To set up ChatGPT brand voice for teams, put one voice document, real examples and clear rules into a shared Project in ChatGPT Business (formerly ChatGPT Team) or Claude, with one project per client. Keep instructions short, store long references as files, test with three standard prompts, and give one person ownership of updates.
Key takeaways
- Put the brand voice in a shared Project, not in each person's custom instructions, so every seat draws from the same source.
- OpenAI says it plans to retire custom GPTs, so build new brand voice setups in Projects and keep your voice document outside any one tool.
- Examples beat adjectives. Three to five approved pieces per format, plus one bad example with a rewrite, teach tone better than any description.
- Agencies should run one project per client, built from the same template, shared only with the people on that account.
- Claude cannot see a project's name or description, so write the client and brand name inside the project instructions.
- Give one person ownership, test with the same prompts across teammates, and log every change.
Five people on your team use the same AI tool and get five different voices. One gets stiff copy, one gets hype, one gets something close to right. The fix for ChatGPT brand voice for teams is not a better prompt saved in someone's notes app. It is one shared setup that every seat draws from, with the voice, the examples and the rules in one place. This guide covers where that setup should live in ChatGPT Business and Claude, what to put in it, how to keep each client separate, and how to keep it from drifting.
Where the shared voice should live
First, the name. OpenAI renamed ChatGPT Team to ChatGPT Business on August 29, 2025, and says "all features, pricing and limits remain the same" (OpenAI). If an older guide says Team, it means Business.
There are three common places teams put a brand voice: personal custom instructions, a custom GPT, or a shared project. Personal custom instructions belong to one person, so each seat drifts on its own. That leaves GPTs and projects.
ChatGPT Business (formerly ChatGPT Team)
A ChatGPT Project keeps related chats, files and instructions together, and ChatGPT can draw on everything in a shared project, including chats, uploaded files and custom instructions (OpenAI). On Business plans a project can hold up to 40 files. When OpenAI launched shared projects for business plans, it named "creating content" and "supporting a customer account" as uses, said client details stay within the project, and noted that on ChatGPT Business they can only be shared with workspace members (OpenAI).
Most older guides say "build a brand GPT." That advice is aging. OpenAI's help page now says it is "planning to retire custom GPTs" and move creators' workflows to plugins, which combine reusable instructions, reference files and connected apps. The same page says a GPT becomes read-only once it is migrated, and that in Business workspaces the ability to create, edit and publish GPTs depends on product availability, workspace settings and permissions (OpenAI). Check that page for the current timeline in your workspace.
My advice: build new brand voice setups in a shared Project. If you already have a brand GPT that works, keep using it, but copy its instructions and files into a voice document you own. Then any future move is copy and paste.
Claude Projects
In Claude, you open a project and use "Set project instructions." Claude applies those instructions to every chat in the project, and anything you upload to the project knowledge base is used across all of the project's chats (Anthropic). Two details matter for teams. Claude cannot see the project's name or description, so naming a project "Acme Voice" tells Claude nothing. And on Team or Enterprise plans you choose visibility when you create the project; "Keep it private" limits it to you and the people you invite.
| Where the voice lives | Who shares it | Good for | Watch out for |
|---|---|---|---|
| Personal custom instructions | One person | Individual preferences | Every seat drifts apart |
| Custom GPT (ChatGPT Business) | Workspace, per settings | Existing setups that work | OpenAI plans to retire custom GPTs |
| Shared ChatGPT Project | Workspace members you share with | One brand or client per project | 40-file cap; off-topic chats in the project become context |
| Claude Project | Set by visibility at creation | One brand or client per project | Claude cannot see the project name |
What goes in the voice document
Write the voice document outside the AI tool first, in a Google Doc or similar. That copy is the source of truth. The project gets a short version as instructions and the long version as a file. You can use this template for every brand or client.
| Section | What to write | Example line |
|---|---|---|
| Who we are | Two sentences on the brand and what it sells | We sell refillable cleaning products to busy parents. |
| Audience | Who reads this and what they care about | Parents with little time who distrust green claims. |
| Voice traits | Three to five traits, each as "this, not that" | Warm, not cute. Confident, not loud. |
| Words we use | Product names, preferred terms, spelling | Say "refill," not "pod." |
| Words we never use | Banned words and phrases | No "eco-warrior," no "game-changer." |
| Formatting | Length, headings, punctuation, emoji policy | Subject lines under eight words. No emoji. |
| Facts and claims | What the AI may and may not state | Never invent numbers. Mark unknowns as [CHECK]. |
| Examples | Approved pieces, labeled by format | See the examples file. |
| When unsure | What to do instead of guessing | Ask one question before drafting. |
Keep the instructions field tight. In my experience, instructions under a page get followed more consistently than long ones, and the long material (full style guide, product sheets, past campaigns) works better as uploaded files the AI can look up.
Examples do more work than adjectives
"Friendly but professional" means something different to every person and every model. Examples settle the argument. For each format your team writes often, add three to five approved pieces with a one-line note on why each one works.
Then add one bad example and its rewrite. Here is an illustrative pair for a made-up cleaning brand:
- Bad: "Revolutionize your cleaning routine with our amazing, planet-saving refills."
