How to get AI to write in your voice
Five ordinary paragraphs of your own writing beat a page of adjectives. The samples method, saving it once, and the edit pass that is still yours.
The instruction almost everybody tries first is a list of adjectives. Warm but professional. Conversational without being stiff. The draft still reads like a company announcement, and the reasonable conclusion to draw is that the model simply cannot do voice.
It can. The adjectives fail for a duller reason than the model being bad at writing: they describe a voice the way a stranger would, and the description leaves out everything that made the voice yours. Your voice is a few hundred small decisions you make without noticing. How long your sentences run before you let one go short. Whether you start with the subject or wander in through a clause. Which words you would never type. Adjectives cannot carry any of that, so the model fills the gaps from the enormous average of everything it has read, and the enormous average of everything ever written sounds like a company announcement.
What follows is the general method: how to actually give a model your voice, how to stop handing it over from scratch every session, how to get the machine residue out of what comes back, and how to find the model whose untouched register already sits nearest yours. Most of it works on the free plan. The part that does not is named plainly when we reach it.
Five paragraphs of yours beat a page of adjectives
Paste five real pieces of your own writing before you ask for anything. That is the whole trick, and it takes about two minutes once.
Five is the number because one sample is a trap. Hand over a single paragraph and the model copies its surface: the topic, the shape, sometimes whole phrases. Five pieces on five different subjects leave only the thing they have in common, which is you. The model has to infer the pattern rather than imitate the instance, and the pattern is what you are trying to transfer.
They should also be ordinary. There is a strong pull towards submitting the best thing you have ever written, and it produces a draft in the voice you use when you are trying hard, which is not the voice you need for a Tuesday email. Pick the boring ones. Two messages you sent to colleagues, a paragraph from something you published, a note explaining a decision, the last long thing you wrote to a friend. Somewhere between 150 and 300 words each is plenty.
Keep the register consistent, because this is where the method most often goes wrong. Mix a work email with a jokey message to your brother and you have asked the model to average two voices, and averaging is exactly the failure you started with. If you want work writing, give it only work writing. Build a second set later for the other register.
What samples carry that a list of adjectives cannot is the negative information, and that is most of what makes writing recognisable. An instruction can only tell a model what to do. Five samples show it what you never do: that you do not use exclamation marks, that you have never once written "delve", that you say "we" when others would say "I", that your paragraphs open cold instead of announcing themselves. Nobody can write that list about their own writing. Everybody can produce evidence of it.
One step earns its place before you ask for a draft. Ask the model to describe your voice back to you in five or six lines, and read the description. If it says something bland, your samples were too short. If it describes two different people, your samples were mixed and you should cut one. If it names a habit you recognise and were slightly embarrassed by, the transfer has worked and you can ask for the draft.
The model whose ordinary voice sits nearest yours
Every model arrives with a house register. One reaches for structure, subheadings and numbered lists at the smallest provocation. Another writes long, warm, slightly over-explained paragraphs, while a third clips everything to the bone. This default does not vanish when you hand over samples. It pulls against them the whole way, which means the easiest improvement available to you is choosing a model whose untouched voice already sits somewhere near your own, before you spend any effort at all on prompting.
Test it on one paragraph. Paste your five samples, give a short brief for something you genuinely have to write, and read the draft. Then change the model in the picker and send the same request again. Changing the model does not start you over: the thread keeps your samples and your brief, and only the writing hand changes. Two or three models is the right number to try. There is no view that puts the drafts beside each other, so you are scrolling back up the thread to compare, and that is fine for three. If you would rather keep the answers apart, branching opens a new thread with the conversation carried over, which is useful for keeping a version you liked but slower to read across.
Give each model the identical brief. The temptation is to improve the wording on the second send, and the moment you do that you are testing your brief rather than the model. Compare the openings first, because the first two sentences show a house register faster than anything else, then compare how long the sentences run. Quality is the wrong axis here and it will mislead you.
We are not going to tell you which model wins, because the question has no general answer. The model that best matches a clipped, plain register is the wrong recommendation for somebody who writes long and digressive, and the draft that is objectively best written is very often the one least like you. The only judgement that matters here is which draft you could send without editing the rhythm. There are 59 models in that picker, from 17 providers, and 22 of them open on the free plan, so the test costs ten minutes and nothing else.
Saving it once, instead of pasting it every Monday
The samples method has one real cost, which is that you do it again next time. Five samples is maybe 1,200 words of pasting, and those words sit in the model's working space alongside the thing you are actually trying to write, competing for the same room. Our explainer on what a context window is covers why that shared budget matters on long jobs.
