Guides

How to build a brand voice sheet (that AI can use)

By Rory Wilmer, InkByte · 6 min read · 16 October 2026

Most brand guidelines have a tone-of-voice page. Most of those pages say the brand is "friendly, confident and human". That's not much use to a copywriter and even less use to ChatGPT or Claude. A brand voice sheet is the working version: one or two pages, specific enough that a new writer, or an AI tool, could produce on-brand copy from it on day one.

Here's how to build one for a client, step by step.

1. Collect real copy before you write a word

Don't start from adjectives. Start from evidence. Gather the client's best recent copy across channels: homepage and about page, three to five recent social posts, a newsletter, a product page, anything the client is proud of. Ask the client, or the account lead, which pieces feel most "them" and which feel out of date.

If the website and the newer material disagree, the newer copy usually reflects where the brand is going. Note the conflict and confirm it with the client rather than picking one quietly.

2. Read for patterns, not vibes

For each piece, note what it actually does on the page:

• Sentence length: short, mixed or long?
• Person: "we" and "you", or third person?
• Punctuation: dashes, exclamation marks, emoji, the Oxford comma?
• Humour: none, dry or playful, and where it comes from?
• How they handle claims: numbers and proof, or big promises?
• Calls to action: "Shop now" or "Have a look"?

After a handful of pieces, patterns appear. Those patterns are the voice.

3. Turn patterns into three traits

Three traits is a workable number. Each one should be specific to this client, and each needs three parts:

TRAIT: Plain-spoken
It means:     Short sentences. Facts first. Names the thing.
It doesn't mean: Blunt, cold or curt with customers.
Evidence:     "Good coffee, no ceremony." (homepage)

The "doesn't mean" line matters most. It stops a trait being pushed too far: "playful" becoming silly, "confident" becoming arrogant. Mailchimp's published style guide is a good model here: alongside its voice traits it says things like "We're weird but not inappropriate, smart but not snobbish."

4. Write how the tone shifts

Voice stays the same; tone changes. Mailchimp's guide explains it as speaking differently to close friends at dinner and to your boss in a meeting, while still being the same person. Your sheet should say how tone moves across the channels the client actually uses:

Social:        lightest. Short. One joke, max.
Email:         warm, direct, one clear ask.
Customer care: calm and plain. No jokes.
Press:         factual, third person, no superlatives.

5. List approved phrases, word for word

Pull lines the client already uses and loves: straplines, sign-offs, recurring phrases. Copy them exactly, and note where each one came from. Don't tidy them. If the client writes "Have a look" rather than "Take a look", that difference is the voice.

6. List banned words and category tropes

Two lists here. First, the client's own no-go words: legal or sensitive terms, competitor names, anything that has caused trouble. Second, the tired phrases every brand in the category uses. Read two or three competitors' sites and note the lines they all share. Those are the ones to avoid.

For general corporate filler, plain-English lists help. The GOV.UK style guide's "words to avoid" list (for example "leverage", "streamline", "utilise") is a sensible baseline for most brands.

Give a reason for each ban, and an alternative where you can. Anthropic's prompting guidance says AI models follow rules better when they understand why, and when told what to do instead of only what not to do. People do too.

7. Add before/after examples

This is the most useful page of the sheet. Take three messages from the channels that matter most. For each, show a generic draft (the kind ChatGPT writes with no context), then an on-voice rewrite, then one or two lines on what changed and which trait it shows. Writers learn faster from contrast than from description.

8. Add a list of approved facts

Write down what the client is happy to see in copy: founding year, location, awards, product specifics, all checked against their own material. Anything not on the list either gets confirmed or doesn't get written. This keeps everyone, and every AI tool, from inventing details.

9. Finish with a quick rubric

End with a pass/fail checklist someone can run in 30 seconds on any draft. For example: no banned words; at least one specific fact instead of an adjective; sentence length matches the trait; tone fits the channel; nothing a competitor could publish unchanged. If it's quick, people use it.

10. Make it AI-ready

Once the sheet exists, turning it into AI instructions is mostly compression. Put the traits, banned list and facts into ChatGPT Custom Instructions (OpenAI allows 1,500 characters on Free and Go, 5,000 on paid plans) or into a Claude Project's instructions, and add the examples to the project knowledge. Anthropic recommends three to five varied examples, wrapped in tags so the model can tell them apart from the instructions.

If you write for more than one client, keep each in its own Project. OpenAI's Custom Instructions apply to all your chats, so a second client would inherit the first one's voice.

A one-page template

CLIENT:            [name] · [what they do] · [main audience]
TRAITS (x3):       means / doesn't mean / evidence
TONE BY CHANNEL:   [channel]: [how it shifts]
APPROVED PHRASES:  "[exact line]" ([source])
BANNED WORDS:      [word]: [reason] → [use instead]
CATEGORY TROPES:   [cliché competitors use]
FACTS WE MAY USE:  [fact] ([source])
BEFORE / AFTER:    generic → on-voice → why
30-SECOND CHECK:   [5 pass/fail questions]

Build it once, keep it updated when the client's copy changes, and every brief after that starts in the right voice.