How to keep client brand voices separate when using AI

You write posts for two or three clients with ChatGPT or Claude, and by now they all start to sound the same. Here is how to keep client brand voices separate when using AI: one project per client with its memory kept inside, a one-page voice file built from the client's own words, and a quick check on every post. Without it, a phrase or a fact from one client ends up in another client's post, and the client is the one who spots it.

Do these three things

About 2 hours per client today, then about 12 minutes a post and 15 minutes per client each Sunday. Judge the setup after 40 posts per client, not after one awkward post.

Show the three actions in fullHide the full list

1. 1Today: make one ChatGPT or Claude project for each client, keep its memory inside that project, and load it with the client's one-page brand voice file and 5 posts they approved.
2. 2Every day: draft every client post inside that client's project only, 3 versions each, and send a post for approval only after a blind check names the right client.
3. 3Every Sunday: turn each client's voice edits from the week into one new rule or banned word in their voice file, and count voice edits per 10 posts.

Get help doing these

ChatGPT or Claude reads this guide and works through it with you.

Do this with ChatGPT(opens in a new tab)Do this with Claude(opens in a new tab)

1. Make one ChatGPT or Claude project for each client and keep its memory inside it

Make one project per client in ChatGPT or Claude, and set each one so it cannot read your other chats. A project is a folder in the chat app with its own instructions, files and chats. The memory setting matters most: by default, both apps can carry what you said in one chat into another, and that is how one client's words end up in another client's post.

Pick one app for all your clients. Both work on the free plan. ChatGPT's free plan has as many projects as you like, with 5 files in each. Claude's free plan has up to 5 projects. If you have more than 4 clients and want Claude, pay for Claude Pro ($20 a month); otherwise use ChatGPT. Do not split clients across two apps. You will forget which client lives where.

In ChatGPT:

1. Click New project in the sidebar. Name it after the client, such as "Client: [name]", and give each client its own colour.
2. Open the project's menu (the three dots), click Project settings, and under Memory choose Project-only memory. Click Save. OpenAI says the change can take a few hours to apply.
3. With project-only memory, the project does not use your saved memories, it sees only the chats inside it, and your other chats cannot see into it. With default memory, on the free and Plus plans, a project's chats can reach your chats outside it.

In Claude:

1. Make a new project per client, named the same way.
2. On the project's page, under Memory, turn off Use account memory. Each Claude project already has its own memory. This switch stops it from also using what Claude remembers from your chats outside projects.

Then turn off training on your chats. You will paste in client plans that are not public yet. In ChatGPT, go to Settings, then Data controls, and turn off Improve the model for everyone. In Claude, open your Privacy Settings and turn off Model Improvement. Never paste passwords or a client's customer data. If the client's agreement or scope of work says how you may use AI, follow it.

Make one more project for your own agency's marketing, set up the same way. Your own LinkedIn posts and cold emails should never sit in a client's project either.

If a contractor drafts for you: in ChatGPT, open one client's project, click Share, invite the contractor by email and give them chat access. They see that client's chats, files and instructions and nothing else. Shared ChatGPT projects always use project-only memory, and the free plan allows 5 people per project. Claude shares projects only on its Team and Enterprise plans. When you hire your first contractor, give them one client's project at a time.

• Tool: ChatGPT or Claude, free plan.
• Time: 30 minutes for all your clients.
• You will have: one project per client and one for your agency, each with its memory kept inside, and training turned off.
• It worked if: you open client 2's project, ask "What do you know about [client 1's name]?", and it says it knows nothing. If it answers with anything true, the memory setting has not applied yet. Wait a few hours and test again before you draft.
• Common mistake: drafting "just this one post" in a plain chat outside any project. That chat goes into your account memory, and from there into the next client's drafts.

2. Write a one-page brand voice file for each client from their real words

Write one page per client that says who they write for, how they sound, what is true about them, and what they never say. AI copies examples and rules much better than adjectives. "Friendly and professional" fits every client you have, so it keeps none of them apart.

