A marketing manager has a language model write four blog articles. She reads them back and can't quite point at what's wrong. The sentences are fine. The information is correct. The structure works. And yet any competitor in the sector could have published these texts. So she rewrites everything anyway, and the time saved is gone.
That's rarely the model's fault. Usually it's the input. The brandbook that goes with it was written for humans: a brand promise, three core values, a colour palette and half a page about being "approachable but professional". A colleague fills in the rest with experience and instinct. A model has no instinct.
Why a classic brandbook breaks down on a language model
Most brand documents consist of adjectives. Reliable, innovative, human, sharp. For a person that's a direction. For a model it's noise. "Professional" leads one model to formal, bureaucratic prose and another to a LinkedIn-style tone with plenty of white space. Both are defensible. Neither is necessarily your brand.
Language models work with patterns and examples. They are good at imitating and poor at interpreting. Which means a brandbook for AI has to be more concrete than a brandbook for your agency. Not more abstract, not longer. More concrete.
There's a brand interest behind this too. The moment AI produces part of your content, your tone scales faster than your editing can correct it. A vague brandbook spreads that vagueness across your blog, your nurture emails and your ad copy. Recognisability is exactly what you need to rise above the generic pile, and it's also what carries your conversion. People buy from brands they recognise.
What belongs in it when AI co-writes
These are the components that, in practice, separate usable output from rewriting work:
- A positioning in one paragraph, without metaphors. What you do, for whom, and how you differ from the three parties you pitch against.
- Your reader, per segment. Job title, what they already know, what they don't, and who they have to convince internally.
- A word list. The terms you use, the terms you avoid, and the explanation you give for jargon. Include the internal words that must never go public.
- Sentence rules. Average sentence length, maximum paragraph length, how you address the reader, and whether you write in first person singular, plural or third person.
- A banned list. Clichés, opening lines, superlatives and punctuation you won't accept.
- Good and bad example texts, side by side, with one line explaining why one works and the other doesn't.
- Channel rules. A LinkedIn post, a product page and an email call for different lengths and different openings.
- A claims framework. What you may assert without a source, what needs backing, and which statements about results are off limits.
That last point is the one most often forgotten, and the most expensive. Models fill gaps with plausible figures. If you never state that percentages, studies and client names may only come from an approved source list, sooner or later you'll publish something that isn't true.
Capturing tone of voice so a model can follow it
A tone of voice made only of adjectives can't be transferred. This route works better and takes roughly half a day:
- 1Collect eight to ten of your own texts that feel right. Emails, proposals and internal notes count. The real tone often lives there, not on the website.
- 2Mark what makes each one good. A sentence, a word choice, the way something is conceded or bounded.
- 3Turn every mark into a rule in the imperative. Not "we are direct", but "put the conclusion in the first sentence and the nuance behind it".
- 4Add an example and a counter-example to each rule. Two sentences will do. The counter-example usually does more work than the example.
- 5Have the model write the same piece twice: once without your rules, once with. Read the difference out loud.
- 6Delete every rule that doesn't change the outcome. A five-page brandbook that gets used beats thirty pages nobody pastes into a prompt.
—A tone of voice you can't test is an opinion. A tone of voice you can demonstrate with examples is an instruction.
The banned list is your most valuable page
Working with AI content shows, again and again, that negative instructions take effect faster than positive ones. Models default to the same constructions: openings referring to a shifting landscape, rules of three, rhetorical questions, hollow closing paragraphs that summarise what you just read.
Write down what you don't want to see, literally, in the wording you keep encountering in the output. Keep the list short enough to remember. Twenty banned phrasings applied consistently do more for your brand than a page about brand values.
Test it before you roll it out
Two tests are enough. The first is a blind read: give five texts to colleagues who don't know which ones a model produced. If they can't pick them out, your brandbook works. The second is simpler: measure how many minutes you spend rewriting one piece. That number is your real score. If it doesn't drop after two rounds, the problem sits in your documentation, not in your tool.
What you can't expect: output that ships without editing. Almost nobody gets there. What you can expect is a first version that's usable instead of disposable.
Who owns this when your team is small
One owner, one file, one fixed location. Keep the brandbook in plain text or markdown so it can be loaded into any prompt or agent without formatting noise. Keep a correction log: every time you fix something by hand, note the rule that would have prevented it. Update the document each quarter with those notes. That costs an hour and is the cheapest quality control you have.
Whether you do this yourself or bring in help depends on your starting point. With a sharp positioning and enough of your own writing, the steps above will get you a long way. If you doubt whether your brand is clear enough to scale in the first place, a documentation exercise won't fix it. The work sits a layer deeper, in the positioning itself. That's usually the point where REX.digital gets involved: establish what your brand is, then automate how it sounds. Start this week with those ten texts that feel right. The rest follows from there.