Most people ask AI to write in their voice by describing it. They type something like professional but warm, or conversational, not too formal. Then they read what comes back and do not recognise themselves in a single line of it.
The description is not the problem. The problem is that a description is a summary of your writing, and the model needs the writing.
What goes wrong when you describe your voice
A model has read an enormous amount of writing that was described as professional but warm. When you hand it those words, it goes and finds the middle of all of it. What comes back is the average of a million people trying to sound professional but warm, which is exactly the thing your readers would never mistake for you.
Think of it like ordering a cake over the phone. You can say chocolate, not too sweet, nothing fancy, and you will get a chocolate cake. You will not get your grandmother’s chocolate cake. The only way to get that one is to hand over the recipe, or better, hand over a slice.
The instruction that cannot work
There is a line that circulates in prompt templates, and it reads something like: never write anything I would not say. It sounds careful. It is also empty, because the model has no idea what you would not say. It has never read you. You are asking it to avoid a list it does not have.
The same goes for most negative instructions. Do not be cheesy. Do not use jargon. Avoid fluff. Each one asks the model to steer away from a thing it cannot locate, and research on prompting has landed in the same place repeatedly: positive instructions outperform prohibitions. Tell it what to do. Short sentences. One concrete example for each point. No summary at the end. Those it can act on.
Give it evidence instead
Find three to five passages you actually wrote. Not your best work, and not something you polished for an hour. Ordinary writing, in the format you want back. If you want emails, give it emails. If you want captions, give it captions.
Paste them in. Then ask the model to study them before it writes anything: the rhythm, the length of the sentences, where they break, the words you reach for and the ones you clearly never use. It can see all of that directly. It does not have to guess, and that is the whole difference.
The prompt
This is the one I use. Copy it, replace the bracketed part with your own writing, and paste the draft underneath.
Below are three passages I wrote myself.
[paste your writing here]
Study them before you write anything. Look at the rhythm, the sentence lengths, where the sentences break, and the words I actually use. Notice what I do not do.
Now rewrite the draft below in that voice. Keep every fact exactly as it is. Do not add anything I did not say. If a sentence of mine is long, let it stay long.
[paste the draft here]
Notice there is no adjective anywhere in it. No warm, no professional, no engaging. The samples carry all of that, and they carry it more precisely than any word you could pick.
What the difference looks like
Here is a paragraph a model wrote for me when all I gave it was a topic and the word professional.
In today’s fast paced world, AI is not just a tool, it is a partner. Leveraging it is crucial for businesses seeking to delve into their full potential.
Read that out loud and listen for who is speaking. Nobody is. It has the shape of a point without making one, it reaches for delve and leverage and crucial inside two sentences, and it opens on a phrase that has been printed on roughly every slide deck since the invention of slide decks.
Then I pasted in three paragraphs of my own writing and asked for the same idea again, with the instruction above. What came back was this.
Most people use AI badly. Nobody ever showed them how.
Two sentences instead of two clauses stacked on two more. No adjectives doing the work. The claim is narrower and it is actually a claim, which means a reader can disagree with it. That is not the model becoming cleverer in the eleven seconds between the two attempts. It is the same model, with evidence instead of adjectives.
The honest caveat: it did not get there on the first try. It took two rounds, and on the first one it kept a sentence of mine and buried it in the middle of something longer. Telling it which sentence to keep and asking again was enough. Treat the first draft as a starting position rather than an answer, and this stops being frustrating.
Which model to use for this
This is the part people skip, and it changes the result more than the prompt does.
When you open an AI app, there is usually a small menu somewhere with names in it. Claude Sonnet. Gemini Flash. Those names look like nonsense until you see that they are only two parts: a size, and a version. One part tells you how big the model is. The other tells you how recent it is.
Every company sells a small one, a medium one and a large one, and every company keeps shipping newer versions of each. The small ones answer quickly and cheaply. The large ones think harder and make you wait.
Matching a voice is a job for a large one. It is a comparison task: hold five passages in mind, work out what they have in common, then apply that to something new. A small model will produce something fluent and generic, which is the exact failure you were trying to fix. Most apps hand you the small one by default, so this is worth checking once.
One thing that trips people up here. Claude Code and Codex are not models. They are tools that drive a model for you, the way a car is not an engine. If you see those names, you are looking at the wrong menu.
Save it so you are not doing this every week
Pasting your samples into every new conversation gets old by the third time. Both Claude and ChatGPT let you save a set of instructions that applies to every conversation inside a project or a custom assistant. Put your samples and the instruction in there once. After that you just ask for the thing you want and it already knows how you sound.
This is also where a short list of your own banned words earns its place. Mine has delve, leverage, seamless, robust and tapestry in it, because those are the words that give a draft away before I have finished the first paragraph.
Where it stops
It will match your rhythm. It will stop reaching for words you never use. On a good day you will read a paragraph back and not be able to tell which of you wrote it.
What it will not do is have your opinions. It can arrange your sentences, but it cannot notice the thing in your week that was worth writing about. That part does not transfer, and no amount of sample text will make it transfer.
Which means the work moves rather than disappears. You stop spending your afternoon fixing sentences that sound like a press release, and you start spending it deciding what is actually worth saying. For most people that is a harder afternoon than the one they were having before.
Forward → Upward ↑ Onward ↗︎
Mstimaj
Sources and Further Reading
- How to Make Claude (and other AIs) Write Like You
- How to Train AI to Write in Your Voice: Prompting Guide for Creators
- How to Get AI to Write in Your Voice: The Voice Brief Method
- Jodie Cook, ChatGPT Prompts That Write LinkedIn Posts In Your Unique Voice, Forbes
Say something back
Comments go through our own form and are read before they appear. Links are not allowed; describe it in words.









