To give AI your company’s tone of voice, do not paste your brand guidelines and hope. Rewrite them as a voice sheet: who you write for, five to eight rules a reader could check, the words you use and avoid, two short samples with the features to copy named, and how the tone shifts by situation. Save it where every chat starts, in ChatGPT, Claude, Gemini or Copilot, and test it on three different jobs before the team relies on it.
Why adjectives fail
An adjective in a voice guide is shorthand for a hundred decisions your writers already share. Friendly at a bank means a short greeting and no jokes. Friendly at a snack brand means puns. A model has no shared history with you, so it resolves the adjective to what friendly usually means across everything it has read, which is the voice of everyone. Our post on why AI writes like a brochure shows the result: style words in a prompt produce style in the answer.
Good human guides know this. Mailchimp’s public voice and tone guide separates a voice that stays the same from a tone that changes with the reader’s situation, and describes the voice as plainspoken and genuine, with dry humor. Those words work for Mailchimp’s writers because the guide goes on to show what they mean. Copy the adjectives into a prompt without the showing and the model gets the label without the meaning.
Turn each adjective into rules
For every adjective, ask what you would see on the page if a writer got it right, and write that down as a rule someone could check by reading. Here is the translation for a made-up company, Fernhill Tea, whose guide says warm, plain and expert:
Example: Fernhill Tea is made up, and so are its rules. Yours come from your own best writing.
The test of a good rule is that two people reading the same draft would agree whether it passed. “Warm” fails that test. “Use their first name once” passes it.
What goes on a voice sheet
The rules are the center of the sheet, but a model also needs to know who is reading, what to do when the rules pull against each other, and how the voice bends when the news is bad:
We write for people who already drink tea and want to know what is in the cup.1 Sentences under 20 words. Use the customer’s first name once. Name the origin and brew time of any tea we mention. No exclamation marks.2 We say steep, cup and leaf. We never say premium, artisanal, indulge or curated.3 Two emails we sent this year are below. Copy their sentence length, their sign-off and their use of digits for amounts. Do not reuse their content.4 Apologies: say what went wrong in the first sentence, and no jokes. Launches: one fact per sentence.5 If a rule and clarity conflict, choose clarity. If you need a fact about a tea that is not given, ask.6
- 1Who it is forOne or two sentences about the reader.
- 2Rules you can checkFive to eight, each visible on the page.
- 3WordsA short list of yours and a short list of banned ones.
- 4SamplesTwo short pieces, with the features to copy named.
- 5Tone by situationHow the voice bends for an apology, a launch or a refusal.
- 6When unsureWhat wins when rules conflict, and permission to ask.
Aim for a sheet someone can read in two minutes, around 200 to 300 words plus the samples. The brand voice guide template drafts one from your best and worst copy, ties every rule to a sample, and ends with a short paragraph to paste into a custom GPT, a Claude project or a Gem.
Here is what the sheet changes in an everyday request. The scores come from the free checks in the Prompt Grader:
Write a product announcement for our new oolong tea. Make it friendly, warm and on-brand.
Write a product announcement for our new oolong tea, for customers on our email list who have bought from us before. Follow the voice sheet below. Facts are listed after it; use only those, and if a sentence would need a fact that is not there, leave it out. Voice sheet: """ [the voice sheet] """ Facts: """ [the facts] """ Under 120 words, with a subject line. Then list any rule from the sheet you could not follow, and why.
The last line is the one people skip. Asking the model to name the rules it broke turns the sheet into something you can audit, and shows you which rules are unclear.
Two samples beat fifty
The tempting shortcut is to upload everything: the last fifty newsletters, the website and the brand book, with “write like this.” It feels thorough, and it has three problems. A model copies what is in examples, including their subjects, their old offers and the mistakes nobody caught. A pile of fifty averages your best writing with your worst. And long inputs make models less reliable. A 2025 Chroma study of 18 models found performance fell as input grew, and that text on a related topic that does not answer the question did the most harm. Fifty newsletters are fifty pieces of related text.
