On this page
The best prompt to make ChatGPT write like a human does three things: it gives the model a real voice sample to copy, it bans the specific habits that make text sound machine-made, and it names the reader and the purpose. A generic "write like a human" instruction does almost nothing. Below is a prompt you can paste today, a worked example, and the edits that make the biggest difference.
The prompt to paste
This is the version we use as a starting point. Fill in the brackets, paste your voice sample, and run it.
You are helping me write [type of text, e.g. a cold email / a blog intro / a LinkedIn post]. Reader: [who will read this, and what they care about] Purpose: [what the reader should think or do afterwards] Length: [word count or number of paragraphs] Here is a sample of how I write. Match its sentence length, vocabulary, and level of formality. Do not copy its content. <sample> [paste 150-300 words of your own writing] </sample> Rules: - Vary sentence length. Mix short sentences with longer ones. - Use plain words. Prefer "use" over "utilise", "help" over "facilitate". - Be specific. Use concrete examples, names, and numbers I give you. Do not invent facts or figures. If you need one, leave a [placeholder]. - No openers like "In today's fast-paced world" and no closing recap of what you just said. - No rhetorical questions as openers. No lists of three adjectives. - Avoid these words: delve, leverage, tapestry, landscape, seamless, unlock, robust, crucial. - Contractions are fine. A fragment is fine if it reads naturally. - Take a clear position instead of hedging every claim. Write the text on this topic: [topic and key points]. After the draft, list any sentences where you were unsure about a fact so I can check them.
That's the whole thing. Each part has a job, and it's worth knowing which ones carry the weight so you can trim the prompt when you're in a hurry.
Why this prompt works
The voice sample does most of the work
Models imitate far better than they follow descriptions. "Write casually" means a hundred different things. A pasted paragraph of your own writing shows rhythm, punctuation habits, how long your sentences run, whether you start with "And" or never would.
Use something you wrote without AI help: an email you were pleased with, a post, a report intro. 150 to 300 words is enough. If you don't have anything, the prompt still works without the sample, but the output will land closer to generic.
Naming the reader changes the register
A text for a busy finance director reads differently from one for a friend. When you tell the model who is reading and what they should do next, it cuts padding on its own, because padding doesn't serve that reader.
The ban list targets the real tells
Most text that reads as machine-written shares a few habits: an opener that announces the topic, evenly sized sentences, triplets of adjectives, a closing paragraph that repeats the body, and a handful of favourite words. Name those and the model avoids them. Keep the list short, though. A list of fifty banned words makes the model contort its sentences around the list, and the result reads stiffer, not looser.
The "don't invent facts" line protects you
When you ask for a human voice, models tend to add confident specifics: a statistic, a quote, an anecdote that never happened. The placeholder instruction and the closing check turn that into a list you can verify rather than a risk you find after publishing.
Before and after: a real example
Here's the same task with a weak prompt and with the template. The task: a short intro for a newsletter about remote work.
The weak prompt:
Write like a human. Write an intro for my newsletter about remote work.
Typical output:
In today's fast-paced digital world, remote work has become an essential part of modern life. Whether you're a seasoned professional or just starting out, navigating the challenges of working from home can be both rewarding and challenging. In this newsletter, we'll explore key strategies to help you thrive.
Every sentence could appear in a thousand other newsletters. Nothing is specific, and the opener is the first thing a reader skims past.
Now the template, filled in:
You are helping me write a newsletter intro.
Reader: team leads at small companies who manage people they rarely see.
Purpose: get them to read this week's piece on running a 15-minute weekly check-in.
Length: 80 words.
Here is a sample of how I write. Match its sentence length, vocabulary, and level of formality. Do not copy its content.
<sample>
I used to think a good manager was always available. Then I spent a year answering messages at 9pm and getting nothing done in the day. Now I block two hours with notifications off, and my team knows when to expect me. It's less heroic. It works better.
</sample>
Rules:
- Vary sentence length.