So save it. Tell the chat to remember how you write and it stores a short note. One detail catches people out here: that note attaches to the conversation you were in and goes no further. To get it into everything you start from then on, open Settings, then Memories, and star the note. The tooltip on that button reads "Include in all new chats by default", and it confirms with "Memory will be included in all new chats", after which your register travels with you.
A memory is short, though, so it holds a description of your voice rather than the evidence for it, and descriptions are the weaker input. So use both: let the saved memory carry the ordinary jobs, and paste the full sample set for anything that will be read carefully. On Pro and Pro+, five samples kept in a folder's Files section reach every chat in that folder with no paste, though they still take their share of the room.
Memories are a paid feature and we would rather say so here than let you find out at the toggle. The free plan holds none at all, Pro holds 200, and Pro+ has no cap. In the chat itself the control is labelled "Chat memories (Pro)".
On any plan, including free, there are two boxes under Settings, then Account, that apply to every conversation you start. "Personal preferences" is a 5,000-character field, described in its own tooltip as instructions "you always want applied", with the line underneath reading "What should MultiChats consider in responses? Your preferences will apply to all conversations." That is room for a description of your register and one or two short samples, which falls short of the full method and still beats adjectives typed fresh each time. "Traits" sits below it, and its own line reads "Up to 50 traits, max 100 chars each", a cap the server enforces. It is the place for the short register words you would otherwise repeat.
Approach | What comes back | What it costs you each session | Where it lives |
|---|---|---|---|
Instructions only: a list of adjectives | A competent register belonging to nobody in particular. Formality usually one notch above yours | Retyping the same paragraph of adjectives, and re-explaining it when the draft misses | Any plan, typed into the message |
Samples only: five real pieces | Close on sentence rhythm, vocabulary and the things you avoid. Needs a light edit rather than a rewrite | Finding and pasting roughly 1,200 words, every new conversation, plus the room they occupy | Any plan, pasted into the message, or partly into Personal preferences |
Samples plus a saved memory | Same as samples on a considered piece. Ordinary jobs start close with nothing pasted at all | Nothing after the first time, until your writing changes and you edit the note | Settings, then Memories. Pro and above |
Hand the draft back and ask what drifted
Even a good sample set leaves residue, and the edit pass is the part that stays yours. Work it in two readings: once for whether the thing is true and says what you meant, then once for texture.
Our guide to using AI on a job application keeps the full table of machine-writing tells, each with something to do about it. Rather than repeat it here, take three that turn up specifically when a model is imitating a person.
The first is that nothing is a fragment. Every sentence arrives whole, subject and verb in place, properly closed. People do not write like that, least of all in their own voice. If a 600-word draft contains no sentence shorter than six words, a person did not write it in a hurry, and hurry is audible in most of what you actually send. Break two.
The second is the connective bolted to the front of a paragraph. Moreover. Furthermore. Additionally. Ultimately. A model reaches for these to signal a relationship between two paragraphs that it has not established in the writing. Delete the first word of every paragraph and check what was lost. Usually nothing was.
The third is the doubled hedge: "can often", "may sometimes", "generally tends to", "is typically likely to". Two hedges do less work than one, because a reader stops believing the sentence is claiming anything. Keep the stronger word, cut the other.
Then a move that uses the model against its own draft. Paste one of your original samples back alongside the new text and ask which sentences sit furthest from it, and why. That produces a short list of specific lines rather than a general verdict, and it catches drift you have gone blind to after three reads. What it will not do is fix them well, so make the edits yourself.
The instruction to avoid is "make this sound more human", which reliably returns contractions, a rhetorical question and an emoji. Name the change you want: shorter sentences, no adverbs before verbs, no opening question.
Where does this stop working?
In a few places, and each one is worth knowing before you invest an hour in this.
Voice is partly a matter of what you choose to say, and no prompt reaches that. A draft that matches your rhythm exactly while raising a point you would never have raised reads wrong to anybody who knows you, and it will read wrong in a way they cannot name, which is worse. The judgement about what goes in stays with you.
The match also weakens as the task moves away from your samples. A set of work emails does very little for a wedding speech. When the register genuinely changes, the honest answer is to gather five samples of that register or accept that you are writing this one yourself.
And a good voice match makes invented detail harder to catch, because the wrong number now arrives in your own cadence, sounding exactly as confident as the things you know are true. Check the facts at the same rate you would check a draft that sounded like a stranger.
One last thing before you paste a folder of your own writing into anything. We do not train on your conversations. What the provider behind a given model does with them varies by model, and the models where we know of a difference carry a notice on their info card in the picker.
Start with five ordinary things you have already written, ask the model to describe the voice in them, and see whether you recognise yourself in the answer. That test costs nothing on a free account. Saving the result is the part that needs a paid plan.