TL;DR
“Rewrite this” is the single worst way to use ChatGPT for writing — it averages your prose toward fluent, generic mush. The workflow that works splits writing into four stages and only hands two of them to the model:
- Pre-writing (audience, angle, outline) — ChatGPT is genuinely useful here.
- Drafting — you write the sentences. This is where your voice lives.
- Editing — ask ChatGPT to critique structure, never to rewrite. Its best stage.
- Polish — headline and opener variants, voice match. Use sparingly; voice gets sanded off.
This guide gives you the exact prompts for stages 1, 3, and 4, plus the June 2026 ChatGPT details that changed how the polish step works (GPT-5.5 model picker, where Canvas went, and how detectors actually behave).
Who this is for
- Writers and content marketers who want AI help without their drafts collapsing into median internet prose.
- Founders and operators shipping emails, posts, and announcements every week.
- Non-native English writers who want to lift their prose without it reading as translated.
- Anyone who has pasted a draft into ChatGPT, accepted the rewrite, and cringed at it a week later.
What changed in ChatGPT for writers (June 2026)
A few things shifted in 2026 that directly affect this workflow. Get these right and the prompts below land better.
- The default model is GPT-5.5 Instant. Since the GPT-5.5 rollout (default for logged-in users from February 13, 2026, after GPT-5 was retired), the picker shows Instant / Thinking / Pro. Instant auto-routes hard requests to Thinking on its own; you don’t have to babysit it for everyday editing.
- Canvas moved. As of mid-2026, Canvas is no longer available under GPT-5.5 Instant — only under GPT-5.5 Thinking. Under Instant, longer writing now comes back as inline writing blocks you can edit in place. If you want the side-by-side document editor with revision shortcuts (adjust length, change reading level, final polish), switch the model to Thinking first.
- Free has limits and ads. The US free tier started showing ads in February 2026 and keeps tight message caps. For weekly writing work, Plus at $20/month (as of June 2026) removes the friction that breaks a drafting session.
- It’s less over-formatted than it was. GPT-5.5 is noticeably better at not dumping aggressively bulleted, boilerplate output. It still defaults to a generic professional register, which is exactly why the voice-anchor step below matters.
The four stages — and which two to delegate
| Stage | What happens | Hand to ChatGPT? | Why |
|---|---|---|---|
| Pre-writing | Audience, angle, outline | Yes | You’re committed to no prose yet, so generic output is fine to remix. |
| Drafting | You write the sentences | No | The model produces the same arc you’ve read ten times today. |
| Editing | Structural critique, cuts, weak spots | Yes — best stage | Critique is high-value and low-risk; you stay the author. |
| Polish | Headline variants, voice match | Sparingly | This is where voice quietly gets averaged away. |
Before you open the chat
- Bring a target audience and goal. “Audience: indie devs. Goal: get them to try our SDK in 5 minutes” gives the model traction. “Write a blog post about X” gives you the median article on X.
- Bring a voice anchor. Paste 5–10 paragraphs you’d want to write like at the start of the session. Better: save them once in Settings → Personalization → Custom instructions, or build a Custom GPT with those samples as Knowledge, so every chat starts from your voice instead of the default professional one.
- Decide the format up front (long form, email, thread). Different formats need different prompts.
- Set one hard constraint — a word count, a banned-phrase list (“we’re excited to announce”), or a no-jargon rule. Constraints are what pull the output away from average.
Step by step
-
State audience and goal in one paragraph. Be specific:
Audience: senior PMs at B2B SaaS companies, 5+ years experience. Goal: convince them to try our analytics tool. They've heard pitches like this 100 times this year. -
Ask for the outline first — 4–6 sections, one sentence of purpose each. Then edit it ruthlessly. The outline is where the piece lives or dies; spend your effort here, not on prose.
-
Draft section by section, yourself. Reach for ChatGPT only when stuck, and when you do, paste your context, the draft so far, and ask for one specific thing: a transition, a stronger opener, a counter-example.
-
After the draft is complete, ask for structural critique only (this is the highest-leverage prompt in the whole workflow):
Critique this draft for structure: where does it drag, where is the argument weakest, what should I cut? Do NOT rewrite anything. -
Make YOUR cuts based on the critique. You decide; the model diagnoses.
-
Polish pass: ask for 3 headline variants, 3 opening-line variants, and “what would I lose by cutting paragraph X?” If you want the document editor for this, switch to GPT-5.5 Thinking and use Canvas; otherwise keep it inline.
