The reason AI-generated content sounds generic is that it was asked to be. 'Write in a confident, friendly, professional tone' describes every brand and therefore none. A model can reproduce a voice reliably, but only if the voice is described the way a model can use it: as examples of the real thing, explicit rules, and a list of what never to say.

Examples over adjectives

Collect twenty pieces the company is proud of — posts, emails, page copy — and twenty it would never publish. The first set is the target; the second is what to avoid. Include them in the prompt or fine-tune on them. A model given real examples matches sentence length, rhythm, how paragraphs open and how much certainty to express, without any adjectives at all.

Rules that can be checked

Sentence length limits. Whether to use the first person. Which words are banned — 'leverage', 'seamless', 'unlock', 'game-changing', and the ones your industry overuses. Whether to open with a question. How to handle numbers, dates and product names. Every rule should be one a second model pass could verify.

What the company refuses to say

Claims it will not make, competitors it will not name, superlatives it will not use, jokes it will not attempt. Stated explicitly, these prevent the confident nonsense that damages trust and keep generated drafts inside what a human would have written.

Keep the guide alive

Every human edit to a generated draft is a signal: log the before and after. Monthly, look at the most common edits and turn them into rules or examples. The guide gets sharper, the drafts get closer, and the editing time drops — which is the entire return on the effort.