- Good: "One refill, one bottle, a year of clean counters. Less plastic under your sink."
- Why: Concrete benefit, short sentences, no hype words, no exclamation point.
Label every example clearly ("Approved email intro," "Bad example, do not copy") so the AI knows which is which. When output drifts later, add a new example before you add a new adjective.
Rules that stop the common drift
Rules handle what examples cannot: hard limits. The ones that matter most for marketing teams:
- Banned words. List the brand's never-words and the generic AI phrases your team keeps deleting.
- Claims. No invented statistics, prices, results or quotes. Anything unverified gets marked [CHECK].
- Format. Length limits per channel, heading style, punctuation and emoji policy.
- Legal and compliance. Required disclaimers and words the client's legal team has ruled out.
- Priority. Say which wins when rules conflict, for example "Compliance rules override voice rules."
In my experience, a short list of "do this" rules plus a tight list of "never" rules works better than a long wall of prohibitions.
One project per client
For agencies, the rule is simple: one project per client, never two clients in one project. Because ChatGPT draws on everything in a shared project, mixing clients means one client's files and chats can shape another client's copy.
Build every client project from the same template so anyone on the team knows where things are:
- Instructions: the short voice document, starting with the client's name and what they sell. This matters most in Claude, which cannot see the project name.
- Files: the full style guide, examples file, product or service sheet, and any compliance notes.
- Sharing: only the people on that account. In ChatGPT Business, shared projects stay inside your workspace. In Claude Team, pick "Keep it private" and invite the account team.
- Naming: a consistent pattern such as "Client Name Voice" so projects are easy to find.
If your agency also has its own house voice for proposals and your own marketing, give it a separate project too.
How to set up ChatGPT brand voice for teams, step by step
Copy this checklist and run it once per brand or client:
- Name one owner for the voice. Only that person edits the instructions.
- Collect 10 to 20 pieces the client or brand has approved, across your main formats.
- Fill in the voice document template above.
- Create a shared Project in ChatGPT Business or a Project in Claude.
- Paste the short voice document into the project instructions, with the client name in the first line.
- Upload the full style guide, examples file and product sheets as project files.
- Share the project with the account team only.
- Run three standard test prompts, such as a social post, an email intro and a rewrite of a weak draft.
- Have two teammates run the same prompts and compare. Fix gaps by adding examples or rules.
- Tell the team that client work starts inside the project, not in a blank chat.
- Keep a short changelog in the voice document and review it monthly.
Keep it current
A voice setup is not finished on launch day. Brands change offers, legal adds words, and new formats appear. Rerun your three test prompts after every change to the instructions, and compare against the last results.
Watch for one quiet failure: people drafting in blank chats because the project feels slower. Ask for drafts to be shared as links from the project so you can see where work starts.
Once the voice is stable, the same document can feed automations. If you build reporting or content workflows in Make, pass the voice document into the prompt step so automated output matches what people write by hand. Our guide to automating client reporting with Make and Claude shows where that fits.
Doing it yourself versus getting help
Most teams can do this themselves. Plan a few focused hours per brand for the first setup, and less for each client after that once the template exists. The checklist above is the whole method.
It makes sense to get help when you have many clients to set up at once, when the setup exists but the team does not use it, or when you want the voice wired into automations. That is what Keeplight's Brand Brain Setup does: your team's Claude or ChatGPT Business set up with the brand's voice, examples and rules, with everything created in your name. If you are not sure brand voice is the right first project, an AI Workflow Audit looks at every recurring workflow first; you can see what that report looks like in the sample audit. Or book a free 20-minute intro call and we can talk through your setup.
Frequently asked questions
Is ChatGPT Team the same as ChatGPT Business?
Yes. OpenAI renamed ChatGPT Team to ChatGPT Business on August 29, 2025, and says features, pricing and limits stayed the same. Any guide that mentions Team applies to Business.
Should I build a custom GPT or a Project for our brand voice?
Build new setups in a Project. OpenAI's help pages say it is planning to retire custom GPTs and move workflows to plugins. If you already have a brand GPT that works, keep using it, but save its instructions and files in a document you own so moving later is easy.
How many examples should a brand voice setup include?
In my experience, three to five approved pieces per format is enough to set the tone, plus one bad example with a rewrite. Label each one clearly. Add more examples when output drifts rather than adding more descriptive adjectives.
How do I keep different clients' brand voices separate?
Create one project per client and never mix clients in one project. Share each project only with the people on that account. In Claude, choose private visibility and invite the account team, and put the client name in the instructions since Claude cannot see the project name.
Is Claude or ChatGPT better for brand voice?
Both work well when set up the same way: short instructions, labeled examples and reference files inside a project. Pick the tool your team already pays for and uses daily. Consistency comes from the setup and the habit of starting work in the project, not from the brand of model.
Bryce Walls, Founder, Keeplight
Bryce runs marketing for a global consumer brand and builds the AI systems he uses every day. He has run paid media, analysis and marketing ops in-house for more than seven years, and Keeplight sets up the same kind of AI and automation for small marketing teams.