Prompts that beat “rewrite this”
Copy these. They ask for diagnosis and options, never wholesale replacement.
TIGHTEN
Read paragraph 2. Suggest 3 specific cuts that would tighten it
without losing meaning. Show me exactly what would remain.
VOICE-MATCH
Below is one paragraph of mine. Edit the draft to match its voice
on three axes only: sentence length, vocabulary register, and use
of contractions. List the changes you made and why.
STRESS TEST
What's the strongest objection a skeptical reader would raise to
section 3? Quote the exact sentence where the argument is weakest.
KILL THE BOILERPLATE
Find every sentence that could appear in any article on this topic.
List them. Don't fix them — I will.Quality check before you ship
- Read it aloud. Anywhere you slow down or stumble, edit. The model can’t hear you.
- Test the opener. Is the first line specific to this piece, or could it open 100 other articles?
- Check the claims. Could a reader actually call you on them? Vague = AI; specific and falsifiable = you.
- Keep your weird phrasings. The slightly-off lines that survived editing are the bits that make a piece read as human. The model wants to smooth them; don’t let it.
Does AI detection catch ChatGPT-edited prose?
Mostly no, if you did the work. Detectors like GPTZero catch roughly 95% of raw, unedited AI text, but accuracy drops to about 60–80% on heavily edited or paraphrased content (as of 2026). Run this workflow — your draft, your structure, your voice anchor — and you’re well inside that gap, because the words are mostly yours.
Two cautions that matter more than the detection question:
- False positives are real. Even the strongest detectors flag some human writing as AI, and there’s a documented bias against non-native English. Treat any single detector score as a signal, not a verdict.
- Detectors are not your editor. A piece that “passes” can still be generic. Your read-aloud test is the better quality gate.
ChatGPT or Claude for this?
Both work; the split is real. In independent blind evaluations from Q1 2026, reviewers preferred Claude’s prose about 47% of the time versus 29% for the contemporary GPT model and 24% for Gemini — Claude tends to need the least cleanup and holds tone best across a full article. ChatGPT (GPT-5.5) is faster for tight iteration loops and has narrowed the gap on over-formatting. Practical rule: outline and iterate in ChatGPT; if the final prose still sounds generic, run the same draft through Claude’s writing workflow and keep the better output.
Common mistakes
- Skipping the outline. The piece meanders because the model has nothing to anchor on.
- Asking for the full piece in one shot. You get generic prose with no surprising sentences.
- Not specifying audience. “Write about X” returns the version of X already written a thousand times.
- Accepting the rewrite wholesale. The model averages toward boring; you’re the variance.
- Editing inside the chat box. Keep your 1,500-word draft in a real editor; the chat is for short prompts.
- Asking it to “make it more engaging.” You’ll get exclamation marks and emoji.
How to reuse this workflow
- Keep a voice-anchor file of 5–10 paragraphs you’d want to write like, or load them once into Custom instructions / a Custom GPT.
- Save the structural-critique prompt — it’s the most reusable, highest-leverage one across any draft.
- Track which AI suggestions you accept vs reject. After ~10 pieces, a pattern emerges and you can tune the prompts to your taste.
FAQ
- Should I use GPT-5.5 Instant or Thinking for writing?: Instant for everyday outlining, critique, and quick variants — it auto-routes harder requests to Thinking anyway. Switch to Thinking explicitly when you want Canvas (the side-by-side document editor with length and reading-level shortcuts), which Instant no longer offers as of mid-2026.
- Where did Canvas go?: It’s still here, but only under GPT-5.5 Thinking. Under GPT-5.5 Instant, longer writing returns as inline writing blocks you edit in place.
- Is the free plan enough?: For occasional use, yes, though the US free tier now shows ads and caps messages. For weekly writing, Plus ($20/month as of June 2026) removes the caps that interrupt a drafting session.
- Should I use Claude instead?: For final prose quality, Claude is often a touch more natural and needs less cleanup; ChatGPT is faster to iterate with. Run the same draft through both and keep the better one.
- Can I write a whole book this way?: Not in one pass — the model loses the macro arc over a full manuscript. Chapter outlines and per-section editing, yes.
- Is it cheating?: Depends on the venue. School: read the policy. Professional: if the ideas are yours and the words pass your own bar, the workflow is fine.