Prompts to Make ChatGPT Write Like a Human

Why ChatGPT's writing sounds generated

Ask ChatGPT for «a friendly, natural tone» and you'll usually get the same voice with a few exclamation marks added. That's because a tone word gives the model nothing to copy. It falls back on its defaults: an opener that delays the point, three tidy examples in a row, a hedge in every paragraph, a closing line that sums up what you just read. A reader spots that voice fast, and it's the voice of nobody in particular.

Writing that sounds like a person comes from material only you have. A request that gets there gives ChatGPT five things:

1. A sample of how you write. Two or three paragraphs you wrote yourself, before any AI touched them. Emails count. The model copies sentence length, vocabulary and how you open far better than it follows adjectives like «warm».
2. The specifics. Names, numbers, dates, the thing that actually happened. Generic text is mostly text with the details missing, and nobody but you knows them.
3. The habits to drop, by name. «No opener, no summary at the end, no lists of three, no rhetorical questions.» Banning a habit works better than asking for a mood.
4. Who's reading and where. A note to your team, a post on LinkedIn and a page on your site want different lengths and different levels of polish.
5. Permission to leave gaps. Tell it to mark a missing detail in brackets instead of inventing one. Otherwise it fills the hole with something plausible and false.

Here's that request in a form you can type from memory:

Here are two things I wrote: [paste]. Match their sentence length,
vocabulary and how they open.
Write [what] for [who], using only these facts: [list].
No opening line before the point, no summary at the end,
no lists of three, no rhetorical questions. Use contractions.
Where a detail is missing, write [NEED: what] instead of guessing.

The prompts on this page split that job into steps. The Voice Profile Builder does the first line properly, once, and gives you instructions to reuse. The others each fix one habit, so you can see what changed and keep what's yours. If you only want an existing text reworded for a new reader or length, our rewriting prompts are the better fit.

Which prompt for which job

One paragraph, from generic to yours

We wrote this example ourselves to show the steps. Picture a bike repair shop owner who asked a chatbot for a newsletter paragraph about new opening hours and got something like this:

In today's fast-paced world, keeping your bike in top condition is more important than ever. Our dedicated team is passionate about delivering exceptional service, and we're excited to announce some changes that will elevate your experience. Whether you're a casual rider or a seasoned cyclist, we've got you covered!

Nothing in it is wrong, and nothing in it is about this shop. Run it through the Filler Phrase Remover and almost every phrase goes: the opener, «dedicated», «passionate», «elevate», the whether-or pair. What's left is one fact, «some changes», and that fact is empty.

That's the job of the Concrete Detail Editor. It asks the questions only the owner can answer: which changes, from when, and what's in it for a rider? In our example the owner answers: open Sundays from May 4, 10 to 4, and a flat dropped off before noon is fixed the same day. Give those back, along with a couple of paragraphs the owner wrote for last year's newsletter, and the paragraph becomes:

We're opening on Sundays. From May 4, the shop's open 10 to 4, so you can get the bike sorted before the week starts. Drop off a flat before noon and you'll ride it home the same day. That's been true on weekdays all along, too.

Notice what made the difference. It isn't a clever word. It's two facts, a date and a promise, and a voice borrowed from the owner's own writing: short sentences, a dry aside at the end. The model supplied the structure; the owner supplied everything worth reading.

Follow-up lines when it still sounds off

You don't need a new prompt to fix a near miss. Reply in the same chat with one of these:

Where this goes wrong

Making text sound human is mostly subtraction, and subtraction has its own failure modes.

• Your voice turns into a caricature. Give the model a sample with one habit, say short punchy lines, and it may use that habit in every sentence. Read the profile the Voice Profile Builder returns and delete any line that describes a tic, not a style.
• The samples were already AI-assisted. If the paragraphs you paste were polished by a chatbot, you'll get that polish back. Use something you wrote fast, for a real person.
• Specifics get invented. Asked for detail, a model will produce a customer name or a percentage that sounds right. Every prompt here marks gaps in brackets; fill them yourself or cut the sentence.
• Removing a hedge changes the claim. «May reduce costs» and «reduces costs» are different promises. The Hedge Remover is told to leave legal, medical and financial qualifiers alone, but check the ones that matter to you.
• Banned words come back in disguise. Strike «delve» and you'll get «dig into» in the same spot. The tell is the sentence shape, so ask about shape when word lists stop helping.

A note on AI detectors

None of these prompts is built to change a detector score, and we wouldn't trust a prompt that claimed it could. Detectors guess, they disagree with each other, and they flag human writing as well as missing machine writing. Our AI detectors page explains how they work and where they fail. The safe ground is simpler: write the text yourself or with real input from you, make sure you can stand behind every line, and disclose AI help wherever your school, employer or client asks you to.

ChatGPT settings that help

Two settings change how much of this you have to repeat. The first is custom instructions, under Settings and then Personalization. Paste the instruction set from the Voice Profile Builder there and it applies to your chats without pasting it each time. As of October 2026, OpenAI lets Free and Go users save up to 1,500 characters there, and paid plans up to 5,000, so a short profile fits on any plan.

The second is the personality picker, labelled Base style and tone, with options such as Friendly, Candid and Efficient. It's tempting to pick one and expect drafts to change. OpenAI's own help page says that for writing like an email or a post, ChatGPT follows your instructions and the context, not necessarily the personality you chose. So the picker shapes how ChatGPT talks to you, and your samples shape how it writes for you. These prompts are plain text and work the same way in Claude, Gemini and other chat models; see our ChatGPT review for plans and limits.