ChatGPT Prompts for Writing That Keep Your Voice

The top result for chatgpt prompts for writing is a Reddit thread whose whole prompt is a list of things not to sound like. Use simple language, short sentences, plain examples. It's a good list. It's also the same list a great many people have pasted into the same model, which is why text written under it tends to sound like one careful, faceless person.

Prompts that keep your voice do three things the lists don't. They carry a description of how you actually write, built from your own paragraphs instead of adjectives like "conversational". They stop the model from rewriting anything that has your name on it and make it mark the text instead. And they hand full drafting only to the jobs where voice was never the point, title options, question lists, premises. Eight prompts below do that, shown in full, and the first one produces the voice description the other seven depend on.

It helps to know what you're prompting against first. OpenAI wrote it down.

What The Model Does To A Paragraph By Default

The guide for OpenAI's current flagship, GPT-6 Astra, has a section on personality and writing style, and it's unusually frank. The model "tends toward detailed, formatted responses and may use recurring phrases across sessions". It "tends to use lists, tables and Markdown to make responses scannable". The same page ships a prompt to rein that in, which bans concluding summary statements, contrastive framing of the "X, not Y" kind that "introduces an unprompted alternative that the user didn't ask about", invented compound labels, vague qualifiers, and canned transitions. The full reading of that document is in openai prompt engineering.

I read that section as a list of what happens to your paragraph when you paste it in and ask for it to be better. Each row is one default the vendor documents, what it looks like from the writer's side, and the line that stops it.

The last row comes from a different part of the same page, which says the model is "more likely to ask the user a question when additional input could materially change the result". Useful in an agent. Annoying in an editor.

Every "stops it" line in that table is short enough to live in your custom instructions permanently. The rest of this article is about what the table can't give you. What your prose does that a generic draft doesn't.

Build The Voice Spec From Your Own Paragraphs

"Match my tone" doesn't work because tone isn't a measurement. The ranking pages suggest asking for "conversational, professional, witty, intelligent, casual", and every one of those words describes a million writers. The model needs things it can count and quote. Sentence length range. Whether you ever write a one-sentence paragraph. Which ten words would identify you in a lineup.

So the first prompt isn't a writing prompt at all. It's an analysis prompt, and the role is deliberately not "editor".

You are a copy analyst, not a writer. Below are three samples of my writing, each written without any AI help. Do not summarize them, praise them, or rewrite any part of them. Produce a voice specification with these fields, each answered from evidence in the samples and nothing else:

1. Sentence length: shortest, longest, and typical range in words. Quote the shortest sentence.
2. Paragraph habits: typical paragraph length in sentences, and whether I use one-sentence paragraphs. Quote one if I do.
3. Openings: how my paragraphs and sections begin. Quote three openings.
4. Contractions: always, never, or mixed, with an example.
5. Punctuation: which marks I use and which never appear. Say explicitly whether semicolons, parentheses, and dashes are absent.
6. Vocabulary: ten words or phrasings that recur and would identify me. Ten words that never appear here and that a generic draft would reach for.
7. Hedging: how I signal uncertainty, quoted.
8. Structure: whether I use lists, headings, bold, or rhetorical questions, and roughly how often.
9. Endings: how sections end. Quote two.
10. Never list: five things a draft in my voice must not do, each justified by the samples.

If a field cannot be answered from the samples, write "not enough evidence" instead of guessing. Output the specification as plain numbered text under 400 words.

Sample 1:
[PASTE]
Sample 2:
[PASTE]
Sample 3:
[PASTE]

Three samples of at least a few hundred words each, and they have to be yours, written before you had a model to lean on. Old emails work. So does a forum post you were proud of. A sample the model helped write will return a spec of the model.

The structure matters more than the wording. "Analyst, not a writer" switches off the improving reflex, which is the same reflex that upgrades your verbs when you didn't ask. "Quote" appears in seven of the ten fields, because a quoted sentence is evidence and a description is an opinion. The "never list" is the field I'd fight for if I could only keep one, since the fastest way to tell a draft isn't yours is usually one thing it does that you never do. The 400-word cap is there for storage, and storage is the next section.

What comes back is a numbered document, mostly quotations, that reads like a style guide written by someone who's only seen three pages of your work. That's the right amount of confidence. Delete anything that's wrong, then wrap it.

