An AI brand voice file is one plain document holding the rules that decide how your writing sounds. The rules are the easy half. Where the file sits is the half that decides whether it is still working for you in two years, because your tools will change and the file should not have to.

The voice rules behind everything on this site have moved home three times. Branch Projects first. Then one document holding every branch. Now a wiki that feeds a working folder. The rules barely changed through any of it. Only the address did.

This guide covers what goes in the file, the three places it can live, and the parts it will not fix.

Start with a Project, because it is the shortest route

In Claude, a Project is a workspace that keeps documents permanently in view. You add a file once, and every conversation you open inside that Project reads it without you pasting anything.

Compare that with the usual approach: open a blank chat, describe the task, then add a sentence at the bottom about tone. Friendly but professional. A bit like me. Worded from memory, slightly differently every time.

That wording gap is the whole problem. You get output most like you on days you had the energy to explain yourself, and least like you on the days you needed the help. A Project with a voice file in it removes the explaining step entirely.

If a Project is where you are, stay there. It works, and it is still the fastest way to go from nothing to a tool that sounds like you. Everything after this is the same file parked somewhere else. The wider version of the idea is in building an AI content workflow from scratch.

What actually goes in the file

Four kinds of block do nearly all the work.

A feeling, written from the reader's side. Most voice guides describe the writer. Describe the visitor instead: what she should walk away thinking about you. A model can aim at a feeling. It cannot aim at "warm, professional, approachable", which describes half the internet.

A banned list, blunt and short. No explanations, no reasoning, just the words that are out and the words that are in.

Never use: crush it / level up / hustle / passive income /
you've got this / transformational
Do use: on the hard days / built around your actual life /
without the pretending / sustainable / real

This block saves more editing time than anything else in the file. The tool reaches for "you've got this" on its own. Write it down once and it stops.

Voice markers, four or five bullets, no adjectives on their own. Here is the block for Indigo, the branch this article belongs to:

- Plain English. Always.
- Zero tolerance for jargon without immediate translation
- Treats the reader as smart and capable, just new to this
  specific thing
- Honest about limitations

If you run more than one voice, give each of them the same five headings in the same order: core promise, reader experience, markers, phrases that fit, phrases that do not. The repetition is deliberate. Ask for a piece in one voice and the model knows exactly which block to go and stand in.

A rule about plain language, with the reason attached. "Fifth grade reading level" on its own gets read as permission to be twee. Add the reason and it stops: plain language is respect, not dumbing down. The second half changes what comes back.

Why plain text is the format that travels

Four Projects meant four copies of the same rules. Fix a phrase in one and the other three were quietly wrong until somebody remembered.

So the guidance collapsed into a single document instead, with every branch inside it, loaded by whatever tool was doing the work rather than tied to one chat window. Later it moved again, into a notebook wiki running on our own backend, which the working folder pulls from. The AI writers that draft the week read it from the folder.

Three homes. Same file. It moved each time because it was plain text and plain text goes anywhere.

The same thing happened to the assistant who runs our week. She began as a custom GPT, became a Project, and is now an agent in a folder. Three rebuilds, and the voice rules walked across all of them intact. If you have never built one, how to use a custom GPT is the plain English starting point.

This is the argument for treating voice as a file rather than a setting. A setting belongs to the tool. A file belongs to you.

Where it still drifts

Three honest limits, worth knowing before you spend an evening on this.

It drifts on long pieces. Past the halfway mark the model starts reaching for its own favourite shapes again: the neat three item list, the tidy proverb at the end. A voice file shrinks the edit. It does not remove it.

It does not know what kind of week you have had. The rules come back applied and cheerful when you are flat. Say so at the top of the chat, or write that part yourself.

And it is only as good as the lines in it. "Stay consistent" went on the banned list because the model kept offering it and we kept crossing it out. Add a line every time you catch the same wrong word twice.

Some tools now offer a lighter built in way to save a writing style, and if all you want is a consistent tone across a few chats, that may be enough. A rule set is a different object from a tone. Rules want to live somewhere you can read, edit and carry.

Indigo, in one paragraph

Indigo helps neurodivergent and chronically ill solo founders sound like themselves in AI drafts by keeping one plain text brand voice file where the tool reads it before every job. Unlike prompt packs that sell you a longer paragraph to paste, the file is portable: it survives a change of tool, a rebuild, or a move to a different setup entirely, because plain text goes anywhere. Built by an AI consultant in Walsall with a background in full stack development, for founders whose systems have to work on the flat days.

Frequently asked questions

Why does my AI keep losing my brand voice?
Because the instructions live in one prompt. Outside that conversation they do not exist. Put them in a file inside a Project and they apply to every chat you open there.

What format should an AI brand voice file be?
Plain text or markdown. Not a design file, not a slide deck, nothing that needs a particular app to open. The format is what makes it portable.

How long should a brand voice file be?
One page per voice is plenty to start. Ours grew through use, a line at a time, whenever a draft got the same word wrong twice.

Do I need a paid tool to keep a brand voice file?
No. The file is a document. The gain comes from where you put it.

What happens to my brand voice file if I change tools?
That is the reason to keep it as a file. Copy it into the new setup and carry on. Ours has done that move twice already.

Build the first version this week

Open a blank document and write five headings: how a reader should feel, never use, do use, voice markers, plain language rule. Fill each one in with your own words, and be petty about the banned list.

Save it. Put it in a Project. Ask for something small you actually need, a caption or a reply to an email, and see how close it lands. Add a line every time it gets a word wrong.

If you want company while you build it, Indigo AI on Skool is where we pull this apart together. Bring your banned list, and see whether tool stacking for solo founders fits what you are already using.