Collect the client's own words first. You need three things:

• 10 pieces the client wrote or loved: their best LinkedIn posts, their homepage, a sales email, a case study. If they have customer quotes, add 3 of those too.
• The "How you sound" answers (questions 8 to 10) from your onboarding questionnaire, and the words they used on your discovery call. If they have ever run customer interviews, take the buyers' exact phrases from the notes.
• The Brand notes tab of their approval sheet: every change they have asked for so far.

Let AI draft it, then you fix it. Open the client's project, paste the three things, and paste the template below. Ask it to fill in every line it can and write "[ASK CLIENT]" where it cannot tell. Then write the three voice rules yourself. Each rule needs a "Like this" line in the client's real words and a "Not like this" line showing the mistake AI keeps making for them.

Never put another client's name, words or posts into this file. The "Not like this" lines show generic AI text, not your other clients. Facts come from the client's ideal customer profile, their value proposition and your notes, never from the AI's guesses.

Client: [name] · Approved by [name] on [date]Who we write for: [kind of buyer], [size], who [the moment the problem hurts].
Who is speaking: [the founder, in first person / the company, as "we"].3 voice rules
[Rule, such as "Lead with a number from our own product."]
Like this: "[a real line from the client]"
Not like this: "[the generic line AI writes]"
[Rule]
Like this: "[real line]"
Not like this: "[generic line]"
[Rule]
Like this: "[real line]"
Not like this: "[generic line]"
Shape: sentences under [15] words · emojis: [none / one at the end] · hashtags: [none / up to 3] · ends with: [a question / nothing / a link]Words we use: [10 words and phrases their buyers and team really say]
Never write: [10 words, claims or topics, including competitors' names and any feature names they do not want public]True facts we can use (never invent others):
[price, number, result or customer line, with where it came from]
... up to 20.
If a fact is missing, write [FACT NEEDED]. Never invent numbers, quotes, customers or results.

Send the file to the client's approver with one line: "This is how I will write for you. Please correct anything that is wrong by [day, 3 days from now]." The same message works when you take on a client who already has posts you wrote for them.

• Tool: Google Docs, and the client's ChatGPT or Claude project to draft it.
• Time: about 1 hour per client.
• You will have: one voice file per client, corrected and approved by the client.
• It worked if: the file has 3 rules, each with a "Like this" and a "Not like this" line, 10 words to use, 10 never-write items and at least 15 true facts, and the client has said yes to it.
• Common mistake: a file full of adjectives ("bold, warm, expert") and no example lines. The other mistake is copying one client's file to start the next one. Old lines stay in by accident.

3. Load each client's voice file and 5 approved posts into their project

Put the voice file into the client's project instructions and add one document of 5 approved posts. Now every chat in that project starts from that client's voice, and you never paste it in again.

In ChatGPT, open the project's menu, click Project settings and paste the voice file into the instructions, with the block below under it. Project instructions apply only inside that project and override your general custom instructions. In Claude, paste the same text into the project's instructions, and add the document to the knowledge base on the right of the project's page.

You write only for [client name]. Use only this voice file and the files in this project.
Use only the true facts listed above or given to you in the chat. If a fact is missing, write [FACT NEEDED].
Never mention any other company, client or product unless I name it in the chat.
Before you answer, check every line against the 3 voice rules and the never-write list.

Put the 5 posts in one document. Pick the 5 posts the client approved fastest, with no changes, or 5 of their own best posts if you are new. Paste them into one Google Doc and upload it as one file. ChatGPT's free plan holds 5 files per project, so one document leaves room for their content calendar and product pages. If you write more than one kind of thing for a client, such as posts and blog articles, add 5 of each to the same document under a heading.

Test it with the same brief in two projects. Paste one short brief, such as "A LinkedIn post about what new users ask in their first week", into two clients' projects. Compare the two posts side by side.

• Tool: your ChatGPT or Claude projects, and Google Docs.
• Time: 20 minutes per client.
• You will have: every client project loaded with its voice file, its instructions block and one document of 5 approved posts.
• It worked if: you can tell which project wrote which post from the first line, and neither post uses a word from that client's never-write list. If the two posts read alike, make the 3 voice rules more different, then test again.
• Common mistake: pasting another client's post into a project "as a style reference". The project now holds that client's words and can reuse them.

