What a brand voice document actually needs to say

You send a first draft to a contractor. It comes back technically correct and somehow completely wrong. The brief said your brand voice is “conversational, confident, and direct.” They delivered conversational, confident, and direct. It just does not sound like you. You spend an hour rewriting it and cannot explain, even to yourself, what you changed and why.

This problem does not go away when AI replaces the contractor. It gets more frequent and more expensive.

The adjective problem

Most brand voice guides are built from adjectives. “Conversational but authoritative. Warm but professional. Bold but never arrogant.” These pairs feel meaningful when you write them and are nearly useless when someone else tries to act on them.

The reason is simple: adjectives describe a destination without providing any route. “Confident but not arrogant” tells a writer nothing about what to do in the moment when a sentence could tip either way. It is a standard to aspire to, not a guide to produce against. The writer makes their own call, and their call is not yours.

Feeding these to AI amplifies the problem. “Write in our brand voice: conversational, confident, and direct” produces something different every time, almost none of which will match your specific register. The model interpolates from its training set of content that used those same adjectives, not from your writing. You get a statistically average “confident and conversational” voice, which is to say no voice in particular.

What works instead

The thing that works, for human writers and AI tools both, is examples with annotation. Not “we sound like this,” but “here is a sentence we would write, and here is a sentence we would not write, and here is what is specifically different between them.”

Real examples from your own published work, pulled deliberately rather than collected at random, are the most useful inputs you can give anyone working in your voice. The annotation matters as much as the example. “We do not use passive voice in subject lines” is a rule. “This subject line reads like we are apologizing for existing” is a lesson. The lesson transfers more reliably, because the next writer understands not just the what but the why.

Anti-examples are underrated. A short list of phrases you would never write, with a note on why each one is wrong, teaches faster than a longer list of phrases you love. It is easier to recognize a mistake in context than to reproduce a success.

This is also the format that AI tools can actually use. When you feed a model several examples of copy you are proud of alongside several examples of the kind of copy you are trying to avoid, you get substantially more useful output than you get from a list of adjectives. The model can match patterns. It cannot match aspirations. Specificity in the brief is what unlocks AI copy generally, and the voice document is part of that brief.

Two documents, kept short

A voice guide that actually functions has two parts.

The first is a sample library: ten to twenty published pieces or passages that represent the voice at its best, each with a short note on what makes it work. This is not for inspiration. It is for calibration. When a draft comes back wrong, you compare it against the library and can usually locate the gap in a single read.

The second is a failure shortlist: common drift patterns that appear when the voice is not tended. Not abstract mistakes, but specific recurring language. Category words like “streamline” and “empower.” Hedges tacked onto the end of confident claims. The habit of explaining the thing you just said in slightly different words. AI-generated drafts produce these patterns reliably, so naming them makes the review faster. The editorial pass post goes through the most common ones in detail.

Both documents should be short enough to actually use. A voice guide no one opens is worse than nothing, because it creates the impression the problem has been solved when it has not.

Voice as a living document

Brand voice changes. Not dramatically, but steadily. The way you talk about what you make shifts as the product changes and as the market around it changes. A voice doc written at launch is not the voice a few years later, and treating it as finished makes the drift harder to catch.

The maintenance habit that keeps it current: add a new example to the sample library when you write something you are proud of, and add to the failure shortlist when you catch a pattern recurring. The document stays useful because you update it when the information is fresh, not when you have scheduled time to rewrite it from scratch.

What AI does to this problem is raise the stakes, not eliminate them. When voice was mostly one person, drift was bounded by how much one person could produce. When AI helps a small team publish at much higher volume, a weak voice document shows up fast. The answer is the same one that was always right: make the implicit explicit, give it a format that travels, and keep it current enough to do real work.