How to Get ChatGPT to Write in Your Own Voice
The fastest way to get ChatGPT to write like you is to stop describing your voice and start showing it: paste three short things you actually wrote, then hand it two or three rules it can visibly break. Tone words do almost nothing. Samples plus checkable rules get you most of the way there in one prompt.
Below is the exact sequence I use, and a real test of it, including the part where the output went slightly wrong.
Why “friendly but professional” produces nothing
Tone adjectives are averages. Every person who types friendly but professional is asking for the same middle of the road, so that is what comes back — the beige paragraph with the warm opener and the tidy summary at the end.
Your voice is not a tone setting. It is a set of habits: how long your sentences run, whether you apologise, what you refuse to say, how you deliver bad news, whether you end on a joke or just stop. None of that survives being compressed into two adjectives. It does survive being shown.
Step 1: paste three things you actually wrote
Not your best writing — your ordinary writing. Three short samples work better than one long one, because three lets the model see what stays constant across different jobs. Pick samples that do different work:
- One where you deliver bad news or say no.
- One plain announcement — hours, dates, a change.
- One where you are pleased about something.
Fifty to eighty words each is plenty. Do not explain them, do not annotate them, and do not tidy them up first. The sloppiness is the signal. If you clean the samples, you will get clean writing back that sounds like nobody.
Step 2: give it rules it can fail
This is the part most people skip. “Keep it tight” is not a rule, because nothing can be measured against it. A rule you can check with your own eyes forces the model to restructure rather than decorate.
Rules that work because failure is visible:
- No sentence longer than 15 words.
- No exclamation marks.
- Never use these words: excited, delighted, journey, leverage.
- No bullet lists — write it as paragraphs.
- Do not add facts I have not given you.
Two or three is the sweet spot. Past that it starts writing to the constraints and drops the voice entirely, which is its own kind of robot.
What three samples and three rules produced
For the test I invented a bakery owner with a blunt, unsentimental way of writing — the rye is off because the starter went sideways, new hours are earlier than you want because that is what I can staff. Three samples, then one job: announce a two-week closure in October for a floor repair. Rules: no sentence over 15 words, no exclamation marks, do not use the word excited.
It held all three. The longest sentence is thirteen words. There are no exclamation marks and no excited. More usefully, it did not invent the dates I had not given it — it wrote “We’ll post the exact dates soon” instead of picking a week in October and stating it as fact. That is the right call, and not one I would have bet on.
What it got wrong is the last line: “Stock up before we close. We’ll be back when the floor is.” That second sentence trails off. It has copied the clipped rhythm of the samples without finishing the thought, which is what a sentence-length cap tends to produce — the model chops rather than rebuilds. It also opened its answer in a separate document panel I had not asked for.
Step 3: read the output against your own samples
Put the draft next to the three things you pasted and look for the seams. The reliable tells:
- Words you never use. One wrong word undoes a whole paragraph of good mimicry.
- Symmetry. Real people write lists of two and four. Models love threes, and love making each item the same length.
- The wrap-up sentence. A closing line that restates what was just said is almost never something you would have written.
- Clipped fragments, if you set a length cap — as above.
If you want a longer list of the giveaways, the same tells are what you look for when you are checking whether something else was machine-written: how to spot AI-written content.
Making it stick past one chat
Retyping three samples every time is how people give up on this. Two ways round it:
Custom instructions. In ChatGPT, Settings → Personalization → Custom instructions takes a block of text applied to every new chat. Paste one short sample and your two or three rules. Do not paste all three samples — the field is not big enough to hold them without crowding out everything else you want in there.
A project or a saved prompt. If your plan has projects, put the full set of samples in the project instructions and start every writing chat inside it. Claude and Gemini both have an equivalent. Otherwise keep the block in a note and paste it — it is five seconds.
Either way it drifts. Voice holds well for the first few exchanges of a chat and then slides back toward the middle, so when a draft starts sounding generic, re-paste the samples rather than arguing with it. And keep real client or employer material out of the samples if you have not checked what your account does with it — worth reading what actually happens when you paste work documents in before you use live examples.
The prompt, in full
Nothing clever in it — the work is in the samples you drop underneath.
Here are three things I have written. Match how I write, not the subject matter.
1. [sample]
2. [sample]
3. [sample]
Now write [the specific thing you need]. Rules: no sentence longer than 15 words; no exclamation marks; do not use the words [your banned words]; do not add facts I have not given you. If you need a detail I have not supplied, leave a blank in square brackets rather than inventing it.
That last line is worth keeping. Without it, a model asked for an announcement about dates it does not have will often supply plausible ones, and plausible-looking invented detail is the failure mode that actually costs you something. It behaved in my test, but I would not rely on it behaving every time — check the specifics in anything you send.
Where this falls down
It cannot give you a voice you do not have yet. If you have never written a newsletter, there are no samples, and what comes back is the average newsletter. In that case write badly for a few weeks first and then come back with material.
It degrades over length. Voice holds for a few hundred words and then thins out, so for anything longer, generate it in sections rather than asking for the whole thing at once.
It copies the habits you do not want along with the ones you do. If two of your samples open with “Just a quick note”, every draft will open with “Just a quick note”. Choose samples deliberately.
And it is the wrong tool when the writing has to be defensibly yours. A condolence note, a performance review, a resignation: the point of those is that a person sat down and wrote them. Use it to tidy your own draft instead.
The short checklist
Three ordinary samples, unedited. Two or three rules you can check by eye. One specific job, not “write something in my voice”. Then read the draft against the samples and fix the seams yourself. That last step is not optional, and it is also the reason the result ends up sounding like you: you are still the one deciding what stays.
Related reading
- How to Write Better AI Prompts — the general version of the “rules it can fail” idea.
- How to Draft Professional Emails in Seconds — where voice matching pays off most often.
- How to Turn Rough Notes Into a Polished Document — starting from your own words instead of a blank page.
Written by Mitch, a software analyst who tests software for a living. Every guide here comes from actually using the tool on a real task — including the parts where it falls over. Tested on ChatGPT Plus, Claude Max and Gemini (free). More about this site · Corrections: acheatsheet@gmail.com.
Get the next guide by email
A new tested how-to when there is one — usually a couple a month. No hype, unsubscribe any time.
Tested in real accounts. No affiliate links.
