A voice doc a model can actually follow

1Most brand voice documents are useless, and it's the same failure every time

Open almost any brand's voice guidelines and you will find a list of adjectives. Warm. Confident. Approachable. Bold but never arrogant.

Now try to act on it. Write me a sentence that is approachable but not approachable in a way that reads as arrogant. You can't, because those words describe how the copy is supposed to feel after it's written, not how to write it. They're a review rubric wearing a style guide's clothes.

A model has exactly the same problem, only it won't tell you. Give it warm and confident and it will produce something fluent, generic and technically compliant, because every word in the brief was satisfiable by anything.

A usable voice doc is made of things you can check. Not sound approachable — use contractions, never open with a question, one adjective per sentence. If two people could disagree about whether a line follows the rule, it isn't a rule yet.

2Derive it from what shipped, not from what they say

Here is the method, and it comes from doing this on a real brand.

The reference version of this — the skill that Form BV-01 is generalised from — was not written from that company's brand guidelines. It was written from an audit of what had actually shipped: the live site, the campaign captions, the outreach documents, the in-app strings. Every rule in it is stated as verified in shipped copy, with the file it appears in.

That was not thoroughness for its own sake. It was because the documented rules and the shipped copy disagreed, and the disagreement was the most useful thing found. The written guidelines said "never use em-dashes." The live website used them constantly — not as an occasional slip, but as its house style, on six pages.

Two things follow from that, and they are the core of this section.

1. Ask for the last ten posts, not the brand deck

The deck tells you who they want to be. The posts tell you who they are. When a client sends you a PDF of brand values, thank them, read it once, and then go and read what they actually published last month. Write your voice doc from the second one.

Then go one step further: read the comments. How the audience talks back is a better guide to register than anything the brand says about its audience.

2. When the doc and the copy disagree, that is a question, not a decision

This is the part people get wrong, and it is where a junior creator quietly burns trust.

You find the rule says one thing and half their output does another. Do not silently pick one. Do not "fix" their site in your head and write to the rule. Do not assume the shipped copy is correct because it's newer.

Put it in front of them: "Your guidelines say no em-dashes; your last six posts use them throughout. Which am I writing to?" That question takes ten seconds to ask and it does three things — it gets you an answer, it demonstrates you read their material properly, and it means the inconsistency is theirs to resolve rather than a thing you guessed at and got blamed for.

Record the answer in BV-01. That's what the form is for.

3Why BV-01 has four sections and not one

Each section catches a different failure. Here is what each one is for and what goes wrong when it's blank.

BV-1 — Brand identity: how they sound

Described by comparison, because comparison is checkable and adjectives aren't. Three brands you'd want to be mistaken for; three you'd never want to sound like; and how they'd talk at a party.

Without it: output drifts to the most statistically ordinary register available, which is the thing Module 2 §1 calls the tell. Nothing is wrong, and nothing sounds like anyone.

The comparison trick is doing real work. "Muji, Serious Eats, Patagonia" transmits more usable information in three words than a paragraph of adjectives, because each name carries sentence length, formality, and attitude to hype all at once — and because you and the client can both check whether a draft got there.

BV-2 — Audience & niche: who it's for

One real person in a situation. Not a demographic.

Without it: you get copy for everyone, which persuades no one. This is the single most common blank field and the most expensive.

The second and third questions in this section — what the audience already believes, and what you're trying to change — are the ones that decide the hook. Module 2 §1's audience requirement comes straight from here.

BV-3 — Guardrails: what they cannot say

The section that stops a video getting pulled. Barred claims, off-limits words, disclosure requirements.

Without it: everything else still works and then one line ends the relationship. This is the only section with legal exposure attached, and the form's own note is right that it's worth more care than the rest of the form combined.

BV-4 — Deliverable specifics: what you actually owe

Format, length, platform, revision rounds, who signs off, turnaround.

Without it: you will do four rounds of revisions for the price of two and discover on the Friday that the person you've been sending drafts to can't approve them. The form calls this the boring half, and the half that arguments are about, which is exactly right.

4Two rules the reference skill had that most people miss

Both are worth stealing directly.

Some text is quoted, not authored — and voice rules don't apply to it. The reference skill has a whole section on this: where a piece of copy reproduces text generated elsewhere, verbatim, the house voice rules must not be applied to it. If you quote a customer, a client's own product copy, or a real generated output you're showing as proof, you leave it alone, including its punctuation and its exclamation marks. Editing quoted material to fit a voice guide turns a quote into a fabrication.

A brand can have more than one register, and the split must be explicit. The reference brand had two — one for consumers, one for business buyers — with different vocabulary, different punctuation rules and different calls to action. They were consistently enforced and easy to get wrong. If your client sells to two audiences, you need two BV-1 answers, not an averaged one.

5What "at scale" actually means

Not more output. More clients, without the rules blurring between them.

One brand's off-limits list applied to another brand's script is the specific failure this whole system exists to prevent — and it does not happen because you're careless. It happens because you wrote for Client A on Tuesday and Client B on Wednesday and both briefs lived in the same head.

The next section fills one in. §5 puts it somewhere it gets loaded automatically, which is the only version of this that survives contact with a busy month.