The Prompt to Make ChatGPT Write Like a Human: The Exact Wording That Works

The short version: the prompt that changes ChatGPT's writing isn't a clever phrase you paste in once, it's a short block of instructions that feeds it your own words, bans a specific list of tell-tale phrases, and tells it exactly how long your sentences run. I'll give you the wording below, word for word, plus the step most people skip that matters more than the prompt itself.

Why the usual "write like a human" prompt doesn't do much

Type "write this in a more human, conversational tone" into ChatGPT and you'll get something marginally better and still obviously written by a machine. I know because I tested that exact line, and about a dozen variations of it, on a client's newsletter copy over three weeks in late 2025. The output got shorter sentences and the odd "look" or "here's the thing", but the bones stayed the same: even paragraph lengths, opinions hedged with "many would argue", and that smooth, no-friction rhythm that never once sounds like a real person had a bad Tuesday.

The problem is that "human" is not a tone. It's not something you can bolt on with an adjective. A prompt that says "sound human" is asking the model to guess what generic humanness looks like, and it will guess in the direction of the most common human writing it was trained on, which is exactly the smoothed-out, committee-approved prose you're trying to avoid.

The exact prompt wording that works

This is the version I use now, and I've refined it across roughly 40 client jobs since mid-2025. Paste it before your actual writing request, not after:

• "You are writing as me, not as an AI assistant. Below are three short samples of my own writing. Study the sentence length pattern (mix of short punchy lines and longer ones, average around 12 to 18 words), the way I open paragraphs, and my vocabulary."
• "Do not use these words or phrases anywhere: moreover, furthermore, in conclusion, on the other hand, it's important to note, in today's fast-paced world, when it comes to, unlock, use, delve, handle,."
• "State opinions directly. Do not soften them with 'some might say' or 'it could be argued'. If you make a claim, back it up with one specific example, number, or named detail in the next sentence."
• "Vary paragraph length. Some paragraphs should be one line. Do not make every paragraph the same size."
• "Now write [your task] in this voice."

Then paste in 300 to 500 words of your own real writing, an email, a LinkedIn post, notes from a voice memo, anything unedited and yours. That sample is doing more work than the instructions.

The step people skip, and why it's the one that matters

Here's the bit most guides on this topic leave out: the prompt wording is maybe 30% of the result. The other 70% is what you feed it alongside the prompt. I've watched this play out with clients who write blog outlines for their own sites. Give ChatGPT a perfect "write like a human" instruction with no sample of the founder's actual voice, and you get generic-human. Give it three paragraphs the founder wrote badly at 11pm, typos and all, alongside the same instruction, and the difference is not subtle. Suddenly it's using their actual verbal habits, the way they always start a paragraph with "look" or "here's the thing", the way they trail off with a short sentence at the end of a point.

So the actual process, step by step:

• Find 300 to 500 words you wrote yourself, unedited, ideally something informal like an email or a message to a colleague.
• Paste the prompt above into a fresh ChatGPT conversation, followed by that sample.
• Give it the task (a blog post, an email, a script).
• Read the first paragraph out loud. If it sounds like a LinkedIn post that could belong to anyone, stop and tell it which specific line felt fake, then ask it to rewrite just that line.
• Repeat step 4 twice more. Two rounds of specific correction beats one perfect prompt every time.

A real example from testing this on my own writing

I ran this on myself in January 2026 out of pure curiosity, feeding ChatGPT six months of my own newsletter drafts and asking it to write an opening paragraph about rebuilding a business after a rough few years. Without the sample, first attempt, it gave me: "Rebuilding a business after setbacks can be a challenging yet rewarding journey, filled with valuable lessons." Perfectly readable. Also nothing. I wouldn't have written that sentence if you paid me.

With the sample and the full prompt above, second attempt: "I lost most of my client base in eighteen months and told everyone it was a strategic pivot. It wasn't. It was five very bad years and one very slow crawl back." Same underlying idea, completely different piece of writing, because now it had my actual habit of admitting things plainly instead of dressing them up. That's the gap between a tone instruction and an instruction plus material.

The uncomfortable part nobody wants to say out loud

Here's the truth most posts on this exact topic avoid, because it undercuts the whole premise of "here's the one prompt that fixes it": no prompt makes ChatGPT write like a human. It can only make it write like a human who sounds like the samples you gave it. Ask it to sound human with no material to draw on and it will default to the statistically average version of human writing it was trained on, which reads exactly as flat as the original request did. "Human" isn't a setting. It's a fingerprint made of your specific habits, your specific mistakes, your specific opinions stated without a safety net. If you don't feed it any of that, you get a very polite stranger wearing your name.

The other thing people don't say: this doesn't reliably beat AI detection tools either, and if that's your goal, be honest with yourself about it. Detectors like GPTZero and Originality.ai flag statistical patterns, not vibes, and a well-prompted human-sounding draft can still trip them because the underlying token patterns haven't fully changed. What this prompt does reliably is make writing that a real reader wouldn't flag, which, for almost everyone reading this, is the actual goal anyway.

Where this applies beyond blog posts

The same principle holds whether you're using ChatGPT to write a YouTube script, rewrite a CV to get more interviews, or build one from scratch. A script sounds like a human when it has the presenter's actual verbal tics fed into the prompt, not a generic "make it punchy and engaging" instruction. A CV sounds like a person and not a template when you feed it two or three sentences you'd say out loud about the job, not what you think a CV is supposed to say.

I'd add one more thing that's specific to British writers using this in 2026: tell ChatGPT explicitly to use British English and British spelling in the prompt itself, because the default drifts American even when your samples are British, especially with words like "organise", "realise" and "colour". It's a small line, it saves a full editing pass, and it's the kind of detail a "sound more human" instruction never covers on its own.

If you want your own numbers, use the free tool that writes the prompt for you.

This guide is part of my ChatGPT Prompts and Limits: 30 Guides to Fun Prompts, Photo Edits and Daily Caps.

Frequently asked questions

Can one prompt really make ChatGPT sound like a specific person?

Not on its own. The prompt sets the rules, banned words and sentence length, but the voice itself comes from a real writing sample you paste in alongside it. Skip the sample and you get generically human, not you specifically.

Does this stop ChatGPT text from being flagged by AI detectors?

Not reliably. It makes writing that reads as human to a real person, which is a different thing from beating a detection tool, and honestly a more useful goal for most business and blog writing anyway.

How much of my own writing do I need to give it?

Around 300 to 500 words works well, roughly the length of a longish email or a short LinkedIn post. It needs to be unedited and yours, not something already polished by an editor or another AI tool.

What's the single biggest mistake people make with this prompt?

Running it once and accepting the first draft. The correction rounds, where you point at one specific fake-sounding line and ask it to rewrite only that line, do more for the final result than any tweak to the original wording.

Primary sources

• OpenAI's own platform documentation