On this page
The prompt that makes ChatGPT sound human is not "write like a human." It's a prompt that gives the model a voice to copy, a list of specific habits to drop, and a reader to write for. Below is a complete version you can paste, a worked before-and-after, and a way to keep it so you don't rebuild it every time.
The short answer: a prompt you can paste now
Here is the prompt. Fill in the brackets and use it at the start of a new chat, or paste it above whatever you want written.
Write [the thing: a cold email / a LinkedIn post / a project update] for [reader: who they are and what they care about]. Goal: [what the reader should think, feel or do after reading]. Voice: write the way I do. Here is a sample of my writing: """ [paste 150-300 words you wrote yourself] """ Rules: - Use contractions. - Mix short sentences with longer ones. Some can be five words. - Be specific. Use a real detail, number or example from the facts below instead of a general claim. - Do not open with a greeting line about hoping they are well, or with a line that praises the topic. - Do not end with a summary or a line that restates what you just said. - Skip words like "leverage", "robust", "streamline" and "tapestry". - No rhetorical questions as openers. No exclamation marks. - Plain words over fancy ones. Facts to use: [bullet points of what actually happened, who is involved, dates, numbers] Length: [number] words.
That's the whole technique. The rest of this post explains why each part is there, so you can cut what you don't need.
Why ChatGPT sounds robotic in the first place
When you give ChatGPT a vague request, it produces the average of everything it has seen on that topic. Average text is balanced, polite, evenly paced, and full of safe phrases. That's the "AI voice" people spot in two lines: the tidy three-part lists, the opener that praises your question, the closing paragraph that sums up what it just said.
So making it sound human means pulling it away from the average. You do that with three things:
A sample of real writing. This does more than any adjective.
Specific facts. Generic text comes from generic input.
A short list of habits to avoid. Concrete bans beat vague instructions.
Adjectives like "casual", "warm" or "conversational" are weak. The model reads them as a mild nudge and still produces its default shape. A sample paragraph of your own writing carries rhythm, vocabulary and attitude all at once.
A worked example: before and after
Task: a short note to a client whose project slipped by a week.
The weak prompt
Write a professional but human email telling a client their project is delayed by a week.
Typical output:
Subject: Update on Your Project Timeline
Dear Sam,
I hope this email finds you well. I wanted to reach out to provide an update regarding your project. Unfortunately, we have encountered some unforeseen challenges that will result in a one-week delay. We sincerely apologize for any inconvenience this may cause. Please know that we are committed to delivering high-quality results and appreciate your patience and understanding.
Nothing is wrong with it, and nobody would believe a person wrote it. "Unforeseen challenges" tells Sam nothing. The apology is stock.
The stronger prompt
Write a short email to a client named Sam, a marketing director who dislikes long emails and wants to know what changes for her. Goal: she knows the new date, why it moved, and that her launch plan is safe. Voice: write the way I do. Sample: """ Quick one. The designs are done, but I'd rather you saw them Thursday than rushed them today. Two things to look at: the pricing page and the footer. Everything else I'd leave alone. """ Rules: - Use contractions. Keep it under 110 words. - No "I hope this finds you well". No "unforeseen challenges". No closing summary. - Say the real reason plainly. - One apology, one sentence, no grovelling. Facts: - Delivery moves from 12 June to 19 June. - Reason: our developer found the checkout form failed on Safari and we won't ship it broken. - Her launch is 30 June, so there are still 11 days of buffer. - Next step: I'll send a staging link on 17 June for her to look at.
Typical output:
Subject: Delivery moves to 19 June
Hi Sam,
Your site is a week late. New date is 19 June, not the 12th.
The reason: our developer found the checkout form breaks on Safari, and I'm not handing you something that fails for a chunk of your customers. Sorry for the slip.
Your 30 June launch is still safe, with 11 days of room. I'll send a staging link on the 17th so you can look before it's final.
Talk soon,
What changed? The subject line states the news. The real reason replaced the vague one. The sentence lengths vary. And the email is shorter, because we gave it a word limit and facts to use instead of letting it pad.
