How do you get ChatGPT to write in your own style?

Ask ChatGPT for a newsletter intro with no other guidance and you get something smooth, polite and oddly familiar: long sentences, a little throat-clearing, a tidy summary at the end. That is not a flaw you have to accept. It is the model's default voice, and you can replace it with yours.

ChatGPT is very good at imitation. The Verge's first look shows people asking it to debug code like a pirate or explain a sorting method like a wise-guy gangster, and getting convincing results. OpenAI's own launch examples include rewriting a short note to a neighbor in a more formal tone. If it can borrow a pirate's voice, it can borrow yours, provided you give it something to borrow from.

Why the default sounds the way it does

OpenAI trained ChatGPT partly by having people rank its answers. The launch post notes a side effect: the trainers tended to prefer longer, more complete-looking replies, so the model leans toward being wordy and repeats certain stock phrases. Left alone, that is the voice you get.

Two other habits matter here. The model guesses when your request is vague, and a request like "write it in my style" is about as vague as it gets: it has never read anything you wrote. And it remembers the earlier messages in the same chat, so samples you paste at the start keep shaping every reply that follows.

Telling versus showing

There are two ways to describe a voice, and they work best together.

Telling means a style description: short sentences, no exclamation marks, first person, one joke at most, Canadian spelling. This is precise about rules, but words like "warm" or "punchy" mean different things to different people.

Showing means real samples. They carry everything you did not think to describe: how you open, how long your paragraphs run, which words you never use. This is the same idea as few-shot prompting, where examples in the prompt teach a pattern better than instructions alone.

A worked example

Imagine you run a small bakery and write a weekly email to customers. Here is a prompt that shows and tells.

I write a weekly email for my bakery. Below are two short samples of my writing, then a description of my style. Write this week's email in the same voice.

Sample 1:
Rye is back. We took it off the shelf in August because the oven couldn't keep up, and you noticed. Six of you asked at the counter. It returns Thursday, still dark, still a little sour.

Sample 2:
Small news: we open at 7 now, not 7:30. Marc says the extra half hour is for the commuters. I say it's so he gets first pick of the croissants.

My style: short paragraphs, plain words, one small joke, no exclamation marks, no "we're thrilled to announce". I mention staff by first name.

This week: our holiday cookie boxes are available to order until December 20. Twelve cookies, four kinds. Pickup is in-store only.

Now compare the kind of draft you get with and without the samples. Both replies are illustrative, not transcripts.

The second column will not be perfect on the first try. That is fine: the next step is a correction in the same chat.

Closer. Cut the last sentence, it sounds like an ad. And I never say "treat yourself".

Each correction teaches it a little more about your voice for the rest of the conversation.

How many samples should you give it?

Two or three short samples are usually enough to start. More important than the number is the spread: pick pieces that show your voice in slightly different situations, such as an announcement, an apology and a casual note, so the model learns your voice rather than the topic of a single piece. If all your samples are about bread, expect it to talk about bread.

Keep each sample short, a paragraph or two of your natural writing, not your most polished work. Choose pieces you would be happy to see copied, because the model will pick up your habits, including the ones you would rather lose.

Avoid pasting anything private. ChatGPT is a public research preview, and OpenAI is collecting feedback on how people use it, so use samples you would be comfortable posting publicly.

If the drafts still sound off after a few corrections, add a line to your style description naming exactly what bothers you. Then start a fresh chat with your samples, your improved description and the request. A clean start with better instructions often beats a long chain of fixes.