Ask any AI to "write this like me" with nothing else to go on and it will guess. The guess is usually a polite, even, slightly formal voice that belongs to nobody. Getting your own voice back takes a few things it cannot guess.
1. Give it real samples
The single most useful thing is two or three things you have actually written to the same kind of person. Not your best writing. Your normal writing: a text to a friend, a quick email to a colleague. The model copies what it sees far better than what it is told.
2. Say who it is for
"Write to my landlord" and "write to my sister" should produce different messages, even saying the same thing. Tell it who the reader is, how well you know them, and what you usually call them.
3. Give it the facts, all of them
When a model is missing a fact it tends to invent a plausible one: a day, a reason, a name. That is one of the fastest ways to make a message sound wrong. Give it the date, the number, the reason, and say to leave out anything you did not give it.
4. Ask for one version, and only the text
Say "just the message, no options, no explanation". Most general assistants default to offering choices and wrapping the answer in commentary. You then have to cut all that out before you can use it.
5. Cut what you would never say
Read the draft aloud. Cut the words you do not use. The common ones are easy to spot: a cheerful opener, a word like "delve", a closing line that offers more help. Then send it.
A prompt that works
Here is a starting point you can paste into any model:
"Write the message I'll send to [who they are to me]. Here are two things I've written to people like them: [samples]. The facts: [facts]. Only use these facts. Match my length and tone. Give me only the message, one version, no preamble and no notes."
Or use a model built for it
Hemmingway does most of this for you. People holds who each person is and how you write to them. Memory holds the facts it should keep in mind, and you can read and change every line. And the model, Hemmingway-1, was trained to give you the message itself, one version, in the register the situation calls for. You write "tell Sam I can't make Thursday, offer next week", and you get the text you would have sent.
In our tests it came first of the models we compared at everyday messages, and a judge picked its replies as the ones a person wrote more often than any other model's. How that was measured.