Guides

How to make ChatGPT and Claude write in your client's voice

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

If you write for several clients, you already know the pattern. You ask ChatGPT or Claude for a launch email, and what comes back is fine. It's tidy, upbeat and grammatical. It also sounds like nobody in particular, and certainly not like the client who signs off your invoice.

The fix isn't a cleverer one-line prompt. The model writes generically because it has no idea who your client is. Your job is to give it that knowledge in a form it can actually use, and to put it somewhere it will be read every time. This guide walks through how to do that in both tools.

Step 1: Decide where the voice lives

Both tools give you a place to store standing instructions, and choosing the right one matters more than most people realise.

ChatGPT Custom Instructions sit in your settings (Personalization on web and desktop, Customize ChatGPT on mobile). OpenAI's help centre says they are "applied immediately to all chats". That's ideal if you write for one brand. If you juggle five clients, it's a problem: client A's voice will leak into client B's copy. OpenAI also caps their length. Free and Go users get up to 1,500 characters, and Plus, Pro, Business, Enterprise and Education users get up to 5,000. That's enough for a tight voice summary, not a full brand book.

Claude Projects are self-contained workspaces with their own chat history, knowledge files and project instructions. Anthropic's help centre says projects are available to all users, with free accounts limited to five. Because each project is separate, one project per client is the natural setup for a freelancer. You can upload source material to the project's knowledge and write the voice rules into its instructions.

Rule of thumb: one client, use Custom Instructions. Several clients, use one ChatGPT Project, Claude Project or Gemini Gem per client so voices never mix.

Step 2: Write rules the model can follow

"Friendly, confident, approachable" tells a model almost nothing. Every brand thinks it is all three. Instead, turn each trait into behaviour you could check on the page.

Weak:  Our tone is warm and confident.
Better: Write in short sentences, mostly under 15 words.
        Use "we" for the company and "you" for the reader.
        State facts plainly; never use "we're passionate about".
        One exclamation mark per piece at most.

Anthropic's prompting guidance makes two points worth borrowing. First, tell the model what to do rather than only what not to do. Second, give the reason behind a rule, because the model can generalise from it. "Don't say 'artisan'" is a rule. "Don't say 'artisan': every competitor does, and the client sees it as a cliché that makes them sound like a supermarket own brand" is a rule the model can apply to words you didn't list.

Step 3: Show, don't describe

Descriptions of a voice are useful. Examples of it are far more useful. Anthropic describes examples as one of the most reliable ways to steer Claude's output format, tone and structure, and suggests including three to five. It also recommends making them relevant to the real task, varied enough that the model doesn't copy one pattern, and clearly marked off from the instructions (it suggests XML-style tags such as <example>).

So pick three to five pieces of the client's best real copy, across different channels: a social post, a product page, an email. Avoid five near-identical Instagram captions, or every output will read like an Instagram caption.

<examples>
<example channel="email"> ...client's real newsletter intro... </example>
<example channel="web"> ...client's real product description... </example>
<example channel="social"> ...client's real post... </example>
</examples>

A before/after pair is even stronger. Show a generic draft, then the same message rewritten on-voice, with a line on what changed. The model learns the difference, not just the destination.

Step 4: Give it the facts, and forbid the rest

A big source of off-brand copy is invented detail: a "family-run since 1987" line that isn't true, or a sustainability claim the client can't back up. Put a short list of approved facts in the instructions or the knowledge files and tell the model to use only those. Add an explicit instruction to flag anything else with a placeholder such as [TO CONFIRM] rather than making it up. This also protects you if the client works in a regulated sector.

Step 5: A ready-to-adapt template

You write for [CLIENT], a [what they do] for [main audience].

VOICE (in priority order)
1. [Trait]: [what it looks like on the page]. Not [the overdone version].
2. [Trait]: ...
3. [Trait]: ...

ALWAYS: British spelling. Sentence-case headings. [House rules.]
NEVER USE: [banned words], because [reason].
AVOID CATEGORY CLICHÉS: [list].

FACTS YOU MAY USE: [short list]. If a fact isn't here, write [TO CONFIRM].

EXAMPLES OF THE VOICE:
<examples> ... </examples>

Before you reply, check the draft against the NEVER list and fix it.

For ChatGPT Custom Instructions, compress this to fit the character limit and keep the examples for a Project or the chat itself. For a Claude Project, put the rules in the project instructions and the longer examples and source material in the project knowledge.

Step 6: Test it like you'd test a junior

Don't judge the setup on one output. Give it three real briefs from different channels and read the results side by side. Ask: would the client recognise this as theirs? Did any banned word sneak in? Did it invent a fact? Each time something fails, fix the instruction that should have prevented it, not just the draft. After two or three rounds, most setups settle down.

Keep a short pass/fail checklist to hand for every draft: banned words, invented facts, sentence length, sign-off. It takes 30 seconds and catches most drift before the client sees it.

The short version

• Put each client's voice in its own space: Custom Instructions for one client, a Project per client for several.
• Write rules as observable behaviour, with the reason attached.
• Include three to five varied, real examples, clearly tagged.
• Give an approved fact list and a placeholder for everything else.
• Test on real briefs, fix the instructions, and use a quick checklist.

It's an afternoon's work per client to do properly. Once it's done, every draft starts closer to the finished thing.