4. Draft every post inside its client's project, 3 versions each

Every day, draft each client's posts for the day inside that client's project, and ask for 3 versions of each. Choosing the best of 3 takes a minute, and it gives you a post that already sounds like the client instead of one you have to rewrite.

Before you paste anything, read the project name at the top of the screen. Most mix-ups happen here: the right brief, pasted into the wrong project.

Give it one true detail every time. A post built only on the topic comes out as the average post on the internet. Add one thing only this client has: a number from their dashboard, a question a customer asked, a line from the founder's last call. The LinkedIn post ideas for B2B companies and post ideas for SaaS companies give you topics. The detail has to come from the client.

Write 3 versions of a [LinkedIn / X / Instagram] post for [date].
Topic: [one line]
True detail to use: [the number, question or line, and where it came from]
Goal: [what the reader should do or think after reading]
Follow the voice file. Mark any fact you are not sure of with [FACT NEEDED].

Draft every post in the client's package. If a client pays for 2 posts a day on a channel, that is 2 briefs and 6 versions a day for that client. AI writes the 3 versions in under a minute. You spend the time choosing one, adding the detail, and cutting what the client would never say. Work client by client, and finish one client's posts before you open the next project. Add the daily brief to your checklist for client posts, so anyone who drafts for you starts the same way.

The same rule covers everything you write for a client, on X, Instagram and everywhere else: short video scripts, repurposed articles, replies on Reddit, ad creatives for the clients you run ads for. All of it starts in that client's project. LinkedIn said on 21 May 2026 that "It's ok to use AI to help you write, but your posts and comments need to represent your voice and your perspectives", and it shows fewer AI posts that have no clear view of their own. The one true detail is what gives each post its view.

• Tool: the client's ChatGPT or Claude project.
• Time: about 10 minutes a post, including choosing and editing.
• You will have: every post for today drafted in the right project, with one version chosen and edited.
• It worked if: every fact in the post comes from the voice file or today's brief, and the version you kept needed only small edits. If you rewrite more than half of a post most days, the voice rules are too vague: fix them on Sunday (step 6).
• Common mistake: keeping one long chat per client for months. Start a new chat in the project each day. The project keeps the voice file, and a fresh chat drops yesterday's mistakes.

5. Run a blind voice check before any post goes to the client

Before a post goes to the client for approval, paste it, with every name taken out, into a separate "Voice check" project that holds only your clients' voice files. Ask which client it sounds like. If it names the wrong client, or cannot tell, the post is not ready.

Set up the Voice check project once. Make one more project, with project-only memory in ChatGPT or account memory off in Claude. Upload every client's voice file into it, and nothing else: no posts, no plans, no chats with client work. Never share this project with a contractor, because it holds every client's facts.

Check every post. Remove the client's name, product names and people's names, then paste the post with this prompt:

Here is a post with the names removed:
[post]Score it from 1 to 5 against each voice file in this project: how well does it match that client's voice rules, words and shape?
Then name the best match, and list every line that breaks that client's voice rules or never-write list.

The pass mark: 4 or 5 for the right client, and 2 or lower for every other client. If you have one client, the post needs a 4 or 5 and no broken lines. If it fails, go back to the client's project, rewrite the lines it listed, and check again.

Then run the 8-line check from the approvals guide for facts, links and platform rules. Line 6 of that check is the voice line, and this step is how you answer it.

• Tool: a "Voice check" project in the same app.
• Time: about 2 minutes a post.
• You will have: every post scored before it is marked Ready for client, with broken lines fixed.
• It worked if: after 40 posts per client, at least 9 in 10 pass the check the first time. Below that, the voice rules of the two clients it mixes up are too alike: make one rule in each file more different.
• Common mistake: asking "Does this sound like the client?" in the project where you drafted it. That project only knows one client, so it always says yes.

6. Update each client's voice file every Sunday from their edits

Every Sunday, read every change each client asked for this week, find the ones about voice, and turn them into one new rule or one new never-write word in their file. A client who has to correct the same thing twice starts to wonder what they are paying for.

