideas.
Do not make it more formal.
Return only the fixed text.
Text:
[paste your text]This prompt works because it narrows the job. The model is not being asked to impress you. It is being asked to repair the draft.
2. The "do not rewrite me" prompt
Use this when the idea is personal, opinionated, or written in a voice you want to keep.
Edit this text without rewriting it from scratch.
Keep my opinion, rhythm, and level of directness.
Change only what is needed for clarity.
If a sentence is already clear, leave it alone.
Return one version only.
Text:
[paste your text]The line "If a sentence is already clear, leave it alone" does a lot of work. It tells the model that unchanged text can be a correct answer.
If this is the main problem you keep running into, read fix, do not rewrite. It is the core editing idea behind Prose.
3. The voice sample prompt
Use this when you want the model to learn from a short sample before editing the real text.
Use the writing samples below to understand my voice.
Notice sentence length, tone, word choice, and how direct I am.
Then edit the draft after the samples.
Keep the draft's meaning. Do not copy phrases from the samples unless they fit naturally.
Writing samples:
[paste 2 or 3 short samples]
Draft to edit:
[paste your draft]Samples help because "sound like me" is too vague. A few real paragraphs give the model something concrete to match.
Keep the samples short. If you paste a giant archive of old writing, the model may focus on details that do not matter for this edit.
4. The professional but not weird prompt
Use this for work messages that need to sound cleaner without turning into corporate mush.
Make this message clear and professional without making it stiff.
Keep it concise.
Keep the ask exactly the same.
Do not add apologies, enthusiasm, or extra context.
Avoid corporate phrases.
Message:
[paste your message]This is useful for emails, customer replies, and internal updates. The phrase "professional without making it stiff" is still a little subjective, so the next lines make it concrete.
5. The shorter but same meaning prompt
Use this when the draft is too long, but you do not want the model to change the point.
Shorten this text by about 30%.
Keep the same meaning, facts, names, and tone.
Cut repetition first.
Do not replace specific details with vague summaries.
Return only the shorter version.
Text:
[paste your text]The important line is "Cut repetition first." Without it, the model may remove the specific detail that made the text worth reading.
6. The explain-the-edits prompt
Use this when the text matters and you want to see what changed before accepting it.
Edit the text for grammar and clarity.
Then list the changes you made in plain English.
Separate the answer into:
1. Edited text
2. What changed
Do not change the meaning or add new facts.
Text:
[paste your text]This is slower, but it gives you a review step. If the model changed the meaning, you can catch it before you paste the result anywhere public.
For quick text fixes, Prose keeps this simpler by fixing the selected text where it already lives. For important drafts, asking for an explanation can still be worth the extra step.
7. The model-specific prompt
ChatGPT, Claude, and Gemini can all handle voice-preserving edits, but they do not always need the same nudge.
| Tool | Prompt line to add | Why it helps |
|---|---|---|
| ChatGPT | Before editing, identify what should stay unchanged. | It can reduce broad rewrites by making the model notice constraints first. |
| Claude | Preserve my natural phrasing unless it is unclear or grammatically wrong. | It reinforces the boundary between polish and replacement. |
| Gemini | Return the result in the same structure and approximate length as the original. | It helps when formatting and length matter as much as tone. |
You can add those lines to any of the prompts above. The point is not to treat each model like a different species. The point is to give it the strongest constraint for the mistake you keep seeing.
A full prompt you can save
If you only want one reusable version, use this:
You are editing my writing, not replacing it.
Task: Fix grammar, punctuation, clarity, and awkward phrasing.
Keep: my meaning, tone, sentence rhythm, specific details, names, and formatting.
Avoid: new facts, extra enthusiasm, corporate phrasing, generic AI wording, and a full rewrite.
Rule: If a sentence is already clear, leave it close to the original.
Return: only the edited text.
Text:
[paste your text]That prompt is boring in the right way. It tells the model exactly what editing means.
Where people usually go wrong
Most bad results come from vague prompts. "Make it better" gives the model permission to decide what better means. It may choose smoother, longer, warmer, more formal, or more generic.
That is why people can spot AI-written prompt packs so quickly. In one recent Reddit thread about business-writing prompts, the top replies were blunt: "Written with AI..." and "This doesn't look like that it killed procrastination." The problem was not only the topic. The writing looked detached from the person posting it.
A voice-preserving prompt should protect the parts that make the text yours:
- your level of directness
- your usual sentence length
- your opinion
- your specific details
- your formatting
If you do not name those things, the model may treat them as problems to smooth away.
When prompting is overkill
Long prompts are useful when the text is sensitive, long, or high stakes. They are annoying when the job is a one-sentence fix.
That is the gap Prose is built for. Select text in Chrome, choose a fix, and keep working in the same place. No new tab, no prompt setup, no copy-paste loop.
For a bigger draft, use the prompts above. For a small fix, the best prompt is often no prompt at all.
If you want the privacy side of this workflow, read where your text goes when AI fixes it. If you want to get better at giving AI context, read context engineering examples for better AI prompts.
