Guide
ChatGPT custom instructions for your brand voice
What works, what does nothing, and where to put it in each tool.
Almost everyone has tried it. You open the custom instructions, type something like "write in a friendly, professional, conversational tone", and what comes back is friendly, professional, conversational and completely interchangeable with your nearest competitor. The instruction did nothing, and it would have done nothing in Claude, Copilot or Gemini either. It is worth knowing why.
Why it fails
You have agreed with the default
Friendly, professional and conversational are impressions, not behaviours. ChatGPT, Claude, Copilot and Gemini all believe they are already being friendly and professional, and they are right. Nothing in that instruction can be failed, so nothing in the output changes. You have not constrained the model. You have described its default.
The same is true of "authoritative but warm", "expert and approachable" and "human, never salesy". Ask two models to write in that voice and you get two different drafts, because there is nothing underneath the adjectives to make them agree. Ask one model twice and you get two drafts too.
An instruction only bites when the model can tell whether it has complied.
What bites
Checkable, numbered, and yours
Does nothing
"Write in a friendly, professional, conversational tone that reflects our brand."
Four adjectives, no test. Every model already thinks it is doing this.
Changes the output
"First person singular, never we. Median sentence twelve words, longest forty. Every claim of benefit carries a figure in the same sentence. Never use: solutions, journey, seamless, passionate."
Four rules, all checkable, all specific to one business.
Numbers bite. Adjectives do not. The second column is what a working instruction looks like, and every line in it comes from evidence rather than preference. The sentence rule comes from counting how your published writing actually lands. The banned list comes from the words you reach for that everybody in your sector also reaches for. The pricing rule comes from noticing that your best-performing pages put the number in the prose.
That is the reason the instruction has to be built from what you have already published rather than from what you would like to sound like. A model can follow "median sentence eleven words, longest fifty" all day. It cannot follow "sound like us", because you have not told it what that means and neither has anybody else.
A skeleton
Eight lines you can fill in yourself
If you want to build one by hand before buying anything, these are the eight lines that do most of the work. Fill each with a number or a list. Never an adjective.
- Pronoun policy: I, we or you, and the rate. Say which is banned.
- Median sentence length, longest allowed, and the share under five words.
- The punctuation you use and the punctuation you never use. Long dashes, semicolons, exclamation marks, bullet lists in prose.
- Twelve words and phrases never to use. Take them from the writing you are trying not to sound like.
- Twelve words and phrases that are yours, quoted from your own published work.
- Where numbers go: in the sentence, with the claim they support.
- Three positions the business takes that a competitor would argue with, in one sentence each.
- Three rulings: how to say no, how to price, how to deliver bad news.
That takes an afternoon with a spreadsheet and your own website. The free test will give you the sentence numbers and the borrowed vocabulary in a minute; the positions and the rulings need a person who knows what the business thinks.
Where it goes
Each tool takes instruction in a different place
- ChatGPT. Settings, then Personalisation, then Custom instructions, for everything you write; or a Project with instructions, if the voice belongs to one client or one publication.
- Claude. A Project with custom instructions, or a Style. A Project is the better fit for a full specification because it holds the worked examples too.
- Copilot. There is no persistent slot. The specification goes at the top of the document or the prompt, every time, so keep it short enough to paste. Every time.
- Gemini. A Gem, with the instructions in the Gem itself.
The wording that works in one is not automatically the wording that works in the next, but the rules underneath are the same. Keep the specification as a document you own, and paste from it, rather than as a setting inside somebody else's product.
The honest caveat
No prompt holds forever
Models get updated. People paste over the instructions. Projects get forgotten. A colleague opens a fresh chat and forgets the project. And the average of the internet exerts a steady pull back towards the middle, because the middle is where every model starts from.
So check. Paste a recent draft into the free test every month or so and see whether the numbers have moved. If landing has dropped to 5% and three of the banned words are back, the instruction has been diluted somewhere, and the fix is to paste it in again rather than to argue with the output.
One thing no instruction can do: tell the model what you believe. It can be made to sound like you. It cannot be made to think like you, and the positions section is the one line of the eight that only a person can write.
Give a model an adjective and it agrees with you. Give it a number and it obeys.
Get the numbers for the eight lines
The free test measures what you have published and gives you the sentence rhythm, the landing rate and the borrowed vocabulary, which is five of the eight lines above. A Wordprint writes all eight, with the prompt pack laid out for each tool.
Written by Ian, with Claude assisting, and scored by the smarticulate method at 94.3 out of 100 before it went up. Ian spent twenty-odd years turning thinking, his own and other people's, into something a board would act on, and built smarticulate to put numbers on it. His practice is ConsultingPartner.io: enterprise experience, right-sized. Who is he? →
More guides: how to write brand voice guidelines, tone of voice examples, how to measure brand voice.