By tone
Tone prompts.
- Natural but faithful: “Rewrite this in natural English. Keep every factual claim, number and caveat. Remove generic AI phrasing and do not add new examples. Text: [paste]”
- Professional, not stiff: “Make this sound like a competent person wrote it for [audience]. Avoid hype, clichés and fake certainty. Keep contractions where they sound normal. Text: [paste]”
- Warmer without padding: “Rewrite this to sound warmer, but do not make it longer. Keep the request and all facts unchanged. Text: [paste]”
By length
Length prompts.
- Shorter version: “Rewrite for brevity. Keep the core meaning, remove throat-clearing, keep specific details, and limit the result to [x] words.”
- Aggressive cut: “Cut this by half. List what you removed so I can restore anything that mattered. Text: [paste]”
- Expand honestly: “Expand this with structure, not filler: add the missing who, when, and what happens next — but only from the facts I give you. Facts: [list] Text: [paste]”
By audience
Audience prompts.
- Non-expert reader: “Rewrite this for a smart reader with no background in [field]. Replace jargon with plain terms, keep every claim accurate, and flag anywhere the simplification changes the meaning.”
- Executive summary: “Rewrite this for a decision-maker with 30 seconds. Conclusion first, then the two facts that support it, then the ask. Nothing else.”
- Existing customer: “Rewrite this for someone who already uses our product. Remove introductory selling, keep the change, the impact and the action.”
By format
Format prompts.
- Email: “Turn this into an email under 120 words. Request in the first two sentences, one clear next step at the end, no ‘I hope this finds you well’.”
- Bullets to prose: “Turn these bullets into two connected paragraphs. Keep every item, add logical connections, no new claims. Bullets: [paste]”
- Prose to bullets: “Turn this into bullets a reader can scan in 15 seconds. One idea per bullet, keep all numbers, cut everything decorative. Text: [paste]”
Every prompt here ends the same invisible way: if the model changed a claim, you want to know. Add “After the rewrite, list anything you changed that could affect meaning.” to any of them when the text matters.
Anatomy
Why these prompts work: the four-part anatomy.
Every effective rephrase prompt on this page is built from the same four parts, and once you see them you can write your own faster than you can browse a library. Part one is the task frame: rewrite what, into what — “this email into a shorter email”, not “improve this”. Part two is the keep-list: the claims, numbers, names and caveats that must survive verbatim. Part three is the change-list: tone, length, structure, audience — the dimensions the model is allowed to move. Part four is the failure definition: what would make the rewrite wrong, stated explicitly — invented examples, softened warnings, added enthusiasm.
The failure definition is the part almost everyone omits, and it is the highest-leverage sentence in the prompt. Models are trained to be agreeable; without a named failure mode, they optimize for “sounds better”, and “sounds better” quietly votes against caveats, hedges and inconvenient numbers. One sentence — “Adding any fact I did not provide counts as failure” — measurably changes what comes back.
A corollary worth internalizing: the shorter your text, the longer the prompt can afford to be relative to it. Spending 60 words of prompt on a 40-word sentence feels absurd and is completely rational when that sentence is the subject line of a campaign or the first line of a cover letter. Prompt effort should track the stakes of the text, not its length.
Troubleshooting
When the output is wrong, fix the prompt — not the output.
Symptom → cause → the line to add. Editing the model's output by hand is fine once; three times means the prompt is missing a part.
| Symptom | Usual cause | Add this line |
|---|---|---|
| Output is still generic | No audience in the prompt. | “The reader is [specific person/role in specific situation].” |
| Facts got smoothed away | No keep-list. | “These must appear unchanged: [list them].” |
| Invented examples appear | No failure definition. | “Do not add examples, quotes or numbers I did not give you.” |
| Tone overshoots (too casual / too stiff) | Tone described in adjectives. | “Match the tone of this sample: [paste 2–3 sentences you wrote].” |
| Length ignored | Soft limit (“keep it short”). | “Maximum [n] words. Count them.” |
| Same structure every time | Model defaulting to its template. | “Do not open with a summary sentence. Start with [the request / the number / the scene].” |
Advanced
Three prompts for when the basic ones plateau.
- Style transfer from your own writing: “Here are three paragraphs I wrote: [paste]. Rewrite the text below so it sounds like the same person wrote it — match my sentence length, directness and vocabulary, not my topics. Keep all facts unchanged. Text: [paste]”. This beats every adjective-based tone instruction, because your sample defines the register no description can.
- The skeptical-reader pass: “Rewrite this for a reader who is skeptical and busy. They will challenge any claim without evidence, and they stop reading at the first sentence that says nothing. Cut or support accordingly. Text: [paste]”. Useful for proposals and anything going to leadership.
- The meaning-diff request: “Rewrite this per the instructions above, then output a two-column table: every sentence whose meaning shifted, original vs rewrite, with one line on the difference.” This turns the model's own change-tracking into your QA artifact — verify it, but it catches most drift before you do.
Combine them: style transfer for the register, skeptical-reader for the content bar, meaning-diff for the safety net. That three-line stack is a complete rewriting system in itself.
FAQ
Paraphrasing prompt questions.
What makes a paraphrasing prompt good?
It names three things: what must not change (facts, claims, numbers), what should change (tone, length, structure), and what counts as failure (invented examples, softened caveats).
Why do my rephrase prompts return the same generic text?
Because the prompt describes a style instead of a reader. “More natural” means nothing to a model; “for a client who is behind schedule” changes the register automatically.
Should I chain prompts or use one big prompt?
For short texts, one specific prompt is enough. For anything longer than a page, separate meaning-check and tone passes work better — see the three-pass workflow.
Do these prompts work in Claude and Gemini?
Yes. The structure — keep, change, failure conditions — is model-independent. Outputs will differ in style, which is exactly why comparing models is useful.
Related