Tavimi

Guides

Brand voice and AI: how to keep one tone when the whole team writes with it

AI-generated copy drifts because each person writes their own prompt and the assistant doesn't know what the client has already approved or rejected. The fix: a short brand voice guide written for an AI (observable rules, words to use and to avoid, approved and rejected examples, rules per channel), stored in one place every assistant reads, and updated after every client sign-off.

Published

On Monday, Julie writes three LinkedIn posts for a client with ChatGPT. On Tuesday, Karim picks up the same client in Claude for the newsletter. The client reads both and replies: “This sounds like two different brands.” Nobody did a bad job. Each of them simply gave their assistant different context, and neither assistant knew that last month the client had rejected the word “solution” and exclamation points. This guide explains why tone drifts when several people write with AI, how to rewrite your brand voice guidelines so an assistant can apply them, and where to store them so the whole team starts from the same base.

Why AI-generated copy drifts when several people write

An AI assistant doesn’t write “in the brand’s voice”. It writes in the voice of whatever it was given in the conversation. Three mechanisms explain the drift.

  • Everyone has their own prompt. Julie pastes the brief and two old posts; Karim types “warm but professional tone”. The two assistants get different instructions and, unsurprisingly, produce different copy.
  • The assistant doesn’t know past sign-offs. The back-and-forth with the client lives in emails, comments on a PDF, or a project manager’s head. None of it reaches the conversation unless someone thinks to paste it in.
  • Assistant memory is personal. Claude, for example, remembers things across conversations and keeps a separate memory for each project. But that memory is tied to each user: Anthropic’s help center states that organization owners can neither see nor edit individual memories. What Julie’s assistant learned doesn’t help Karim’s.

On top of that, there’s wear: after round after round of edits, each piece moves a little further from the last, and nobody notices the slope.

Writing a style guide an AI can apply

Most brand style guides are written for humans: values, adjectives (“authentic, approachable, expert”), a few pages of inspiration. An assistant gets little out of them, because “approachable” doesn’t tell it whether to use contractions. A style guide an AI can use fits on one or two pages and contains checkable rules.

The five sections that matter

  1. Tone, as observable rules. Not “warm”, but: address readers as “you”, sentences of twenty words or fewer, at most one exclamation point per piece, no rhetorical question as an opener.
  2. Words to use. The brand’s words, the client’s words, the words the client uses with their own customers.
  3. Words to avoid. Words rejected during sign-offs, jargon the brand steers clear of, legally sensitive terms (“guaranteed”, “the best”).
  4. Approved and rejected examples, with the reason. This is the most effective section: a piece the client accepted shows the tone better than any description. A rejected piece, with the reason, shows where the line is.
  5. Structure per channel. A newsletter, a LinkedIn post and a product page don’t share the same length or the same call to action.

The Claude Code docs offer advice that holds for any instruction you give an assistant: the more specific and concise the instructions, the more consistently they’re followed. A twenty-page style guide gets applied less reliably than a one-page sheet.

Example voice sheet (fictional brand)

# Voice sheet: Cidrerie du Vallon
Updated: September 12, 2026, approved by Claire M. (client)

## Tone
- Address readers as "you" everywhere, Instagram included. No slang.
- Short sentences (20 words max). At most one exclamation point per piece.
- Talk about the orchard, the seasons and the people. No superlatives.

## Vocabulary
- Use: farmhouse cider, orchard, cider apples, the cidery, our growers.
- Avoid: "solution", "artisanal" (rejected Sept. 4: seen as overused),
  "premium", "must-have", any marketing jargon.

## Examples
- Approved: "The first apples have fallen. Pressing starts Monday."
- Rejected: "Discover our premium artisanal cider, the must-have for your happy hours!"
  Reason: three banned words, ad-like tone.

## Per channel
- Newsletter: 150 words, one topic, signed "The Vallon team".
- LinkedIn: 600 characters, a concrete fact in the first line, no emoji.
- Product page: apple origin, pairings, storage. No call to action.

This sheet takes a minute to read, can be checked point by point, and says where each rule came from.

Where to store the style guide so every assistant reads it

A good sheet is useless if everyone keeps a different copy on their laptop. You need one location that the assistant reads before it writes. Here are the options available today in Claude and ChatGPT, based on their help pages.

