Brand voice as a specification: turning style guides into agent instructions
AI-assisted research. This note was prepared with the assistance of AI tools. It may contain errors and is not professional advice. See our Legal Disclosures.
As agents take on more of the drafting in marketing teams, the brand voice guide has quietly become an operating document. Ask for one and you will usually get a page of adjectives: confident, warm, clear, a little playful.
Those words help people who already know the brand. They do little for an agent, because every reader interprets an adjective differently, human or machine. "Confident" to one writer means short declarative sentences, and to another it means superlatives.
To make voice work for agents, we rewrite voice guides as specifications: documents that can be followed, checked, and tested. This post covers how to do that translation, how to test the result, and where human judgment still belongs.
From adjectives to behaviors
Each adjective in a voice guide should become a set of observable behaviors. A behavior is something a reviewer can point to in a draft and say "this follows the rule" or "this breaks it." Here is how three common adjectives translate:
- Confident: state approved claims directly, without hedges like "might" or "perhaps," and avoid superlatives that cannot be supported.
- Clear: keep one idea per sentence where possible, hold average sentence length under 20 words, and define any jargon on first use.
- Warm: address the reader as "you" and acknowledge their situation before making a pitch.
Behaviors can be checked. Adjectives cannot. The translation also forces the team to agree on what the adjective meant all along, which often surfaces disagreements that were hiding inside a shared word.
Keep the list short. Three to five adjectives, each with two or three behaviors, is enough for most brands. A specification with 40 rules is hard for an agent to apply consistently and harder for a reviewer to remember.
Show pairs, not just examples
Examples help, but contrasting pairs help more. For each behavior, include a short "instead of this, write this" pair. The pair below illustrates "confident" and "clear" together:
Instead of: "Our revolutionary platform transforms your entire marketing operation." Write: "The platform handles weekly reporting, so your team spends Monday planning instead of copying numbers."
Pairs teach the boundary, which is what agents most often get wrong. A single good example shows one point inside the acceptable zone. A pair shows where the edge is and which side of it the brand sits on.
Choose pairs from real drafts your team has corrected. The most useful pairs come from mistakes that recur, because those are the boundaries the agent has already shown it cannot find on its own.
Define the edges
A useful voice specification also says what the brand never does. Edges are easier to write than positive rules, and they are often what reviewers spend the most time fixing. Common entries include these:
- Banned words and phrases: terms that are overused, vague, or misleading for your category.
- Off-limits topics: subjects the brand does not joke about or comment on.
- Claims needing review: statements that must go to legal before they ship.
- Formatting rules: constraints such as no exclamation points in subject lines.
These edges are the easiest rules to check automatically after generation. A banned-word list or a formatting rule can run as code on every draft, without relying on the agent to remember it. Our post on guardrails as code covers how to move these checks out of the prompt and into the pipeline.
Structure the document for an agent
A voice specification is read by agents and people, and the layout should serve both. We recommend a fixed order: a two-sentence summary of the voice, then each adjective with its behaviors and pairs, then the edges, then channel adjustments.
Channel adjustments matter more than teams expect. The same brand sounds different in a subject line, a product page, and a support reply. Rather than writing separate guides, state the base voice once and list the specific shifts per channel, such as shorter sentences in push notifications or no humor in billing messages.
Mark clearly which text is a rule and which is an example. Agents tend to imitate examples closely, including details you did not mean to teach, such as a product name or a sentence length. Label examples as illustrations of a stated rule, and vary them so no single surface feature looks like the pattern.
Make it testable
Once voice is a specification, you can build a small test set. A dozen inputs with known good outputs is a reasonable start, plus a checklist reviewers use when grading new drafts. When you change the voice guide, update a model, or adjust a prompt, run the test set and compare.
This is regression testing applied to tone. We cover building these sets in more depth in golden datasets.
Over time, three measures tell you whether the specification is working:
- Rule pass rate: the share of drafts that clear every automated edge check.
- Edit rate: how much reviewers change before approving, tracked by behavior where possible.
- Recurring corrections: fixes that reviewers make again and again, which point to a missing behavior or pair.
The edit rate is often the most informative. When reviewers keep softening the same kind of sentence, the specification has a gap. See measuring human edit rate for how to collect it without slowing reviews.
Treat the specification as a versioned asset. Every change should come with a note and a test run, for the reasons in versioning agent instructions. A voice guide edited informally will drift, and the drift will show up in the content before anyone notices it in the document.
Keep humans as the final judge
Voice is still a matter of judgment, and a specification will never capture everything. The aim is to protect human taste, by making sure people spend review time on the subtle calls instead of the same obvious fixes every day.
A specification also gives reviewers shared language. A note that a draft "breaks the warm rule about acknowledging the reader" is easier to act on than a note that it feels off. Feedback in that form can feed straight back into the document as a new behavior or pair.
The open question we are studying is how voice specifications hold up under volume. Launches and year-end push teams to produce far more content in less time, and small drifts compound. We expect the answer to involve sampling reviews against the specification instead of reading everything, and we are testing how small a sample can be while still catching drift early.
How to apply this: take your three most important voice adjectives and translate each into two or three observable behaviors, one contrasting pair, and one banned pattern. Give that document to your content agents and your new hires, and test it on a dozen inputs.
Related research
The anatomy of a marketing agent: goal, tools, memory, and guardrails
Every dependable marketing agent we study has the same four parts. Get each one right and most failure modes become easier to prevent and easier to diagnose.
Context windows are a budget: managing what your agents remember
Giving an agent more information does not always make it better. Treating context as a limited budget leads to sharper, cheaper, and more dependable agents.
What we mean by agentic marketing, and what we don't
The word agentic is everywhere. A precise definition helps teams decide what to build, what to buy, what to govern, and what to ignore.
Put this research to work
Marketix turns these patterns into software: agents, automated workflows, and a hub that keeps people in control. Sign up for early access.
