Ask a model to “polish” a draft and it usually regresses everything toward the mean: hedged, smooth, voiceless. The fix is not a better model — it is a sharper instruction. The 12 prompts below each target one specific weakness (passive voice, jargon, the 2026 AI tells, dead adverbs) so the editor cuts the rot without flattening the author’s voice.
TL;DR
- Run these on a draft that already says something. They are a finishing pass, not a writing pass.
- One prompt = one job. Stacking “tighten + de-jargon + voice-match” in a single request produces mush.
- The single biggest voice lever is pasting 1-2 real samples of the voice you want. “Be friendly” is noise.
- For the final human-sounding read, Claude Sonnet 4.6 needs the least cleanup on long-form (as of June 2026); GPT-5.5 is faster for short, mechanical passes; Gemini 3.1 Pro is strong on very long documents inside its 1M-token window.
- Always end with a human read-aloud. The model is confident, not always correct.
Which model for the editing pass (June 2026)
You do not need a top-tier subscription to edit. All three of these run a competent editing pass on the free or $20 tier.
| Model | Best editing job | Notes (June 2026) |
|---|---|---|
| Claude Sonnet 4.6 | Long-form voice + tone match | Holds tone across a full article; least cleanup afterward. 1M-token context. Free tier limited; Pro $20/mo. |
| GPT-5.5 | Fast mechanical passes (tighten, active voice) | ChatGPT default since ~Apr 2026. Plus ($20) in-app context ~320 pages; full 1M only on $200 Pro. |
| Gemini 3.1 Pro | Editing across a very long doc | 1M context on Google AI Pro ($19.99/mo, the tier formerly called Gemini Advanced). |
Paste the draft, paste the prompt, paste your voice sample. That is the whole workflow. Skip “humanizer” tools — they swap synonyms and leave the structural tells (uniform rhythm, negative parallelism) untouched.
Prompt anatomy: the six elements
Every polishing prompt should carry six things. Drop one and the output drifts generic.
- Audience — one specific reader, not “users.”
- Goal — one outcome: read to the end / click / agree / share.
- Voice — 2-3 anchor adjectives, plus a sample.
- Constraints — word count, banned phrases, must-keep facts.
- Format — paragraph, bulleted, headed, or table.
- Examples — 1-2 tone samples. This is the strongest single signal for voice.
The 2026 AI tells worth hunting
Before you run any prompt, know what you are removing. The recognizable “AI accent” in mid-2026 is less about single words and more about clustered patterns:
- Negative parallelism: “It’s not X, it’s Y” and “not just X, but Y.” Wikipedia’s editors now flag this as a top sign of machine writing, and one analysis found “not just X, but Y” in roughly 6% of all AI messages in a sampled month.
- Em-dash overuse: even after the late-2025 toggle, models still over-deploy the em-dash as a rhythm crutch.
- Tier-1 words: delve, tapestry, leverage (as a verb), harness, robust, navigate, landscape, multifaceted, “it’s important to note,” “in today’s fast-paced world.”
- Uniform sentence rhythm: stretches of same-length sentences that read like a metronome.
No single word proves AI authorship. Three or more of these clustered in a short passage does. See Wikipedia’s Signs of AI writing for the full, regularly updated list.
12 copy-ready prompt templates
Each prompt is self-contained. Paste your draft below it. Swap any [bracketed] placeholder for your own value.
1. Tighten by 20%
Tighten this draft by 20% without losing meaning. Cut: dead adverbs, hedges (probably, perhaps, somewhat), filler phrases ("at the end of the day"), redundant pairs ("each and every"). Keep voice intact. Output side by side: a cut count, then the final version.2. De-jargon for a non-expert
Audience: [non-expert audience]. Find the jargon in this draft. For each term: replace with plain English, OR explain it in a parenthetical the first time it appears. Do not kill specificity — keep the terms this audience already uses.Swap: [non-expert audience] (e.g. “small-business owners with no finance background”).
3. AI-tell remover (2026 edition)
Strip the 2026 AI tells from this draft. Find and rewrite: (1) negative parallelism ("it's not X, it's Y", "not just X, but Y"); (2) overused em-dashes used for rhythm; (3) Tier-1 words: delve, tapestry, leverage, harness, robust, navigate, landscape, multifaceted, "it's important to note", "in today's fast-paced world". Replace each with the specific thing the cliche was avoiding. Then vary the sentence rhythm. Output the rewritten draft only.4. Voice-match polish
Brand voice anchors: [3 adjectives]. Here is a real sample of the voice: [paste 2-4 sentences]. Re-polish this draft to match that sample's diction, sentence length, and energy. Do not invent new claims. Output changed lines only, as a diff.Swap: [3 adjectives], [paste 2-4 sentences]. The sample matters more than the adjectives.