- Use plain words.
- Do not invent facts or figures. Use [placeholders] if you need one.
- No openers that announce the topic. No closing recap.
- Contractions are fine.
Write the text on this topic: most remote teams don't have a meeting problem, they have a "nobody knows what anyone is doing" problem, and a short weekly check-in fixes more than extra calls do.Typical output:
Most remote teams don't have too many meetings. They have too little clarity. You find out on Thursday that two people spent the week on the same thing, and nobody's quite sure who asked for it. This week I'm covering a 15-minute check-in that fixes most of that. No agenda deck, no status theatre. Just three questions and a hard stop.
What changed: the opening makes a claim instead of announcing a topic, sentence lengths jump around, and the example ("two people on the same thing") is concrete. It also borrowed the short, slightly dry rhythm from the sample. You'd still edit it. But you're editing, not rewriting.
A version for essays
For students searching for a prompt to make ChatGPT write like a human for an essay, the sensible use is drafting help, not handing in generated text. Check your school's or employer's policy on AI first, since rules vary. If AI help is allowed, this prompt keeps your argument yours:
I'm writing an essay on [topic] for [course or audience]. My thesis is: [your thesis in one or two sentences]. Here is my outline: [paste your outline] Here is a paragraph I wrote myself, so you can match my voice: <sample> [paste a paragraph] </sample> Draft the paragraph for point [number]. Use only the evidence I list below. Do not add sources, quotes, or statistics. Keep the sentences varied and plain. Do not summarise at the end of the paragraph. Evidence for this point: [your notes]
This works because you supply the thesis, structure, and evidence. The model only handles phrasing. It also avoids the invented citations that essay prompts tend to produce.
Edits that matter more than prompt wording
No prompt gets you all the way. A few passes by hand usually decide whether the final text sounds like a person.
Cut the first sentence. Models love a warm-up line. Delete it and see whether the paragraph starts better.
Replace one generic phrase per paragraph with something only you know. A client's name, a number from your own data, a mistake you made.
Read it aloud. Anywhere you stumble or run out of breath, split or rewrite.
Delete the last paragraph if it repeats the rest.
A two-step approach also helps. Ask for the draft first, then run a second prompt that edits:
Here is a draft. Rewrite it to sound more like the voice sample below. Keep every fact and claim the same. Do not add anything new. Voice sample: <sample> [paste sample] </sample> Draft: <draft> [paste draft] </draft> Then list every sentence you changed in meaning, if any.
Separating drafting from voice-matching gives better control. The first pass handles content, the second handles style, and you can see exactly what shifted.
What to skip
The advice floating around on Reddit and TikTok often includes tricks like "tell ChatGPT to beat AI detectors" or "add random typos." We'd skip both. Detectors are unreliable, and no prompt can promise to pass them. Typos make writing look sloppy, not human. Aim for text a reader enjoys, and let the detectors argue among themselves.
Also skip very long rulebooks. Ten rules the model can hold in mind beat forty that contradict each other. If you want a deeper look at why shorter, goal-focused prompts tend to behave better, see Prompting in 2026: Write Goals, Not Rulebooks. Results also differ between models, so test on the one you use. For that, How to Get Consistent AI Results Across Claude, ChatGPT, and Gemini covers the differences.
Keep the version that works
Once a prompt produces text you'd actually send, don't leave it in a chat history. Save it with your voice sample attached and note what it's for. In PromptPal, you can store it in your prompt library, and when you tweak the ban list or the sample, version control lets you compare the new draft against the old one instead of guessing which edit helped.
If you write for a team, keep one voice sample per brand or per person. A shared prompt with the wrong sample makes everyone sound like the same stranger.
Next step
Copy the template above, run it with your own writing sample, and save the version that works. Then browse the public prompt library for writing templates by role and task to build from.
Stay Updated
Get the latest blog posts and AI prompt engineering tips delivered to your inbox.




