Ask any chatbot for an Instagram caption and you’ll get the same brand every time. Upbeat, a little breathless, two exclamation marks and a rocket. It isn’t your voice. It’s the average of the internet’s voice, and it’s what you get when your tone of voice guide says “friendly, approachable and bold”.
When I ran a brand across a hundred offices, voice was the thing that drifted first. Colours you can see from across a room. Words slip quietly. Now that first drafts come from models, they slip faster, because a model fills every gap in your guidelines with its own defaults.
Adjectives don’t survive
Every brand is warm, human and confident. Those words don’t narrow anything down, for a person or a model. The test I use: could a competitor paste your voice section into their guidelines and change nothing? If yes, it isn’t doing any work.
Keep your adjectives if you like them, but put a fence around each one. “We are warm, not gushing. We are direct, not blunt.” The second half is the useful half. It tells the model where to stop.
Pairs survive
The most reliable voice rule is a word to use next to the word to avoid. It’s specific, it’s easy to remember, and a checker can find it in a draft without any judgment at all.
In Brandbook these live on the Voice page as say/avoid pairs. Every check looks for the avoid side. When it finds one, the finding names the word and offers the replacement as a one-click fix, so the edit takes a second instead of a conversation.
Punctuation survives
Models overuse three things: exclamation marks, emoji and capital letters for emphasis. If your brand doesn’t, say so in exactly those words. “No exclamation marks in headlines.” “No emoji.” “Never write in all caps.” These are the easiest rules in the world to check, and the ones a model breaks most often.
Limits survive
Length is a voice decision. A brand that speaks in six-word headlines feels different from one that speaks in twenty. Give a number. “Headlines under eight words” is a rule a model will follow on the first try and a checker can count.
Examples survive best
Nothing teaches voice like a handful of lines that are exactly right. Pick four headlines you’d be proud to see on a bus shelter and keep them in the guidelines. When an agent asks Brandbook for your brand context, those headlines come with it, so the model has something real to imitate instead of something generic to avoid.
Make every sentence checkable
Here’s the habit worth building. For every sentence in your voice section, ask whether a tool could tell you that you broke it. When you write a rule in Brandbook, it tries to turn the sentence into a check: a banned word, a required phrase, a word limit, no exclamation marks, no emoji, no shouting. If it can’t, that’s a hint the sentence is a mood, not a rule.
- “Be conversational” → “Use contractions: we’re, you’ll, it’s.”
- “Avoid jargon” → “Never say ‘leverage’, ‘synergy’ or ‘solutions’.”
- “Keep it punchy” → “Headlines under eight words.”
- “Stay calm” → “No exclamation marks in headlines.”
Put the rules where the writing happens
A voice guide only works if it’s in the room when the draft gets written. That used to mean training sessions. Now it means connecting your tools. With Brandbook’s MCP server, Claude or Cursor fetch your voice rules before they write and run a check before they show you. In ChatGPT or a Claude project, add your brand.md as a file. Either way, the model starts from your voice instead of its own.
The model will always fill the gaps. Leave fewer gaps.
You won’t get it perfect on the first pass. Watch what your checks keep flagging and what people keep rewriting, and turn the pattern into a rule. Brandbook suggests some of these for you when the same fix keeps coming up, but you’ll spot plenty yourself. Six months in, your voice guide will be shorter, sharper and actually used.