Location What it does well Where it stops
Custom instructions (profile preferences in Claude, “Custom Instructions” in ChatGPT) Apply to all your conversations, nothing to paste again Tied to your account and shared across all your topics: a poor fit once you handle several brands
Projects (Claude and ChatGPT) Instructions and files specific to one project, so to one brand Sharing limited to Team and Enterprise plans in Claude; open to every plan in ChatGPT
Skills (Claude) A reusable instruction, such as brand rules, that can be shared with colleagues Sharing on Team and Enterprise, enabled by an owner; requires code execution
File in a repo (CLAUDE.md for Claude Code) Shared through version control, loaded every session, can import other files Only for people who work with a coding tool
MCP server A shared source the assistant reads and updates, whatever the account You have to choose or host a server, and connect it in each client

A few details:

  • Projects. In Claude, a shared project can grant “Can view” access (read and chat) or “Can edit” access (change instructions and knowledge). Put the voice sheet in the project knowledge and sum it up in the instructions: “Before writing anything, apply the attached voice sheet.” In ChatGPT, a project can be shared within a workspace and its instructions apply to every conversation in the project.
  • MCP server. Claude accepts custom connectors to a remote MCP server on all its plans (only one on the free plan); on Team and Enterprise, an owner adds it, then each member connects. ChatGPT also accepts custom MCP servers in developer mode, in beta, with permissions (read-only or write) that vary by plan: check OpenAI’s help page for your subscription.
  • The common thread. Whichever option you pick, name one person who edits the sheet, and keep a single reference version. Two copies always drift apart.

Keeping a record of client sign-offs

The voice sheet states the rule; you also need to keep why it exists. Otherwise the “no artisanal” rule gets dropped by the first newcomer who finds it arbitrary.

Keep a simple log of editorial decisions:

  • the date and who signed off on the client side;
  • the piece in question (or an excerpt);
  • the decision: approved, rejected, approved with changes;
  • the resulting rule, copied into the sheet.

Example entry: “September 4, Claire M. Rejected the ‘premium artisanal cider’ post. Reason: ‘artisanal’ seen as overused, ‘premium’ off-positioning. Rule added: both words move to the avoid list.”

Two habits make this log useful to the AI. First, turn every rejection into a rule or a rejected example in the sheet: that’s how the assistant stops repeating the mistake. Second, never delete anything: mark a dropped rule as dropped, with the date, so older copy still makes sense.

Agencies handling several clients

An agency adds one more risk: mixing brands up. A luxury client’s tone bleeding into an insurer’s newsletter, one brand’s vocabulary slipping into another’s.

  • One location per client, never brand rules in global instructions. Custom instructions apply to all your conversations: putting one client’s tone there forces it on every other client.
  • Separate the agency’s method from the client’s voice. The method (review process, delivery format, house typography rules) applies to everyone; the voice belongs to each client. Two separate documents.
  • Name things unambiguously. “Voice sheet: Cidrerie du Vallon”, not “tone v3 final”. The assistant and the intern alike need to know which brand they’re dealing with.
  • Have it state the client up front. Ask the assistant to start by naming the sheet it’s applying. A wrong client then shows up in the first line.
  • Restrict access. A freelance writer working on a single client doesn’t need to see the other clients’ sheets, or the internal notes about them.

Where Tavimi can help

The options above are often enough for one person, or for a team that already has a Team plan. Tavimi is for the case where several people, and their assistants, need to share the same style guide and the same history of sign-offs, client by client.

Tavimi is the shared working memory for a team and its AI assistants, over MCP. Each client gets a space, and you invite the people involved by email; an invited person only sees that space. The voice sheet lives there as a versioned note: nothing gets overwritten, and every old rule can be found again. Sign-offs can be filed under the “decision” category, where you can require headings (date, approved by, reason): an incomplete note is rejected. Claude, connected through an MCP connector, reads the sheet before writing and, when it’s missing a choice, asks the question on the task instead of guessing; only a human answers.

To see how this works in an agency, see the marketing agencies page.

FAQ

What people ask us

Should I give the AI the full brand guidelines?

No. A forty-page document buries the instructions that matter. Pull out the editorial part (tone, vocabulary, examples, rules per channel) into one or two pages, and keep the full PDF as a reference for humans.

Are ChatGPT or Claude custom instructions enough for a brand?

For one person working on a single brand, they're a good start. But they apply to every conversation on your account and aren't shared: as soon as several people are involved, or you handle several clients, you need a shared location specific to each brand.

How do I check that the AI actually followed the style guide?

After each piece, ask it to review its output against the voice sheet and flag any deviations: banned words, sentences that run too long, a missing call to action. Human review is still needed, but it can focus on substance instead of hunting for verbal tics.

What should we do when the client changes their mind about the tone?

Update the sheet, date the change and note the reason. Keep the old rule in the history instead of deleting it: six months later, someone will ask why the brand stopped using emoji on Instagram.

Give your team and your AI assistants one shared memory

Tavimi keeps the decisions, notes and tasks of every project, and Claude reads them back over MCP.

The Tavimi web app is in French for now; an English interface is on its way. Your assistants work in any language.