Voice spec. Follow this over any general writing advice you have:
[PASTE THE SPECIFICATION]

Rules that override the spec where they conflict: no lists unless I ask, no closing summary sentence, no alternative raised in order to dismiss it, no coined labels, no transitions between sentences, no questions back to me.

That block goes at the top of every writing prompt from here on. I'll call it the voice block and write [VOICE BLOCK] where it sits.

Where To Keep It

Two places in ChatGPT hold instructions between chats, with different limits and scopes. The numbers are from the help center pages as archived in September 2026, since the live pages refuse automated readers.

The custom instructions field is under Settings, then Personalization on web and desktop, and under Customize ChatGPT on the phone apps. On the free plan the 1,500-character limit is roughly 250 words, which is why the ban list lives there and the spec doesn't. A 400-word spec fits the 5,000-character field on paid plans with room to spare.

Projects are where I'd put the spec, and the help page's own example of a project instruction is a marketing-mentor persona, so that's the intended use. Two facts from that page matter. Project instructions override your global custom instructions inside the project, so the writing project can have rules your general chats don't. And projects have built-in memory across their chats and files, which cuts both ways. You don't paste the spec every time. A bad edit you accepted last week is also still in the room, so when the spec changes I'd start a fresh project instead of talking the old one out of its habits.

Write, Mark, Or Keep

The lists on the results page share one assumption, that the model's job in a writing prompt is to write. Half the time it shouldn't. The middle column is the whole method.

The "you keep it" rows are the ones I'm most sure of. I self-publish books, and the rule I hold there is that the first draft of anything with my name on the cover is mine, sentence by sentence, before the model sees it. Judgment, not a study. What I notice is that once a model paragraph is on the page, my next paragraph starts matching it, and by the end of a chapter I can't tell whose habits I'm reading.

Marking keeps the useful half. The model is a good reader. It's a mediocre co-writer of you.

The Prompts

Each of these starts with the voice block where it's relevant. None of the mark-only prompts are allowed to write a replacement, and the writing prompts are restricted to jobs from the table above.

The Mark-Only Editor

[VOICE BLOCK]
Read the draft below against the voice spec. Do not rewrite anything. For each place where the draft departs from the spec, quote the exact words, name the spec field it breaks, and say in one line what kind of change would fix it, without writing the fix. Separately, quote any sentence a reader would recognise as generated rather than written, and say why. Order the whole list by how much each item matters. No praise, and no summary at the end.
Draft:
[PASTE]

This is the prompt I'd copy if you copy one. "Name the spec field" turns every note into a check against your own rules instead of the model's taste, so "this sentence is 41 words and your typical range is 12 to 20" is a note you can act on, and "this could be more engaging" isn't. "Without writing the fix" keeps it an editor. The closest prompt on the results page says "Do not suggest any rewriting of sentences, unless unclear", and that "unless unclear" is a door the model will walk through on every third sentence. Close it.

The Cut

[VOICE BLOCK]
The draft below is [N] words and needs to be [M]. Do not rewrite or compress any sentence. List whole sentences that could be deleted without losing a claim, in the order you'd delete them, quoting each one and giving the running word count after each deletion. Stop when the target is reached. Then, separately, quote any sentence that repeats a point made earlier in the draft, alongside the earlier sentence.
Draft:
[PASTE]

Deletion rather than compression, for the reason in the table. A compressed sentence is a new sentence. The running count is there because you'll disagree with the order, and it lets you skip an item and still know where you stand.

Outline Into Questions

Here is an outline for a piece I'm writing. Do not write any of it. For each heading, write the three questions a skeptical reader would want that section to answer, in the order they'd ask them. Then, for the whole piece, write the one question a reader would still have if every section answered its three. Plain numbered list, no commentary.
Outline:
[PASTE]

The lists on the results page expand outlines into drafts. This one expands an outline into a reader. That's the harder half of drafting, and the half that doesn't touch your voice. You answer the questions yourself, and the section writes itself around the answers.

Title Options

Give me 20 title options for the piece described below. Rules: under 60 characters, no colon, no question, no "how to", no number at the start, and none of these words: [PASTE THE TEN NEVER-WORDS FROM YOUR SPEC]. Ten should name the specific thing the piece does. Ten should name the mistake it prevents. Do not rank or explain them.
Piece: [ONE SENTENCE ON WHAT IT ARGUES AND WHO IT'S FOR]

Twenty because the first five are always the obvious ones. The never-words do double duty here, since a title is the first place a generic word gets noticed.