Find the voice edits. Open each client's approval sheet and filter for the posts that came back as Changes asked. Tag each comment "Voice" (a word, the tone, how it ends, who is speaking) or "Fact" (a number, a name, a claim). Fact edits go into the facts list. Voice edits are what this step is for.

Add one rule, and swap one example. Paste the week's voice edits into the client's project and ask: "What do these changes have in common? Write it as one voice rule with a Like this and a Not like this line." Check it, and add it to the file and the project instructions. Then swap the weakest of the 5 example posts for the post the client approved fastest this week.

Count two numbers per client: voice edits per 10 posts, and posts that failed the blind check the first time. Put them in a Sunday row next to the approval numbers you already count, and bring them to your monthly review. Use these rules:

• More than 2 voice edits per 10 posts: book a 15-minute call. Show the client 5 posts and ask them to mark each one "sounds like us" or "doesn't", and to say why. Write their words into the file.
• The same edit twice in one month: add it to the never-write list that day, not next Sunday.
• 2 voice edits or fewer per 10 posts for 4 Sundays: stop changing the file for that client, and only add new facts. Ask the client for one line on what it is like to work with you: it is proof for your next client.
• Still more than 2 per 10 after the call and 4 more Sundays: the client is paying for more rounds than the price covers. Raise the price at renewal, or add a third round as a paid extra.

Voice edits that never stop are hours you are not paid for. They show up in your hours per client, they belong in your monthly report as "approved with no changes", and a client who corrects the same thing every week is one you are close to losing. Once each client's file is steady, you can take on more clients without their posts starting to blur.

• Tool: the approval sheet, Google Docs and the client's project.
• Time: 15 minutes per client, every Sunday.
• You will have: one new rule or word per client per week, a fresh example post, and two numbers in the Sunday row.
• It worked if: after 40 posts per client, voice edits per 10 posts have gone down for 4 Sundays in a row.
• Common mistake: rewriting the whole file every week. Change one rule at a time, or you cannot tell which change helped.

Where Scout7 helps

Scout7 does part of steps 1 to 4 for each client brand.

Steps 1 to 3. Scout7 gives each client brand its own workspace, with its own brand profile, connected accounts and calendar, switched from the top of the app. How many workspaces you get depends on your plan. When you add a client's website, Scout7 reads it and builds the brand profile: positioning, audience, voice, visual preferences and guardrails, held as facts you can edit and filter by voice and guardrails. You correct it with the client's voice file and add the facts it could not see. A workspace only reuses facts someone recorded, so keep it in step with the file each Sunday.

Step 4. Scout7 makes image posts with captions in the brand voice of the workspace you are in, and drafts replies to live conversations in that voice. You approve the idea and the finished post or reply, and you still run the blind voice check. Scout7 also connects to Claude and ChatGPT, so you can work in a client's workspace from the assistant you already use. Nothing goes live without a person approving it.

Your client voice checklist for this week

0 of 11 done

Work through this list with an AI assistant

It drafts the work and keeps track of what is done.

Do this with ChatGPT(opens in a new tab)Do this with Claude(opens in a new tab)

Step 1 · Make one project per client and keep its memory inside it

• Make one ChatGPT or Claude project per client and one for your agency, with memory kept inside each.
• Run the leak test in every client project.

Step 2 · Write a one-page brand voice file for each client

• Write each client's voice file from 10 of their real pieces, their form answers and their past edits.
• Send each voice file to the client and get it approved within 3 days.

Step 3 · Load each voice file and 5 approved posts into the project

• Load the voice file, the instructions block and 5 approved posts into each client's project.
• Test one brief in two client projects and compare.

Step 4 · Draft every post inside its client's project

• Every dayDraft every client post for today in that client's project, 3 versions each.

Step 5 · Run a blind voice check before any post goes to the client

• Set up the Voice check project.
• Every dayRun the blind voice check on every post before it goes to the client.

Step 6 · Update each client's voice file every Sunday

• Every SundayTurn each client's voice edits into one new rule and swap one example post.
• Every SundayCount voice edits per 10 posts and check fails for each client.

Copy listDownloadReset