The problem with most brand voice guides is that they describe a feeling instead of a behavior.
“Confident but approachable.” “Witty but not try-hard.” “Authoritative without being stuffy.” These phrases tell you nothing about what words to choose when you’re staring at a blank draft. And when you paste them into a system prompt, you get back the same average-of-the-internet voice that every other brand gets.
Useful brand voice prompting is about encoding behavior, not adjectives.
Start With Anti-Examples
The fastest way to teach a model your voice is to show it what you’re not.
Bad: “Write in a professional yet friendly tone.” Better: “Avoid corporate hedging language (‘synergies,’ ‘leverage,’ ‘best-in-class’). Never use passive voice when active is possible. Don’t end sentences with questions designed to seem engaging — it reads as desperate.”
Anti-examples work because they’re specific and actionable. “Witty” means nothing to a language model. “Avoid puns, but dry understatement is fine” actually narrows the possibility space.
Build a list of:
- Phrases your brand would never use
- Sentence structures to avoid
- Topics you handle differently than your competitors
This negative space does more work than most positive descriptions.
Encode the Reading Level and Register
Your voice isn’t just word choice — it’s the implied relationship between writer and reader.
Register questions to answer in your prompt:
- Are we writing to someone as a peer, a student, or a client?
- How much does the reader already know? What can we skip?
- What are we allowed to assume they care about?
- Are we trying to teach, persuade, entertain, or validate?
A brand that writes to its audience as peers sounds different from one that writes as teacher-to-student — even if both use “simple, clear language.”
Example register instruction:
“Write as if you’re explaining this to a smart colleague who has domain knowledge but hasn’t been following this particular trend closely. Don’t over-explain the basics. Get to the point.”
The Reference Text Method
If you have a body of existing content that sounds right, this is your most powerful tool.
- Pull three to five pieces that represent your voice at its best
- Ask the model: “Analyze the voice and style of these examples. Describe it in terms of: sentence length patterns, tone, typical openings, how claims are supported, any recurring structural choices.”
- Take that analysis, edit it for accuracy, and use it as your style guide
- Include excerpts as few-shot examples in your prompts
The model-derived analysis often catches things you didn’t consciously notice about your own voice — characteristic sentence lengths, a tendency to use em-dashes, how you handle the transition from observation to implication.
Handling Edge Cases
Brand voice guides break down at the edges. What do you sound like when:
- Covering a difficult or controversial topic
- Addressing a mistake you made
- Writing about competitors (if you do)
- Responding to criticism
These scenarios need their own explicit handling. Most brands avoid documenting them and then wonder why their AI-assisted content sounds tone-deaf in exactly these moments.
For each edge case, write a short instruction: “When covering topics where our audience has strong opposing views, we acknowledge the tension without trying to resolve it. We don’t performatively validate every perspective.”
Structure as Voice
Voice isn’t just word-level — it’s structural.
Some brands open with a story. Some open with the claim. Some open with a counter-intuitive statement and spend the rest of the piece defending it.
Document your structural preferences explicitly:
- What does a typical opening look like?
- How long is a typical intro before you get to the meat?
- Do you use headers? How many, and what style?
- How do you close — summary, call to action, open question, or a kicker line?
When you include this, you stop getting essays that drift into the AI’s default structure and start getting pieces that fit your editorial pattern.
Calibration Loop
None of this is set and forget. Treat your system prompt as a living document.
After each piece, note: what needed to be edited and why? Was it a word choice problem, a structural problem, or a register problem? Track the patterns. Update the prompt.
After three to four cycles, you’ll have a prompt that reliably gets you to 70-80% of the way there. The remaining editing becomes fast — not because the AI has learned your voice, but because you’ve gotten better at specifying what you actually want.
The real unlock is understanding that “brand voice” is not a vibe — it’s a set of constraints. And constraints are exactly what language models work well with.
The more precise you are about what you don’t want, the more room there is for something worth publishing to emerge.