The Argument Against

Read the draft below as a reader who disagrees with it and knows the subject well. Find the weakest claim, quote it, and write the strongest two-sentence objection to it. Then find the claim that most needs evidence I haven't given, and say what evidence would settle it. Do not suggest wording. Do not soften anything.
Draft:
[PASTE]

No voice block on this one, because the model isn't touching prose. It's a reader with a grudge, a role the model plays well and one that rereading your own draft will never give you.

A Story Premise, Since People Search For The Template

Generate 10 story premises. Each premise is three lines: a character with one specific want and one specific thing they are wrong about; a setting given as a place and a time of day; a first complication that arrives inside the first scene. Constraints: no chosen ones, no prophecies, no amnesia, no dead spouses, no "little did they know". No two premises in the same genre. Do not write any prose, dialogue, or titles.

The constraints are the prompt. Without them you get the ten premises everyone gets. "Wrong about" is the field that produces a story instead of a situation, and "time of day" is a cheap way to force a concrete image. Everything after the premise is yours.

Academic Clarity Without Touching A Claim

Below is a paragraph from an academic draft. Do not change any claim, number, citation, or technical term, and do not add or remove a sentence. Quote each sentence a reader outside my field would need to read twice, and name the cause in one line, choosing from: the subject is far from the verb, a noun is doing a verb's job, two ideas share one sentence, a term is used before it is defined. Do not rewrite. I will.
Paragraph:
[PASTE]

The related searches want an academic-writing PDF, and the GitHub repository on the results page is the closest thing to one. Its prompts are mostly of the "improve the language and style" kind. This one names four causes and forbids the fix, because in academic prose a helpful rewrite that moves a hedge is a changed claim.

Canvas Changes Which Prompts You Need

If you write in ChatGPT's canvas rather than the chat, some of the above is built in and some of it is built in the wrong direction. The help page describes writing shortcuts on a menu at the bottom right. "Suggest edits" adds inline suggestions, the mark-only editor as a button. "Adjust the length" has a slider. "Change reading level" runs from Kindergarten to Graduate School. "Add final polish" checks grammar, clarity, and consistency, and the same page says that shortcut, like the emoji one, updates the contents of your document immediately. No suggestion step.

So on your own draft, I'd use Suggest edits and nothing else from that menu. A polish that applies itself is the "you keep it" rule being broken by a button. What makes canvas worth it is the version history and the Show changes button, which the page says shows additions and deletions between versions, so when you do accept an edit you can see which of your sentences moved. Two limits from the same snapshot. Canvas was listed as available on web, Windows, and macOS with mobile "coming soon", and it isn't available with the pro-series models.

The Questions From The Results Page

What are the best writing prompts for ChatGPT? The analyst prompt, once, and the mark-only editor, every time. Those two change what the model is allowed to do. Everything else on this page is a job-specific convenience.

What are the top 10 ChatGPT prompts, and what are some good ones generally? A top ten is a top ten of somebody else's tasks. The six slots that make any prompt work are in chatgpt prompts, and the reusable versions with the variable slots marked are in chat gpt prompts.

What are some creative prompts, and what prompts make ChatGPT go crazy? The premise prompt above is the creative one I'd keep. The "go crazy" family is mostly about giving the model a persona and watching what it does with it, and the ones that teach you something about how it works are in fun chatgpt prompts.

What are the trending ChatGPT prompts for writing? The one at the top of the results page, the "write naturally" rules list. It works, in the sense that text written under it reads less like a press release. It's also a floor. Every rule in it is a thing not to do, and a voice is made of the things you do.

How do I make ChatGPT write like a human? The table at the top is half the answer, since the vendor documented the tells and you can ban them by name. The spec is the other half, because "like a human" isn't a target. Like a particular one is.

What I'd Do First

Find three things you wrote before you had a model to write with. Run the analyst prompt on them tonight. Read the never list twice and cross out anything you think is wrong.

Then take a draft you're already unhappy with and run the mark-only editor, and do the rewriting yourself, sentence by sentence, with the notes open beside it. It's slower than "make this better". The difference is that when you're done, the piece is still yours, and the prompt didn't have to guess what that meant.