Rewriting with AI usually kills voice. There is now peer-reviewable evidence for why: an April 2026 study, “Voice Under Revision,” ran 300 personal narratives through three frontier models (GPT-5.4, Claude Sonnet 4.6, Gemini 3.1 Pro) and measured 13 stylistic markers. Every model, in every condition, drifted the same way: fewer function words, fewer contractions, fewer first-person pronouns, longer words, fancier punctuation. The kicker: an explicit “preserve the author’s voice” instruction shrank the drift but never reversed its direction (van Nuenen, arXiv 2604.22142).
So “preserve voice” alone is not a fix. The prompts below add the missing scaffolding: one rewrite goal per pass, an explicit voice spec, hard preservation rules, and a forced change log. Before you rewrite at all, it usually pays to run AI writing feedback on the draft first; a senior-editor critique tells you which problems are worth a rewrite and which are not, so you don’t burn voice fixing things that did not need fixing.
TL;DR
- AI rewrites systematically normalize prose toward “AI-grey” (drop contractions and first-person, lengthen words). “Preserve voice” alone only dampens this; it does not stop it.
- Use one goal per prompt. Bundling tighten + restructure + translate + SEO into one pass produces a worse version of all four.
- Pin a 3-4 rule voice spec and demand a change log every time, so you can audit drift instead of trusting the model’s “I kept your voice” claim.
- Model pick (June 2026): GPT-5.5 for the most natural-sounding prose; Claude Opus 4.7 when no fact, number, or quote may shift; Gemini 3.1 Pro when the source is huge and you need the full 1M-token window.
- The single biggest AI tell in 2026 is cadence uniformity, sentences all landing in the 18-24 word band. Put a sentence-length-variety rule in your voice spec.
What these prompts solve
Article rewrites fail when they bundle five goals into one prompt: tighten + restructure + add depth + translate + SEO. Each goal flattens voice a little; together they erase it. These templates isolate one rewrite goal per prompt and force the model to declare what it changed.
Which model to use (June 2026)
The right model depends on what you are protecting in the rewrite:
| Rewrite goal | Best pick | Why |
|---|---|---|
| Natural-sounding prose, tone shifts, conversational rewrites | GPT-5.5 | Highest fluency; adapts tightly to tone instructions |
| Fact-preserving rewrites (numbers, quotes, legal/technical) | Claude Opus 4.7 | More cautious; least likely to silently change a claim |
| Huge source articles (long reports, whole archives) | Gemini 3.1 Pro | Full 1M-token context, lowest API price ($2/$12 per 1M tokens) |
| Free / casual one-offs | ChatGPT Free (GPT-5.5) | Works, but tight limits and ads on US Free since Feb 2026 |
All three carry a 1M-token context window as of June 2026, but in-app limits differ: ChatGPT Plus ($20) caps usable context at roughly 320 pages, with the full 1M only on the $200 Pro tier. For most article rewrites that is plenty; the cap only bites when you paste a whole archive.
Who this is for
Content editors refreshing a 3-year-old archive, indie creators porting blog posts to a newsletter or another language, founders rewriting AI-drafted copy for human voice, technical writers turning internal docs into public posts, anyone whose drafts come back from AI sounding identical to every other AI-edited piece.
When not to use these prompts
Skip them on first drafts — rewrite prompts assume there’s an article worth editing. Skip them if the original has a fact problem; rewrite first, fact-check second is backwards. And don’t run multiple rewrite prompts in one chat without resetting context; each rewrite eats voice signal from the prior turn.
Prompt anatomy / structure formula
A rewrite prompt should always carry six elements:
- Source article: pasted in full, or a clear pointer.
- One rewrite goal: tighten OR restructure OR translate — never bundled.
- Voice spec: humor level, sentence length, signature phrases, banned phrases.
- Preservation rules: which sentences must survive; which facts cannot change.
- Change log: model must list what it changed, not just hand back the new text.
- Stop conditions: “do not add new claims”, “do not introduce new sources”, “no new examples unless requested”.
Best for
- Old blog posts needing a content refresh
- Translated articles needing real localization (not literal translation)
- Voice-preserving length cuts (1,200 → 800 words)
- Restructuring to inverted-pyramid for SEO
- Stripping AI-grey phrasing from AI-drafted copy
- Reading-level adjustments (Grade 12 → Grade 8)
- Adding honest counter-arguments to one-sided posts
17 copy-ready prompt templates
1. Tighten without changing voice
Below is a 1,200-word article. Cut to 800 words while preserving the author's voice (track: humor level, sentence length, signature phrases). Output (a) the rewrite, (b) a list of what you cut and why. Do not introduce any new claims.