5. Active-voice pass
Find passive-voice constructions in this draft. For each, rewrite in active voice — unless passive is genuinely correct (the actor is unknown or the action deserves the emphasis). Do not flip every passive. Judge case by case and list the ones you intentionally kept.6. Paragraph flow audit
Audit paragraph flow and report only: (1) paragraphs that abandon their topic mid-way; (2) paragraph breaks in the wrong place; (3) missing transitions where the topic actually shifts. For each, say split / merge / bridge and show the one-line fix. Do not rewrite the whole draft.7. Headline + first-line strengthening
Rewrite the headline and first two sentences. Headline: 8-12 words, concrete, no colon-clickbait. First sentence: a specific hook. Second sentence: pays off the hook. Do not start with "In this article" or a rhetorical question. Give me 3 options.8. Read-aloud test
Read this draft aloud in your head and flag every sentence that: (a) trips on a comma, (b) runs past 25 words, (c) is the 3rd-plus long sentence in a row, or (d) forces a re-read. For each flag, suggest a split or a cut. List line numbers.9. Sentence variety
Audit sentence-length variety. Flag any stretch of 4+ consecutive sentences of similar length. Suggest a rewrite that alternates short and long. Do not make everything short — that reads choppy. Show before/after for each stretch.10. Cliche-to-specific replacement
For each cliche in this draft, generate 3 specific replacements using real nouns and numbers. Example: "moving the needle" -> "doubling weekly trial sign-ups". Skip any cliche that genuinely has no concrete replacement, and say so.11. Cut a sentence per paragraph
Apply this discipline: from every paragraph, cut at least one sentence — the weakest one. Output a two-column result: the cut sentence, then the leaner paragraph. If a paragraph cannot lose a sentence without breaking, flag it instead of cutting.12. Tone calibration
I think this draft reads [current tone] but I want [target tone]. Find the 5 paragraphs where the tone is most off-target. Rewrite each one to land on target, and explain in one line what you changed. Do not rewrite the whole piece.Swap: [current tone], [target tone] (e.g. “stiff and corporate” -> “direct and warm”).
How to push results further
- Give 1-2 real voice samples. Adjectives alone produce the same flavor every time.
- Constrain ruthlessly: cap the word count, ban specific phrases, name the must-keep facts.
- Run one job per prompt, then chain them — tighten, then de-jargon, then voice-match.
- Use AI for passes one and two; do pass three by hand. Pass three is what ships.
- Test the headline standalone. If it doesn’t carry meaning alone, it won’t in a feed.
- Read the final draft aloud. Your ear catches what the model’s “polish” smoothed over.
Common mistakes
- Vague audience — “users” produces generic output every time.
- No voice sample — every variant comes back the same flavor.
- No constraints — without a word cap and a banned list, the model adds, never cuts.
- Stacking jobs — “tighten + de-jargon + voice-match” in one prompt yields mush.
- Trusting the first pass — models land on the safe middle; push for a second version.
- No fact-check — the model is confidently wrong sometimes; verify every claim it touched.
FAQ
- Will these prompts remove the em-dash problem?: Prompt 3 catches most of it, but models slip em-dashes back in even with the dedicated suppression toggle on. Do a final find-and-replace by hand before publishing.
- Which model needs the least cleanup?: As of June 2026, Claude Sonnet 4.6 holds tone best across a full long-form draft. GPT-5.5 is quicker for short mechanical passes; Gemini 3.1 Pro shines on very long documents.
- Can AI do the whole edit?: Use it for the first two passes. Do the third by hand — that is the one that ships.
- Do “humanizer” tools work?: Not really. They swap synonyms and leave the structural tells (uniform rhythm, negative parallelism) in place. Prompt 3 plus a human read does more.
- One voice or per-channel?: One brand voice; shift the tone axes (formal/casual, brief/detailed) within it per channel.
- Can I reuse these for other content types?: Yes. Swap the audience, goal, and voice sample — the structure carries over to email, ads, and docs.