Brand voice
Three adjectives are not a voice
A brand voice is a documented set of concrete writing rules, examples, and forbidden words that let different people produce text that reads as one consistent author.
Nearly every brand voice document says confident, human, and approachable, which describes almost all writing and constrains none of it. A voice document earns its existence only when a competent writer who has never met you can use it to produce something you would have signed, and that requires specifics rather than adjectives.
The forbidden list does more work than the aspiration list
Telling a writer to be clear does nothing. Telling them you never say leverage, unlock, delve, seamless, or game changing changes the next paragraph immediately. A list of words and constructions the brand does not use is the most operationally useful part of any voice document, and it is the part almost nobody writes.
Voice lives in rhythm, not vocabulary
What makes writing recognisable is sentence length variation, how often it starts with a conjunction, whether it uses the second person, and how it opens. Two writers with the same vocabulary and different rhythms sound like different people. Record the rhythm: typical sentence length, paragraph length, and whether the writing hedges or asserts.
Examples beat rules
Pairs of before and after are worth ten rules, because they carry the judgement a rule cannot express. Keep a small file of sentences you rewrote and the reason, and hand a new writer that instead of a style guide.
One voice per person, not per company, when the byline is a person
A company account can have a company voice. A founder account cannot, because readers can tell and they judge inconsistency harshly on a personal profile. If several people write under one person name, the document has to describe that person and not the brand.
How to do it
Collect thirty pieces you would sign
Real published writing, not aspirational examples from other brands. The voice already exists in what you have written; the document is an act of description, not invention.
Extract the patterns
Average sentence length, how posts open, how often the first person appears, whether questions are used. Count rather than guess, because writers are unreliable narrators of their own habits.
Write the antivocabulary
Twenty words and phrases that would never appear. This is the fastest quality gate you will ever build and it can be checked mechanically.
Add annotated examples
Five pairs of a weak sentence and its corrected version, each with one line about what changed. This is what teaches judgement.
Test it on somebody else
Give the document to a writer who does not know you and ask for three paragraphs. Whatever they get wrong is what the document failed to say, and that is the only reliable review it will ever get.
Common mistakes
- Adjectives instead of rules, so the document describes every brand and constrains none.
- No forbidden words list, which leaves the single most checkable rule unwritten.
- Aspirational examples copied from brands you admire rather than from writing you actually published.
- A document nobody tests on an outside writer, so its gaps are only discovered in published work.
How Quillen handles brand voice
Quillen builds a persona from your own writing rather than asking you to choose a tone from a dropdown. It is stored as a SOUL.md covering your phrasing, sentence rhythm, and an antivocabulary of words you would never use. A workspace can hold several personas, each connected account has a default persona it composes with, and a campaign is drafted with the persona assigned to it, so one team running several accounts does not have to converge on a single voice. Every generated draft is linted against that persona before it reaches your queue, and a draft that fails the lint is flagged rather than sent. Your edits to drafts feed back into the persona, so a voice that starts approximate gets closer with use.
Common questions
- How long should a brand voice document be?
- Two pages of specifics beats twenty pages of principle. The binding constraint is whether a writer reads all of it before writing, and almost nobody reads twenty pages twice.
- Can an AI tool learn a voice from examples?
- It can learn the observable parts well: rhythm, vocabulary, openings, and the words you avoid. What it cannot infer is judgement about what you would refuse to say, which is why a human approving before anything publishes is still the safeguard that matters.
- Should each person on a team have their own voice?
- If they post under their own name, yes. A shared voice across personal accounts is the single clearest tell that posts are being produced centrally, and readers spot it quickly.