{paste article}2. Add depth without bloat
Take this article and add one concrete example per claim. Do not lengthen the total by more than 30%. Keep all original sentences unless they directly conflict. List the examples you added with one-line justifications.
{paste}3. Restructure to inverted-pyramid
Restructure this article into inverted-pyramid format: lead with the strongest conclusion, then evidence, then nuance. Preserve all sentences but reorder. Mark with [MOVED] any sentence whose position changed.
{paste}4. Localize for a new audience
This article was written for {original audience}. Rewrite for {new audience}, swapping examples, references, and vocabulary. Keep the structural argument intact. Flag any cultural references that don't translate.
{paste}5. Strip AI-grey phrasing
Below is an AI-generated draft. Strip generic AI phrasing ("in today's fast-paced world", "it's important to note", "moreover", "delve into", "ultimately", "robust", "leverage", "navigate the landscape"). Replace with concrete, specific phrasing. Output diff list: what you replaced and why.
{paste}6. Lower reading level
Rewrite for a Grade 8 reading level: shorter sentences, simpler words, no jargon without definition. Preserve all facts. After the rewrite, list jargon you replaced and the simpler word used.
{paste}7. Translate without flattening
Translate this English article to Simplified Chinese for Xiaohongshu / Zhihu readers. Preserve voice and rhythm — don't produce machine-literal translation. After the translation, list 5 idioms you adapted rather than literally translated, and 3 phrases you deliberately kept in English because translation would lose meaning.
{paste}For the self-critique loop that catches where this translate pass quietly drifts on tone or culture, run the output through our AI translation self-critique workflow before shipping.
8. Add SEO without keyword stuffing
Optimize this article for the keyword "{keyword}" without stuffing. Add at most 3 natural mentions in body, 1 in H1, 1 in first 100 words. Suggest 5 semantically related terms to weave in. Keep voice intact. Output the diff only.
{paste}9. Add an honest counter-argument
This article argues one side strongly. Add a 150-word section "When this advice doesn't apply" between the conclusion and FAQ. Be specific about edge cases; don't weaken the main argument.
{paste}10. Refresh for current year
This article is from {year}. Update for 2026: refresh data points, tool versions, dead links, and outdated screenshot references. Mark each change with [UPDATED]. Do not rewrite untouched paragraphs.
{paste}11. Match a specific author’s voice
Below is a target voice sample (200 words by author {name}). Below that is my article. Rewrite my article in that voice: rhythm, sentence length distribution, lexical choice. Preserve the factual content exactly. Output the rewrite, then a 3-line note on what voice features you matched.
Voice sample: {paste}
Article: {paste}12. Make it more professional (without stiffening)
Rewrite this article more professionally for a {industry/role} audience. Cut casual asides and emoji. Keep one signature device (e.g., short opening line, em-dash rhythm) so it still sounds human. List what you removed and what you preserved.
{paste}13. Make it more conversational (without dumbing down)
Rewrite this article in a more conversational voice. Use contractions, second person ("you"), and at least 3 short questions to the reader. Do not lose technical precision. Note which sentences you reworded and which you kept verbatim.
{paste}14. Multi-language adaptation (not just translation)
This article was written for {source language / culture}. Produce a {target language} version that an editor in the target market would publish — re-pick analogies, swap region-specific tools, recalibrate examples for the target audience's expectations. Output: the adapted article, then a "what I localized vs translated" list.
{paste}15. Fact-preserving rewrite
Rewrite this article in {voice spec}. Hard constraint: every numerical claim, name, date, and URL must survive unchanged. After the rewrite, output a checklist verifying each numerical claim appears in the new version with the same value.
{paste}16. De-clickbait the title and lead
The title and first paragraph below feel like clickbait. Rewrite both to make the same claim without the clickbait pattern (no "you won't believe", no vague numbers, no fake urgency). Title ≤12 words. Lead ≤80 words. Keep all factual claims.
Title + lead: {paste}17. Repurpose for a different format
Below is a 1,500-word blog post. Produce a {target format} version: {options: Twitter thread (10 tweets) / LinkedIn post (250 words) / podcast outline (10 beats) / 60-sec script}. Keep the strongest 2 ideas; drop the rest ruthlessly. Output the new artifact only.
