How many samples? Five
There is a measurement for this, and it is surprisingly clear. A 2025 study ran on six frontier models, four datasets and texts from more than four hundred authors, with over forty thousand generations per model. The setup was the same throughout: five writing samples plus a content summary of what the new text should be about.
The key finding concerns the number of samples: "including more writing examples in the prompt affects the four metrics very little". Raising it to ten produced no meaningful improvement. So the ceiling is not at the number of examples, and anyone loading in twenty articles is not doing it for the style but for their own reassurance.
From the other direction, though, the gain is enormous. Against zero — that is, merely describing your style with no samples — a handful of examples produced up to a more than twenty-fold improvement in style matching, and the same measurement found that the prompting strategy matters more than the size of the model. With instructions and no samples, models "default to an average, generic tone" — precisely what you are trying to avoid. This is few-shot logic applied to style.
Which five to pick
The selection matters more than the count. Five criteria, in this order.
- The same genre as what you are about to write. If you want emails drafted, supply emails. A mixed selection produces an average, and the average is what we are working against.
- Whole texts, not extracts. A clipped paragraph loses the opening and the close — the two places where style shows most.
- Something that is genuinely yours. Do not supply what an editor rewrote, or what you wrote to satisfy a template. The model will imitate what it sees.
- Something that worked in practice. A piece that got replies, that was picked up, that did its job. That is not a style criterion but insurance against fixing one of your bad habits in place.
- Not all written on the same day. Five texts from one week capture a mood, not a voice.
What not to do when selecting
Three tactics that look obvious and, by measurement, either do not help or actively hurt.
- Do not select by content. In the study cited above, samples chosen for topical similarity made the result worse. Style does not come from subject matter.
- Do not match the length to the target text. In the same measurement, length matching improved one metric while degrading the style model. Better to let the samples keep their natural length.
- Do not mix languages in one sample set. If you write in two languages, you need two sets. Sentence rhythm and register differ by language, and from a mixed set the model derives a middle that does not exist.
Fine-tuning is not the answer in 2026
This is the section that would have read differently two years ago. Fine-tuning for style is today not merely "not recommended" — at several major providers it is a service being withdrawn.
From May 2026 one of the largest providers stopped admitting new organisations to its fine-tuning platform; in July it cut off long-inactive customers as well; and by its own statement, from January 2027 existing customers will not be able to start new training runs either. Another large European provider already marks its fine-tuning documentation as deprecated. More telling still: on that first provider's own use-case list — classification, nuanced translation, output in a specific format, fixing instruction-following failures — style and tone do not appear at all.
It is still on sale in places, with published prices: around $25 per million training tokens on the larger model, $5 on the faster one, and a dollar and a half on the smallest. But that comes with a dataset, an epoch count, evaluation and maintenance — and the style you are after is available with five samples, for nothing. The decision logic is worked out in fine-tuning, RAG or a better prompt, and for style the answer is nearly always the third.
What exists instead
Four mechanisms that are genuinely available in 2026, all of them free or part of a subscription.
- Global custom instructions. At account level, for every conversation. The style sheet goes here — the rhythm and the banned list — which fixing your own voice sets out in detail.
- Project or workspace instructions. A different voice per client or brand, with the reference texts in the same place. This is the best home for the five samples.
- A named style built from your own samples. Several assistants let you upload a few of your own texts, from which the system builds a style you can save under a name. This is the closest thing to what you wanted from fine-tuning.
- Persistent memory. The system keeps what it has learned about you from conversations. Convenient, but the least reliable for style: you do not control what it retains.
Where it fails
Style imitation does not work equally across genres, and the spread is enormous. The same measurement reached 86 to 93 percent author identification on journalistic text and 95 to 97 on email — but 17 to 21 percent on a personal blog post and 50 to 66 on a forum comment. Email is therefore the best case, and email drafting is the place to start applying this. The rule is simple and unwelcome: the more structured and factual the genre, the better it goes; the more of you there is in it, the worse.
There is a deeper layer too. The same studies measured a split between surface features and the statistical shape of the text: the average "unpredictability" of human essays was 29.5, against 15.2 for machine text matched to the same style. Your turns of phrase come out right; the variation does not. In the authors' phrasing, stylistic fidelity and detectability are separate things.
And a third: machine assistance flattens on its own. An analysis of 880,000 texts published in August 2026 found that machine polishing cut the spread of writing complexity by 21 to 50 percent and erased the traces pointing to the author's age, gender and worldview. Your samples push against this, but they do not cancel it.
How to check it
Four steps, half an hour in total, after which you will know whether your sample set is worth anything.
- Have it write something you have already written. An old email or post that is not in the sample set. That gives you something to compare against.
- Put the two side by side. The question is not which is better but which is yours. If you cannot tell at a glance, the set is good.
- Give it to a colleague, unlabelled. Someone who knows your writing. This is the fastest and strictest test, and it is usually settled by the first paragraph.
- Look at what is missing. Do not hunt for what it wrote badly but for what it did not put in: the half-sentence aside, the short sentence for emphasis, your own example. That list becomes the next version of your style sheet.
What happens to the texts you upload
This rarely comes up, yet the sample set is your own — or a colleague's — past work. The 2026 ground rule is clear: on consumer plans the conversation falls within the scope of model development by default, and on business and API plans it does not. This is readable from the providers' own documentation, and it is switchable per plan.
Two details tend to surprise people. First: retention outlives deletion. At one large provider, conversations selected for human review are kept, detached from the account, for years even after you delete your history; at another, a conversation flagged with feedback can be kept for up to five. Second: a thumbs-up can count as consent to that conversation being used — meaning one click can override the setting.
The practical conclusion is not to avoid doing this but to look at which plan you are doing it on. If you are working with client texts or a colleague's writing, that belongs on the business tier. The territorial and legal side is unpacked in EU data residency.
What to take away
Five of your own samples, from the same genre, as whole texts — that is the recipe, and ten will not improve it. Do not select by subject matter, because that made results worse in measurements. Fine-tuning for style is a poor investment in 2026: several major providers are withdrawing it, and it does not even appear on their own use-case lists. Put the samples somewhere durable at project or style level instead. And expect email and professional writing to come out almost perfectly, while your personal voice does not — that part you still have to write in yourself.