The One Prompt, If You Only Use One
Paste this with your draft:
Rewrite the text below. Delete every instance of: three-item lists used for rhythm rather than content; the construction "it's not just X, it's Y" and its variants; stacked hedges like "it's worth noting that" and "it's important to remember"; any closing paragraph that restates what the piece already said. Vary paragraph length deliberately - some one sentence, some six. Do not add warmth, enthusiasm or personality. Only remove. Return the text with nothing else.
That will get you most of the way. The rest of this page is why it works and what to do next.
Why "Write More Naturally" Does Nothing
Almost every prompt circulating for this asks the model to add something - be conversational, write like you're talking to a friend, use a warmer tone, sound more human.
The result is reliably worse. You get the same underlying structure with enthusiasm painted on: more exclamation, more second person, more "let's dive in". The tells are structural, and adding personality on top leaves every one of them in place.
The tells are also specific, finite, and deletable. Which means the useful instruction is not "sound human". It is "remove these constructions".
The Tells, Named
You need to be able to name them for the prompts below to work. These are the ones that actually mark text as machine-written to a reader who reads a lot:
- The rule of three. Three adjectives, three examples, three clauses - used for rhythm rather than because there are three things. Once you see it you cannot stop seeing it.
- "It's not just X, it's Y." Along with "X isn't about A - it's about B". A pivot construction that sounds insightful and usually carries no information.
- Stacked hedges. "It's worth noting", "it's important to remember", "that said", "of course" - often two in a row.
- Uniform paragraph length. Human writing lurches. Four tidy three-sentence paragraphs in a row is a signature.
- The closing restatement. A final paragraph that summarises what you have just read, adding nothing. Writers end on a point; models end on a summary.
- Vocabulary. Delve, leverage, robust, landscape, realm, tapestry, navigate, underscore, pivotal, seamless.
- Symmetry. Every section the same depth, every list the same length, every example matched. Real writing is lopsided because the writer cared more about some parts.
Prompt 1 — Strip the Tells
Here is my draft: [paste].
Remove, without replacing: three-item lists used for rhythm; "it's not just X, it's Y" constructions; hedges such as "it's worth noting" and "it's important to remember"; the words delve, leverage, robust, landscape, realm, navigate, underscore, pivotal, seamless; and the final paragraph if it only restates earlier content.
Change nothing else. Do not improve, expand or add transitions. Return only the edited text.
Why it works: "do not replace" is the critical clause. Without it the model substitutes a different flourish for each one it deletes and you end up where you started.
Prompt 2 — Match My Voice
Below are two things: a sample of my own writing, and a draft to edit.
First, describe my writing in specifics - average sentence length, how I open paragraphs, punctuation habits, words I use and words I avoid, how formal I am, whether I use contractions, how I handle lists.
Then rewrite the draft to match that description. Where my sample and the draft conflict, follow my sample.
MY WRITING: [paste 300+ words you actually wrote]
DRAFT: [paste]
Why it works: this is the single biggest improvement available and almost nobody does it, because it requires supplying a sample. "Sound human" has no target. "Sound like this" does.
Prompt 3 — Break the Rhythm
Rewrite this so the rhythm is uneven. Specifically: at least one paragraph of a single sentence. At least one that runs past five. At least one sentence under six words. Do not make every section the same depth - go deeper on the part that matters most and briefer on the rest.
Keep every fact and all the meaning. Change only length and structure.
TEXT: [paste]
Why it works: uniformity is the tell most people feel without being able to name. Attacking it directly fixes something the other prompts leave behind.
Prompt 4 — Cut the Summary
Read this and find every sentence that restates something already said - section summaries, recaps, the closing paragraph, any "in short" or "to sum up".
List them, then give me the text with all of them deleted.
If removing one makes the piece end abruptly, say so rather than writing a replacement ending.
TEXT: [paste]
Why it works: models restate because it reads as thorough. Readers experience it as padding, and an abrupt ending is almost always better than a summarising one.
Prompt 5 — Specific Over General
Find every vague claim in this text - anything using "various", "numerous", "a range of", "many", "significantly", "substantially", or any benefit stated without a number.
For each, either ask me for the specific detail, or tell me the sentence should be deleted because it carries no information.
Do not invent figures. Do not guess. Asking me is the correct answer.
TEXT: [paste]
Why it works: vagueness reads as machine-written because a person writing from experience has the specifics. This prompt surfaces where you were vague, which is usually where you did not know.
Prompt 6 — The Read-Aloud Test
Go through this sentence by sentence and flag anything a person would not say out loud - constructions that only exist in written text, phrases that need a second reading, sentences where the subject and verb are far apart.
For each, give me a version you would actually say.
Do not flag things that are simply formal. Flag things that are unnatural.
TEXT: [paste]
Why it works: it is the oldest editing test there is and it catches a different category from everything above.
The Order Matters More Than the Prompts
Running all six at once produces mush. The sequence that works:
- Draft however you like. Do not try to make it good at this stage.
- Prompt 1 — strip the tells. Subtraction only.
- Prompt 4 — cut the summaries. Usually removes 10-15% of the word count.
- Prompt 2 — match your voice, using a real sample. This is where it stops sounding generic.
- Prompt 3 — break the rhythm. Last, because earlier edits flatten it again.
- Read it aloud yourself. Prompt 6 is a backstop, not a substitute.
Strip before you style. Every attempt to do both at once gives you a polished version of the thing you were trying to remove.
One Thing Worth Being Straight About
These prompts are for writing you are going to edit, own and put your name on - your marketing copy, your emails, your drafts.
They are not a way to pass off AI output as unaided work where that matters. If you are submitting something that has to be your own writing, the problem is not the prose style. And if you are wondering about AI detectors: they are unreliable in both directions, they flag plenty of human writing, and neither passing nor failing one tells you much. Write something you would defend and the question mostly goes away.
FAQ
How do I make ChatGPT sound like a human?
Tell it what to remove rather than what to add. Ask it to delete three-item lists used for rhythm, "it's not just X, it's Y" constructions, stacked hedges, and any closing paragraph that restates the piece — without replacing them. Then give it 300 words of your own writing and ask it to match that.
What is the best prompt for ChatGPT to make my text sound human?
The one at the top of this page, which strips the structural tells in a single pass. If you only have time for one more, use Prompt 2 and supply a sample of your actual writing — a target beats an instruction.
How can I make AI text sound more human-like?
Vary the rhythm. Uniform paragraph length is the tell most readers feel without naming it. Mix single-sentence paragraphs with long ones, go deeper on the section that matters and briefer elsewhere, and delete anything that restates what you already said.
What is the best prompt to humanize AI-generated text?
A subtraction prompt, not an addition one. Asking for warmth, enthusiasm or a conversational tone leaves the structure intact and paints over it. Naming the specific constructions to delete changes the structure, which is what reads as machine-written.
Do these work on Claude, Gemini and Grok?
Yes. The tells are shared across models because the training data and the preferences behind them overlap heavily. The vocabulary list varies slightly by model — add whatever words you notice yours overusing.
Will this get past an AI detector?
Possibly, and that is not what they are for. Detectors are unreliable in both directions and regularly flag human writing. These prompts make text read better; they are not a compliance tool, and anywhere disclosure is required the honest route is disclosure.
These six are a starting set. Our full library has 744 prompts across 123 collections for ChatGPT, Claude, Gemini, Grok and Kimi — the ChatGPT writing collection is the closest neighbour to this page.