Prompts, voice profiles, or fine-tuning: how to keep AI copy in your brand voice
AI copy drifts into a generic voice fast. Compare the three ways to hold your brand voice in AI writing — prompt examples, a stored voice profile, and fine-tuning — and which to reach for first.
Your visuals can be pixel-perfect on brand while the caption underneath sounds like everyone else's AI. Keeping AI-written copy in your brand voice comes down to three mechanisms: describe and demonstrate the voice in the prompt, store it as a reusable voice profile, or fine-tune a model until the voice is baked in.
The short answer, before the detail: build a voice profile and pair it with real examples of your copy in every request. That combination gets most teams to publishable without training anything. Fine-tuning holds a voice tightly at high volume, but it demands a clean dataset, and the industry's own tooling is quietly moving away from it as the default answer.
We've written about locking down the visual side in reference images, brand kits, or fine-tuning. This is the same decision for words, and the trade-offs land in different places.
What are the three ways to keep AI copy in your brand voice?
Prompting means putting the voice into each request: a role, instructions, and above all examples. Anthropic's prompting guidance is blunt about the ranking: examples are one of the most reliable ways to steer a model's tone, and 3 to 5 well-chosen ones outperform paragraphs of description. A role in the system prompt ("you write product copy for a plainspoken hardware brand") shifts tone with a single sentence.
A voice profile is that same material written down once and reused everywhere, instead of living in whoever wrote last week's prompt. It is the verbal half of a brand kit: tone rules, vocabulary, banned words, formatting habits, plus a small library of canonical example copy. Every generation reads it automatically, so the intern's product description and the founder's launch email draw from the same source.
Fine-tuning trains a model on hundreds or thousands of examples of your copy until the voice no longer needs explaining. OpenAI's model optimization guide lists exactly this use case (its preference-tuning method is aimed at generating messages "with the right tone and style") and notes the real advantage: you can train on far more examples than fit in any single prompt, then run with shorter prompts at lower cost.
The three methods at a glance
| Prompting | Voice profile | Fine-tuning | |
|---|---|---|---|
| What it is | Voice instructions + examples per request | Stored voice spec read on every run | A model trained on your copy |
| Setup | None, write and go | Codify once, reuse | Curate hundreds of examples, train |
| Consistency across people | Low: every prompter improvises | High: one source of truth | High: baked in |
| Best at | Quick wins, one-off pieces | Team-wide consistency, no training | Very high volume in one tight register |
| Where it breaks | Voice lives in individual prompts | Only holds what you can write down or exemplify | Dataset quality, retraining, platform churn |
| Examples needed | 3 to 5 per request | 3 to 5, stored once | Hundreds, all clean |
Prompting with examples: the fastest way to sound like you
If you do only one thing, paste real copy into the prompt. Models imitate what they see far more faithfully than they follow adjectives, which is why "friendly but professional" produces the same beige output for every brand that asks for it. Three of your actual paragraphs carry the sentence rhythm, the word choices, and the jokes you don't make: things no adjective list captures.
The role line matters too. Anthropic's docs note that setting a role in the system prompt focuses behavior and tone even at one sentence long. Stack them: role first, a few rules, then examples marked clearly as examples.
The weakness is organizational, not technical. Prompting keeps the voice in people's heads and chat histories. Two marketers on the same team will assemble different prompts, get different voices, and neither version improves the other's next attempt. Fine for a solo writer; leaky for a team.
A voice profile: your brand voice, written down once
A voice profile moves the prompt's contents into shared infrastructure. The test of a good one is whether someone who has never read your blog could generate a passable paragraph from it.
The most useful framing we know comes from Nielsen Norman Group's research on the four dimensions of tone of voice: every piece of copy sits somewhere on four spectrums: formal vs. casual, serious vs. funny, respectful vs. irreverent, and matter-of-fact vs. enthusiastic. Their survey work found that shifting copy along these dimensions produced statistically significant differences in how users rated a company's friendliness and formality, from changes as small as swapping "we apologize" for "we're sorry." Readers register tone and adjust their impression of you accordingly, whether or not you chose it.
