How to train AI on your brand voice so output ships without a rewrite

Why adjectives do not work

Telling a model to be "professional yet approachable" produces the average of everything ever described that way, which is precisely the texture people recognise as machine-written.

Voice is carried by specifics: sentence length, whether you use contractions, how you open, what you refuse to say, whether you lead with the claim or the context. Those are demonstrable rather than describable, which means examples beat instructions.

Three annotated examples of your real writing outperform three paragraphs describing how you write.

What actually goes into a voice system

• Twenty to thirty pieces of your best published writing, annotated with what makes each one right.
• A banned-phrase list. This does more work than anything else on the list, because most of the machine-written feel comes from a small set of recurring words.
• Sentence and paragraph rhythm, stated concretely: typical length, how much variation, whether fragments are allowed.
• ICP personas, so the model knows who it is speaking to and what they already know.
• Product and factual guardrails, including what must never be claimed.
• A worked example of a rewrite: a weak draft next to the corrected version, with the reasoning.

The website-echo test

Generate a piece on a topic your site already covers, then put it beside the existing page. If a reader could not tell which was which, the voice layer is working. If they can, the difference tells you exactly what is missing.

Run it before producing anything at volume. Finding a voice problem after four hundred pages is expensive, and finding it after ten is a Tuesday afternoon.

Keep a human gate, at least at first

Draft and approve, not generate and publish. Not because the output cannot be trusted, but because the approval step is where you learn what the system still gets wrong, and that feedback is how it improves.

Once the first-time approval rate is consistently high you can widen the gate. Starting wide and narrowing after an incident is a much worse sequence.

What good looks like

One client’s system produces social content the team publishes without rewriting it. The rewriting step was the actual bottleneck, not the writing, and removing it was worth more than doubling output would have been.

That is the bar. Not "sounds fine". Ships without a rewrite.