Teaching a Model to Write Like Someone Else
Building reusable tools that learn a person's writing voice, and deciding where that should stop.
You believe your writing voice is yours alone. It is mostly a set of habits, and habits can be learned. I know, because I taught a model to counterfeit mine, then did it for colleagues, then built the thing that does it on its own.
It began as a favour. Someone wanted help making AI drafts sound like them rather than like a press release. Done by hand, prompt by prompt, that works exactly once and drifts back to generic the moment you stop babysitting it. So instead of fixing one post I built a small reusable tool that learns a voice and writes in it, and then a second tool that builds those tools from a person's own writing. The second one is the part I am still a little pleased with, and I will get to it.
Voice is a pattern, which is the good news and the deflating news
The reason this works is faintly humbling: a writing voice is largely a bundle of measurable tics.
Sentence length and how much it varies. Punctuation habits, whether you lean on commas, full stops, or the em-dash as a structural crutch. How you open and how you close. The hedges you reach for. The handful of phrases you reuse without noticing. The formatting reflexes, lists against prose, headers against none. Stack enough of them and you have a fingerprint, and a model can match a fingerprint if you describe it precisely and show it good examples. That is the good news for building the tool and the slightly deflating news for anyone attached to being a singular literary presence. You are, statistically, a fairly consistent set of moves.
“You are, statistically, a fairly consistent set of moves.” THE DEFLATING NEWS ABOUT VOICE
Why a portable instruction set beats fine-tuning
The instinct some people have is to fine-tune a model on someone's writing. For voice, that is the wrong instrument. Fine-tuning is heavy, awkward to update, and excessive when the thing you are capturing is style, not knowledge.
The lighter and better approach is to carry the voice in a portable instruction set: explicit rules plus a curated handful of strong examples, handed to a capable model at the moment it drafts. It is few-shot prompting, packaged so it is reusable and version-controlled instead of retyped each time. It updates instantly, it travels as a single file, and you can read exactly what it is doing, which you cannot say of fine-tuned weights. For something a colleague has to trust and tweak, legible beats clever every time.
So each voice became a small, shareable artifact: the rules of the voice, the do's and don'ts, and the examples that anchor it. Drop it in the right place and the assistant writes in that voice on demand.
The builder, or making the tool make itself
Building one of these by hand still takes time. You read a great deal of someone's writing, infer the patterns, write the rules. So the second tool removes that step: it reads a person's own sent mail, infers the voice, generates the personal tool from it, and then drafts new mail in their style.
A tool that builds tools is a satisfying place to arrive, and the moment the project stopped being a favour and became infrastructure. Help one person and you have helped one person. Build the thing that does the helping and you have helped everyone, once. The honest cost is that I have now read more of my colleagues' sent mail than is strictly healthy, in the name of science.
Where it goes wrong
It is not magic, and the failure modes are instructive.
The first is the uncanny valley: a draft that is ninety percent right and ten percent off reads worse than one that is generically fine, because the misses feel like impersonation rather than a clumsy email. The second is recency bias, where the tool over-indexes on whatever you wrote last and starts to sound like your worst recent week. The third, and the funniest, is that it will faithfully reproduce the verbal tic you would have edited out had you known you had it, and seeing it played back is a small, useful humiliation I recommend to everyone.
The part that actually takes judgment
Cloning a voice is easy. Deciding whose, and whether to admit it, is the real work, and it is the bit I would refuse to skip.
Cloning your own voice is uncontroversial. Cloning someone else's, even helpfully, raises questions worth being deliberate about. Consent is the obvious one: people should know a tool exists that writes in their voice, and opt into it rather than discover it. Ownership is murkier than it first appears, because a voice is personal in a way a template is not. And there is a line I will not cross, which is anything that misleads a reader about who actually wrote something. A tool that helps you draft faster in your own style is fine. A tool that lets one person impersonate another is not, and the technology does not care which one you are building, so you have to.
That is the article, really. The engineering is a weekend. The decisions worth making are about people, which is true of most of this once the novelty wears off. ▮