Brand voice for AI: why it belongs in a file

What this makes

You answer questions about how your brand writes. At the end you get a folder of files. One of them is a short guide a person can read. The rest are the same rules in the exact formats that ChatGPT, Claude, Gemini, Copilot, Cursor and a style checker read on their own.

The answers are put together in your browser and stay there unless you ask for a share link. There is no account and no email gate. Page views are counted, which the notice at the foot of the page turns off.

If you already have a brand guide, none of this replaces it. What to do with the guide you already have covers the parts of a fifty-page document that an assistant can act on and the parts it cannot.

Why a file, and not a chat

Most people teach an assistant their voice by describing it, which works until the conversation ends. The next one starts cold. One person on the OpenAI forum described spending hours on it and finding that the next day it had forgotten almost everything. You cannot teach it. You can hand it the same short briefing every time and this makes the briefing.

There is a second reason. Attaching a file to a project does not make an assistant follow it. OpenAI's own model specification says file attachments have no authority by default and must be treated as information rather than as instructions, unless you delegate that authority in your own words. Your download includes the sentence that does it. It is one line. Without it, the file sits there.

The wider version of that problem has its own page: why an assistant ignores instructions it has been given sets out the published causes, and what happens when two of your rules contradict each other covers the case where the file fights itself.

How many questions, and why?

Free brand voice tools tend to ask one question and hand back three adjectives. Across forty-nine products surveyed, the median number of questions was between one and four and the most common number was one. Three adjectives are not enough to change what a model writes, because “warm” means something different to everybody. This tool starts with eleven questions, which is enough to write every file. Two optional rounds follow. The first asks where you write and which words you swap out, the second how long your sentences run and what you must never claim. That makes up to thirty-six questions and you can stop after any round.

The core questions do two things adjectives cannot. Four of them show you the same message written two ways and ask which one you would rather send. The paragraph you pick goes into your file as a worked example, and examples do more work than rules do: three to five of them, chosen for spread, roughly double how accurately a model reproduces a voice compared with instructions alone.

Three of the rest are mechanical. Do you say “we” or “I”. Short forms or full forms. British or American spelling. They look small. They change every sentence you will ever get back and they are the cheapest questions in the set. Almost nobody writes them down.

What a finished document should contain, field by field, is set out in the long version of what a tone of voice document holds, including which fields are rules and which are decoration.

What it will not do

It will not hide the fact you used AI. Nothing does. Style tests still identify machine writing after it has been rewritten, translated and cut down, at around ninety-seven per cent accuracy. An explicit ban on long dashes still leaves one model using nearly four of them per thousand words. This makes the writing sound like you instead of like the default. Anybody promising the other thing is selling you something that does not exist.

It is also a first draft. A good one, and better than most brands have ever written down. It is not the same as a week with a writer who reads your work and argues with you about it. Your file says so on its face and it lists what you left unsettled rather than guessing.

The catalogue of things people treat as proof of machine writing is worth reading before you rely on any of them: what the tells say, and how fast they expire goes through them with their sources.

Why the file is short

Long files get ignored. Anthropic documents that bloated instruction files cause the model to skip the instructions that matter and the compliance research is blunter still: as verifiable constraints stack from one to twenty, the rate at which a model follows all of them falls from about ninety-six per cent to as low as twenty.

So your file caps at fifteen binding rules. Anything past that is kept, marked as non-binding and written into the long version only. More captured detail past that point makes the result measurably worse, which is why the cap is in the software rather than in a paragraph of advice.

The published limits behind that cap, per tool, are collected in how long an instruction file can be before it stops working.

The lengths

Every place you might paste this has a different limit and they are real, published limits. Canva's brand voice box holds five hundred characters. ChatGPT's custom instructions hold one thousand five hundred on the free tier and five thousand on a paid one. Claude's organisation instructions hold three thousand. Perplexity's project context and Microsoft 365 declarative agents hold eight thousand.

You get a version rendered to each of those, and none of them is a truncation of a longer one. The five hundred character version drops the tone description entirely and keeps the mechanics, because mechanics are checkable and change every sentence while a tone adjective costs a quarter of the budget and changes very little. No other tool surveyed renders to any of these caps.

How to know it worked

The most common question about a file like this is whether it is doing anything and the failure is silent by design. So the download includes a ninety-second check: a short prompt, an answer written without your file and the same prompt to run again once the file is installed. Put them side by side and count how many of your own rules each one breaks.

That count describes the text. It does not judge it. There is no score, no gauge and no percentage, because a brand-fit number out of ten would be the most shareable and least defensible thing here. The same counts are computed by the style checker in your download, so you can re-run them yourself and get the same answer.