Pick two short pieces you would be happy to see copied, each under 200 words, and name what to copy from them: sentence length, how you open, how you sign off, digits for numbers. Then add one line telling the model not to reuse their content. The case for a big library is real for people: a content team benefits from a curated set of approved examples to choose from. Keep that library for your writers, and give the model the two that fit the job in front of it.
Where to save it
A voice sheet that lives in a shared drive gets pasted half the time. Put it where every chat starts:
- In ChatGPT, in a project’s instructions or a custom GPT the team shares.
- In Claude, in a project’s instructions, or as a custom style built from your samples.
- In Gemini, in a Gem’s instructions, with the samples added as files under Knowledge (Google’s guide to Gems).
- In Copilot, in its custom instructions, where your organization has turned them on.
The custom instructions generator writes the standing part for any of the four and says where each one goes. Keep one master copy with a date on it, so the version in the GPT and the version in the Gem do not drift apart.
Test it on three jobs
A voice that works for a launch can fall apart in an apology. Before the team relies on the sheet, run three jobs that pull in different directions: a product announcement, an apology for a late order, and a refusal of a discount you will not give. Read each draft against the rules, not your mood, and fix the rule rather than the draft when the same miss appears twice.
Run the same three in two models side by side in Model Compare, because the sheet should survive a change of model. Then put each draft through the Robot-Speak Checker, which catches the stock phrases any model falls back on when a rule runs out. For day-to-day use, edit a draft without losing your voice and change the tone of an email both take the sheet as an input, and support teams can start with rewriting a saved reply so it sounds like a person.
The checklist
- Every adjective in the old guide has become one or more rules you can see on the page.
- The sheet names the reader in a sentence or two.
- It has a short list of your words and a short list of banned ones.
- It carries two samples under 200 words, with the features to copy named and a line saying not to reuse their content.
- It says how the tone changes for an apology, a launch and a refusal.
- It lives where every chat starts, and it passed three different jobs in two models.
Key takeaways
- Adjectives such as friendly or confident leave a model to guess, and it guesses the most typical version.
- Rewrite each adjective as rules two readers would agree a draft passed or failed.
- A voice sheet holds the reader, the rules, a word list, two samples, tone by situation and what to do when unsure.
- Two short, chosen samples beat a pile of fifty, which copies old content and dilutes the signal.
- Save the sheet where every chat starts, and ask the model to list any rule it could not follow.
- Test on an announcement, an apology and a refusal, in two models, before the team relies on it.
Questions
How do I train ChatGPT on my brand voice?
You do not need to train it. Give it standing instructions: a voice sheet with rules you can check, a word list, two short samples and the features to copy from them, saved in a project, a custom GPT or custom instructions. The same sheet works in Claude, Gemini and Copilot. The brand voice guide template drafts one from your best copy.
How many writing samples should I give AI?
Two or three short ones usually do better than a large pile. Models copy what is in examples, including their topics and mistakes, and long inputs make them less reliable. Choose pieces you would be happy to see copied, name the features to copy, and say not to reuse their content.
Where do I put brand voice instructions in ChatGPT, Claude, Gemini or Copilot?
Where every chat starts: a project’s instructions or a shared custom GPT in ChatGPT, a project or custom style in Claude, a Gem in Gemini, and custom instructions in Copilot. The custom instructions generator writes the standing part and says where to paste it.
Why does AI ignore my brand guidelines?
Usually because the guidelines are adjectives. Words like friendly and confident leave the model to guess, so it writes the most typical version of them. Rewrite each adjective as a rule you could check on the page, such as sentences under 20 words or no exclamation marks, and add two samples with the features to copy.
Prompts to use with this
- Write a brand voice guide from your best copyYou need tone rules, examples and words to avoid so every writer and AI tool sounds like your brand.