Most brand voice documents are unusable. They describe a personality — "confident but approachable, expert but never condescending" — which tells a writer approximately nothing and tells a machine literally nothing.

A voice guide that works reads more like a specification than an essay.

What a usable voice guide contains

Vocabulary rules, with actual words. Not "we use plain language." A list: we say customer, never client. We say problem, never pain point. We never use leverage as a verb. Words with a defined meaning here — what "platform" means for us and what it does not.

Sentence-level patterns. Typical length and whether we vary it. Whether we use fragments. Whether we address the reader as "you." Whether we use contractions — a small decision that changes register more than most.

Structural conventions. Do we open with a claim or with context? Do we use headings as questions or as statements? Do we number things? Do we end with a summary — and if not, say so, because that is the default a model will produce.

Claim discipline. This is the most important section and the one that is always missing. What may we assert without qualification? What requires a source? What may we never claim? Where do we use numbers, and which numbers are approved?

Paired examples. Ten pairs of before and after — a sentence in the wrong voice, the same sentence in ours. This teaches more than every abstract description combined, to humans and machines equally.

Why "confident but approachable" fails

Because it cannot be checked. Two people reading it will produce different text and both will defend theirs as compliant. There is no test.

"We do not use exclamation marks" can be checked. "We never claim to be the best at anything" can be checked. "We use contractions" can be checked. A specification is a set of statements someone can verify against the output.

The test for your own guide

Hand it to a competent writer unfamiliar with your company along with a topic. If what comes back sounds like you, the guide works. If it sounds like generic professional writing, the guide described a feeling.

Then do the same with a model. The result is usually informative — most voice guides produce identical output regardless of whether they were supplied.

Why this matters more now

When all writing was human, an impressionistic guide was survivable because writers absorbed the voice from reading existing material and from being edited. That transmission mechanism is slow and it works.

A model has no such apprenticeship. It gets the guide and the examples, and nothing else. Everything left implicit is filled in with the median of the internet.

Which means the discipline of writing a real specification is no longer optional — and the exercise is worth doing regardless, because most companies discover during it that they do not actually agree on their own voice. The document was vague because the agreement was.

Maintaining it

A voice guide is not written once. Every time an editor makes the same correction twice, that correction belongs in the guide as a rule. The document should grow from the edits, which is the same feedback loop that improves a model — and it works on the humans too.