How to Build a Brand Voice Claude Actually Follows
Most teams hand Claude their brand voice as three adjectives: "friendly but professional, confident but not salesy." Claude does something reasonable with that. The problem is that the same three adjectives describe a hundred other brands, so the output drifts a little further off every time a different person types their own version of them.
A tone description works once. A voice profile works every time.
A voice profile is not a longer description. It is a reference concrete enough that Claude, or a new hire, can check a sentence against it and get the same answer you would.
Why "friendly but professional" never sticks
Nobody sets out to write off-brand copy. It slips one decision at a time: one writer reads "confident" as punchy one-liners, another reads it as long declarative sentences, and both are following the brief honestly. An adjective is a category. Without something more specific to point at, every writer, human or AI, fills the gap with their own reasonable guess, and the guesses do not match.
The three parts of a voice profile
A description gives Claude adjectives to interpret. A profile gives it evidence to match. Three pieces, and you need all three:
| Part | What it is | What it lets Claude do |
|---|---|---|
| Samples | 5 to 8 real pieces you have shipped and still stand behind | Match a pattern instead of guessing at a word |
| Rules | Explicit do and don't lines, including specific words you avoid | Check a sentence and catch a violation |
| Anti-example | One piece that is close but wrong, with a sentence on why | See the boundary, not just the target |
Samples alone leave Claude to reverse-engineer your rules, and it will get some wrong. Rules alone are still abstract. The anti-example is the part almost everyone skips, and it does the most work: "what almost fits" defines a voice more sharply than another example of what does.
Tip
If you can only build one part today, write the rules, and make each one a sentence Claude can test a draft against. "We write 'you'll save time,' never 'streamline your workflow'" beats "we sound efficient."
Building the profile
- 1
Pull 5 to 8 genuinely representative samples
Real headlines, emails, or product copy your team has shipped. Not your favorite pieces, the ones that show how you actually sound week to week.
- 2
Turn the patterns into explicit rules
Ask Claude to read the samples and propose the do and don't rules it sees, then edit them. It is faster to correct a draft set of rules than to write them from a blank page.
- 3
Add one off-brand example and say why
A piece that is close but wrong, plus one sentence on what tips it over. Rejected freelancer copy is perfect for this.
- 4
Store it where every conversation can reach it
A Claude Project's knowledge files, so everyone drafting copy starts from the same reference instead of their own memory of the brand guide.
- 5
Test it on something low-stakes first
An internal email or a social caption, not the homepage headline. This is where you find out whether the profile holds before it matters.
What a real voice profile looks like
Here is a complete one for a fictional project-management tool whose whole positioning is "for people who are sick of project-management software."
Samples we stand behind
- "Your projects, minus the ceremony."
- "We didn't add a feature this month. We removed three."
- "Set it up in the time it takes to make coffee. We timed it."
- "You'll get a weekly digest. That's the only notification we'll ever send by default."
Do and don't
- Say "set up," not "onboard." Say "plan," not "workflow."
- Lead with what we removed or simplified, not what we added.
- Short declarative sentences. One idea each. No semicolons.
- Never use "all-in-one," "powerful," "robust," or "next-generation."
- No exclamation points in headlines. One is allowed in a sentence, rarely.
Anti-example
"Trellwise is the all-in-one platform that helps modern teams optimize their workflows and get more done." Wrong on three counts: it is a category sentence (swap the name and it sells any SaaS), it leads with what we do instead of what we removed, and "all-in-one" is on our never list.
Notice that the rules are testable and the anti-example names the exact failure. A writer, or Claude, could take a draft headline and tell you in seconds whether it fits.
Prompt templates
Here are [6-8] pieces of copy we've shipped and stand behind: [paste]. Here's one we rejected, and why: [paste the piece] because [reason]. From this, write: 1. A "do and don't" list, each line a sentence I could test a draft against. 2. The specific words or phrases we clearly avoid. 3. One sentence describing the voice for someone who hasn't seen the samples. Keep it under one page.
Using the brand voice profile above, write [what you need].
Before you show me the draft, check it line by line against the do and don't
rules. Flag any line that breaks a rule or lands close to the anti-example,
and tell me which rule and why.What a complete profile covers
5 to 8 real, on-brand samples, not a written description
Explicit do and don't rules, including the specific words you avoid
At least one off-brand example with a one-sentence reason
Stored where every conversation can reach it, not in one person's notes
Common mistakes
- Adjectives instead of rules. "Confident, warm, human" is a vibe, not something Claude can check a sentence against. Every line should be testable.
- Samples that are your favorites, not your average. The award-winning campaign is not how you sound in a Tuesday product email. Pick representative pieces.
- Skipping the anti-example. It is the single most useful part and the one people leave out. Without it, Claude only knows the target, not the edge.
- Writing it once and never touching it. A voice that was accurate a year ago describes a brand you used to be. Swap in a fresher sample when an old one starts to feel off.
- Keeping it in a doc nobody opens. If it does not live in the Project or the prompt, people fall back to the three adjectives.
Inside Claude Tutorial
Build it once, stop re-explaining it.
Setting something up once so you never restate it in a conversation is a pattern that goes well beyond brand voice. The app teaches dozens of variations on it through short, practical lessons.
Frequently asked questions
How do I give Claude a brand voice?
Paste 5 to 8 real samples, a set of do and don't rules, and one off-brand example into the prompt, or store all three in a Claude Project so every chat inherits them. Then ask Claude to check each draft against the rules before showing it to you.
Can Claude learn my brand voice permanently?
Not on its own between separate chats. It applies whatever voice reference is in the current conversation or the Project. Store the profile in a Project's knowledge and every new chat there starts with it.
How many writing samples does Claude need?
Five to eight is enough for a consistent voice. Fewer and it over-fits to one piece; many more and you are mostly adding noise. Quality and representativeness matter more than count.
Where should I keep a brand voice guide for Claude?
In a Claude Project's knowledge files, so it is available to every conversation any team member starts, not pasted in by hand each time.
What is the difference between brand voice and tone of voice?
Voice is constant: how the brand always sounds. Tone shifts with context: warmer in a welcome email, plainer in an outage notice. A good profile locks the voice and gives Claude room to adjust tone.
Is a brand voice profile different from matching my own writing?
Yes. Matching your personal voice, for a reply or an update with your name on it, means giving Claude a sample of your writing. A brand voice profile represents the organization and has to survive any one person leaving. Getting Claude to sound like you covers the personal version.
Continue reading
- Claude for Marketing: where this fits in the full workflow, from research to distribution.
- Writing Better Marketing Copy With Claude: the drafting and feedback loop that runs on top of a locked voice.
- Turn Competitor Pages Into a Positioning Map: voice is how you sound, positioning is what you say. This is the other half.
- Building a Reusable Context Brief You Stop Retyping: the general pattern behind saving any reference once instead of restating it.