Your outputs will differ, and the exact wording depends on the model and its version. The pattern holds: specifics and a voice sample beat "make it human."
The three levers, and which to pull first
Lever 1: give it a voice sample
Paste 150 to 300 words you wrote yourself. An email you were pleased with works. So does a post, or a message to a colleague. Don't use text that was itself written by AI.
If you don't have a sample, ask ChatGPT to describe your style from one first:
Read the writing sample below and describe the voice in 8 bullet points I can reuse as instructions. Cover sentence length, formality, how I open and close, words I favour, words I avoid, and how I use humour or opinion. Be specific, and quote short phrases from the sample as evidence.
"""
[paste your sample]
"""Save the bullets. You can reuse them in place of the sample when you don't want to paste it again.
Lever 2: give it facts instead of asking for ideas
Most robotic writing is the model filling space it has no material for. If you give it the date, the reason, the number, and the next step, it has nothing to pad with. If a draft still feels vague, the fix is usually more facts, not more tone instructions.
Lever 3: ban the habits, not the feeling
"Don't sound robotic" does little. "Don't open with a compliment, don't end with a recap, don't use three-item lists unless there are three items" works, because the model can check its output against each one.
Keep the ban list short, around six to eight items. Long lists get partly ignored, and sometimes the model swaps one cliché for another. When that happens, add the new offender to the list.
Fixing a draft you already have
Sometimes the text exists and sounds stiff. Rewriting it is a different job from drafting it, and it needs a different prompt, because you want the meaning kept and only the voice changed.
Rewrite the draft below so it sounds like me. Keep every fact and the order of the points. Do not add claims, numbers or examples that aren't in the draft. My voice sample: """ [paste 150-300 words] """ Changes to make: - Cut filler openers and closing summaries. - Replace abstract words with the plain version. - Break up any sentence over 25 words. - Keep contractions. Draft: """ [paste draft] """ After the rewrite, list anything you changed that affected meaning, in one line each.
The last line matters. Rewrites can quietly change a claim, and asking for a change list makes that visible. Read it before you send anything.
What to skip
"Act as a human." It adds nothing. The model already knows it's meant to produce human-readable text.
Long persona backstories. "You are a 42-year-old copywriter from Leeds who loves football" gives you cosplay, not voice. A real sample beats it.
Detector-evasion prompts. Prompts promising to beat AI detectors are unreliable, and detectors themselves make mistakes. Aim for text that reads well and is true, and edit it yourself.
Endless polishing in chat. If the third round of "make it more natural" isn't helping, the input is the problem. Go back and add facts or a better sample.
In every case, a quick human edit is the final step. Read it aloud. Wherever you stumble, rewrite that sentence in your own words. Two minutes of this does more than another round of prompting.
Keep your voice prompt so you don't rebuild it
The annoying part of all this is that the good version takes effort, and then it disappears into a chat history. A few habits help:
Split the prompt into a fixed part (voice sample, rules) and a variable part (task, reader, facts).
Save the fixed part somewhere you can reach from any chat.
When you fix a recurring problem, such as a new cliché the model keeps using, edit the saved version and note what changed.
In PromptPal, you can save the template to your prompt library, use the Chrome extension to refine a prompt inline inside ChatGPT, and keep earlier versions with prompt version control so you can go back if an edit makes the writing worse. If several people write under one brand, the same template can live in a shared team workspace so everyone gets the same voice.
If you want more on structuring the task part of the prompt, How to Write Better ChatGPT Prompts covers it, and Few-Shot Prompting explains why a pasted sample works as well as it does.
A note on different models
The prompt above is written for ChatGPT, but the approach works in Claude and Gemini too. How strictly each follows a ban list, and how long a sample it uses well, varies by model and version, so test your prompt on the one you actually use. Run the same task three times and compare. If two of three drafts still carry a habit you banned, tighten that line, for example by moving it to the top of the rules.
Next step
Pick one email or post you write often, build the prompt above around it, and save it. Browse the public prompt library for writing templates to adapt, then keep your own version in PromptPal so it's there the next time you need it.
Stay Updated
Get the latest blog posts and AI prompt engineering tips delivered to your inbox.




