How to capture your brand voice for AI

Most marketing teams know their brand voice when they hear it. They can review a piece of content and immediately identify that something is off — wrong register, wrong energy, wrong word choices. That recognition is real and valuable. It is also tacit knowledge: expertise that lives in professional judgment rather than in explicit rules anyone could follow.

Descriptive brand voice documentation — adjectives like "warm," "confident," "conversational" — is an attempt to make that tacit knowledge explicit through description. It works, partially, for humans who bring domain judgment to the interpretation. For AI, description is insufficient. AI needs directive rules: specific, named constraints that apply the same way regardless of who is prompting.

The capture process is about converting tacit to explicit. Here is how to do it.

Start with content that passed, not content that failed

The most common mistake in voice capture is starting from scratch — trying to describe brand voice by thinking about it abstractly. Abstract descriptions produce abstract rules, which produce abstract AI output.

Start instead with content your team has approved. Collect five to ten pieces that represent your brand voice well: pieces you are proud of, pieces that required minimal editing, pieces that feel right. Then study them.

What you are looking for: patterns you did not consciously put there. The rhythm of sentences. The ratio of short to long. The words that appear consistently. The words that never appear. The way arguments are sequenced. Whether assertions or questions dominate. Whether numbers appear and how specific they are.

Write down what you observe as rules, not descriptions.

Wrong: "We use plain language." Right: "Voice rule — Plain language: Average sentence under 18 words. No noun strings longer than 3 words. No passive construction when active is possible."

The rule version is derivable from the actual content. The description version is what you would tell a human copywriter in a briefing meeting. Both are useful in their domains; only the rule version constraints AI output.

The editing method

If you do not have a strong corpus to analyze, use the editing method. Take a piece of AI-generated content — any content, from any tool — and edit it to match your brand voice. Edit it thoroughly, the way you would before publishing.

Every edit you make is a rule the brief does not yet contain. Categorize each one:

Substituted a word → The original word is on your ban list. The replacement is a required term or a model for voice. Add both.

Rewrote a sentence structure → Your brand has a structural preference the AI violated. Name the preference and specify it.

Deleted an explanation → You assumed audience knowledge the AI did not. Add an audience rule: what this audience already knows; what never needs to be explained to them.

Moved information earlier or later → Your brand has a sequencing preference. Specify it as a structure rule.

Added specificity → The AI generalized where your brand always uses concrete numbers or examples. Specify the concreteness requirement.

Changed the CTA → Your brand has specific call-to-action language. Document the exact phrasing.

Do this exercise with three pieces. Each round surfaces failure modes the previous round missed. After three rounds, you have a rule set grounded in real editing decisions, not in abstract preferences.

The five sections every voice rule set needs

Once you have a set of rules from the process above, organize them into five named categories. Each category corresponds to a different type of AI failure.

1. Vocabulary bans — Words that never appear. Be specific: if "leverage" is off-brand, add "leverage." If "innovative" is also off-brand, add it separately. Do not add generic categories ("jargon") that require interpretation.

2. Required terms — Words that always appear. Your product name, your signature phrase, your specific category term. Include context: not just the word, but where and how it is used.

3. Sentence structure rules — Specifications for how sentences are built. Lead placement, hedge prohibition, sentence length ceiling, passive voice rules. One rule per structural preference.

4. Argument structure rules — How content is organized at the paragraph and section level. Does your brand open with the conclusion or build to it? Does it use numbered frameworks or flowing prose? Does it end with a question or an assertion?

5. Audience rules — What your audience knows and does not need explained. What they assume and do not need convinced of. How they want to be addressed. These rules prevent the AI from misjudging the audience's expertise level, which is the most common register failure in AI content.

What not to include

Voice rules that require interpretation defeat their own purpose. "Write with empathy" requires knowing what empathy looks like in your brand's specific context. "Use empathy markers — acknowledge difficulty before offering a solution" is a directive that constrains AI output without requiring interpretation.

Similarly, negative examples are only useful if the positive version is also specified. "Don't sound like a press release" tells the AI what to avoid but does not tell it what to do instead. "Lead with the specific finding, not the announcement of the finding" tells it what to do.

The test for any rule: could an AI system follow this rule without knowing what your brand is? If the rule requires context to interpret, it needs to be made more specific before it is useful as a constraint.

How the rules connect to AI generation

Copper Sun's brand context system stores voice rules as structured data and injects them into every content generation. The rules apply across modules — blog writing, email sequences, social content — without re-pasting into each prompt. The structured brand brief template covers the complete brief format; voice rules are one section within it.

The capture process above produces the raw material. The brief template organizes it into the format that constrains AI generation effectively.

Frequently Asked Questions

How many voice rules are enough?

Enough to cover the failure modes that appear in your AI content. Most teams find a working set is between 8 and 15 rules across all five categories. Fewer than 8 typically leaves significant gaps; more than 15 often includes rules that are aspirational rather than grounded in actual editing decisions. Start with what the capture process surfaces and trim later.

Should voice rules apply to all content types?

One set of rules works when your brand produces content in one register. If you produce content in significantly different registers — a technical integration guide and a social campaign — you may need separate rule sets with shared vocabulary bans but different sentence structure and argument structure rules. Start with one set and split it only when you have documented evidence that a single set produces consistent failures in one content type.

What is the difference between a voice rule and a style guide?

A style guide is written for human writers who will exercise judgment about application. A voice rule is written for AI that will follow it literally and consistently. A style guide can say "avoid passive voice when possible" and rely on writer judgment about when passive voice is appropriate. A voice rule needs to specify: "Replace passive construction whenever the agent is known and the sentence can be rewritten without passive." The specificity requirement is the core difference.

How often should voice rules be updated?

Add a rule whenever you catch a consistent failure mode in AI content — a word that keeps appearing, a structure that keeps going wrong. Do not delete rules unless you have evidence they are causing problems. A stable rule set is more valuable than a frequently updated one, because stability allows you to distinguish rule failures from prompt failures. Quarterly review is appropriate once the initial set is stable.