NN/g also recommends choosing anti-tone words: the qualities you refuse to sound like. "Authoritative but never pedantic" does more work in a voice profile than five positive adjectives, because it names the specific failure you're steering away from. Generic AI copy is exactly that kind of failure: a default nobody chose.
A workable profile for AI, then: your position on the four dimensions, a handful of target tone words, your anti-tone words, banned vocabulary (every brand has words it never says), mechanical rules (contractions, sentence length, how you write numbers), and 3 to 5 canonical examples. Store it where generation happens, in a brand kit that text steps read automatically, rather than in a PDF nobody opens. We covered why the written-down format matters so much in machine-readable brand guidelines.
Fine-tuning for voice: when it earns its keep
Fine-tuning is the strongest hold and the biggest commitment. Train on enough of your published copy and the model produces your cadence without being shown examples each time, with shorter prompts and lower per-request cost at scale. For a support team answering thousands of tickets in one register, or a catalog operation writing ten thousand product descriptions, that math can work.
Three caveats before you commit. First, the dataset is the product: mixed-quality training copy teaches the model your bad habits alongside your good ones, at scale. Second, a tuned model freezes a moment in time: reposition the brand and you retrain, and the work is tied to a base model that will eventually be deprecated. Third, the ground is shifting under the technique itself: OpenAI's optimization guide now states it is winding down its fine-tuning platform, with new users no longer able to access it. Frontier models have become good enough at following prompts and examples that the vendor's own recommended order is evals first, prompting second, fine-tuning only where a measured gap remains. Betting your brand voice on the third step of that list deserves hesitation.
Which should you reach for first?
Write the voice profile first. It costs an afternoon, it upgrades every prompt anyone on the team writes from then on, and it is the prerequisite for everything else, because you can't curate a fine-tuning dataset for a voice you haven't defined.
Pair it with examples immediately, because the profile states the rules and the examples carry the music. In our experience this combination is where "sounds like AI wrote it" turns into "sounds like us on a fast day."
Consider fine-tuning only when three things are simultaneously true: volume is high, the register is narrow, and you already have hundreds of clean, on-voice examples. If any of the three is missing, a profile plus examples is faster, cheaper, and easier to change your mind about.
And wherever you land, keep a human read before anything ships, because voice drift is exactly the kind of failure that's obvious to a person and invisible to the model that produced it. Our review checklist for AI content covers what to look for. For the wider system that voice fits into, alongside color, imagery, and motion, start with the complete guide to on-brand AI content.
Your brand kit is one URL away.
Paste your site and Orisu builds the kit — colors, fonts, voice, logo — then holds every generation to it.
Common questions.
How do I make AI write in my brand voice?
Two moves cover most of it: write a short voice profile that states your position on formality, humor, reverence, and enthusiasm plus a few banned words, and attach 3 to 5 real examples of your published copy to every request. The profile sets the rules; the examples carry the rhythm rules can't describe.
Do I need to fine-tune a model to match my brand voice?
Rarely. A voice profile plus example copy gets most teams to publishable, and fine-tuning only pays off at high volume with a large set of clean, on-voice training examples. Vendors themselves treat it as the last resort: OpenAI's own optimization guide orders the workflow as evals, then prompting, then fine-tuning where useful.
What should a brand voice profile for AI include?
A useful profile states where you sit on the four tone dimensions from Nielsen Norman Group's research — formal vs. casual, serious vs. funny, respectful vs. irreverent, matter-of-fact vs. enthusiastic — plus a handful of target tone words, the anti-tone words you refuse to sound like, banned vocabulary, and formatting rules like sentence length or how you write numbers.
Why does AI copy sound generic even with a style guide in the prompt?
Usually because the guide describes the voice in adjectives instead of showing it. Models imitate examples far more reliably than they obey descriptions, so a prompt with three real paragraphs of your copy typically beats a page of tone adjectives. If the examples are there and it still drifts, the examples are probably inconsistent with each other.