{paste}Common mistakes
- Not pasting the original article. The model invents a plausible one instead of editing yours.
- Asking for “improvement” without specifying what. Voice? Structure? SEO? Length? Vague intent → generic output.
- Letting AI add new claims without flagging them. Always require a change log so you can spot hallucinated facts.
- Stacking goals into one prompt. “Tighten + translate + SEO” produces a worse version of all three. One goal per pass.
- Skipping the voice spec. Without 3–4 explicit voice rules, every rewrite drifts to neutral AI-grey.
- Trusting “I preserved voice” claims. Spot-check sentence-length distribution and signature phrases by hand for the first few rewrites.
- Running multiple rewrites in the same chat without resetting. Voice signal degrades each turn; start a fresh chat per article.
How to push results further
- Pin a voice spec block at the top of every rewrite: sentence-length range (not a single target), banned phrases, allowed devices. Compose it once per author and reuse.
- Put sentence-length variety in that spec explicitly. Cadence uniformity, sentences all landing in the 18-24 word band, is the single biggest AI tell in 2026, and it survives most “preserve voice” prompts. Ask for a deliberate mix of short and long.
- Demand a change log in every prompt. The “Voice Under Revision” study showed models normalize even when told not to, so an “I preserved your voice” claim is not evidence; the change log is.
- For translations, prefer the localize template (#4 / #14) over the translate template (#7) when the target audience is clearly different from the source.
- For length cuts, ask the model to output a first pass at the target length AND a second pass 10% shorter. Comparing forces ruthlessness; you keep whichever lands.
- After each rewrite, run a 5-line diff audit: “list every place the meaning could have shifted, even slightly.” This catches sneaky semantic drift.
- Pair rewrites with a tone rewrite prompt only if voice was the goal; otherwise you double-flatten the piece.
- For refresh-for-current-year rewrites, ask the model to list every fact that needs verification before doing the rewrite. Saves you a fact-check pass later.
FAQ
- Why does AI keep adding phrases like “in conclusion” no matter what I say? Constraint drift in long prompts. Add the banned phrases at the END of the prompt too; recency in the prompt matters for the model’s filtering pass. GPT-5.5 suppresses em-dashes and stock connectors better than older versions, but it still slips them back in on long rewrites.
- Does a “preserve voice” instruction actually work? Partly. The 2026 “Voice Under Revision” study found that instruction reduces stylistic drift but never reverses its direction; rewrites still trend toward fewer contractions and first-person and longer words. Treat the instruction as a dampener, not a guarantee, and back it with a voice sample plus a change log.
- How do I keep voice across multiple rewrite passes? Pin a voice sample (200 words) at the top of every prompt, even passes 2 and 3. Don’t rely on the model remembering voice from turn 1.
- Should I trust the model’s translation as-is? No for any high-stakes piece. Use the “localize, then list adaptations” pattern (#7, #14). Then have a native speaker spot-check the adaptations list, not the whole article; that’s where errors hide.
- What’s the minimum article length these prompts work on? Roughly 300 words. Below that, just edit by hand — the prompt overhead outweighs the benefit.
- Can I run a rewrite + SEO + translate flow in one chat? Yes, but reset the chat between each rewrite. Carry forward only the artifact, not the model’s prior reasoning, to prevent voice drift.
- How long is too long for a single rewrite prompt? Roughly 2,000 words of pasted article. Above that, chunk by section — the model loses track of voice and starts neutralizing.
Related
- Blog outline prompts — outline before drafting so rewrites are minor, not structural
- Tone rewrite prompts — for voice-only rewrites that don’t change structure
- FAQ writing prompts — common rewrite target when refreshing articles
- Landing page copy prompts — the rewrite-equivalent set for landing pages
- ChatGPT writing assistant — full chained workflow including rewrite passes
- Article rewrite use case — end-to-end example with before/after
- AI content repurpose tutorial — apply template #17 across formats systematically
- Comparison Article Prompts: 12 Templates for “X vs Y” That Doesn’t Suck
- Customer Education Article Prompts for Help Articles
- Policy & Terms Plain-English Rewrite Prompts
- Problem-Solution Article Prompts That Earn Reader Trust
- Short-form Summary Prompts: Compress Without Losing Truth
- Scaling Content with AI Without Tanking Quality
- Using AI to Review and Improve Existing Articles
- Plan a Multi-Part Article Series With AI
- Long-form Article Expansion Prompts for Real Depth
- Press Release Rewrite Prompts for a Journalist’s Inbox