How to give any AI tool your brand (so the output stops being generic)

Short answer: AI tools default to generic because you give them nothing to work from. The fix is to hand any AI tool your brand as context — a single structured brand file with your identity, voice, and rules — attach it at the start of every task, and check the output before it ships. Give AI your brand and the first draft already sounds like you.

Every AI tool you use, Claude, ChatGPT, Cursor, Figma AI, starts from the same place: the average of everything it has been trained on. When you give it a task with no brand context, it makes something competent and generic. It cannot do anything else. It does not know who you are.

The fix is not a longer prompt. It is giving the tool your brand, as a proper brief, the same way you would brief a freelancer who has never worked with you before. Do that once and the output changes completely.

Why do AI tools default to generic?

An AI produces the most likely response to what you give it. No brand context means the most likely response is the one that fits the widest possible audience, which is the one that fits no one in particular. That is not a flaw in the tools. It is what happens when you send any intelligent collaborator, human or AI, into a job without a brief.

Most teams hand the AI a task and skip the brand. Then they spend the next twenty minutes editing the output back toward their voice, their words, their rules. The editing is not the value-add. It is the tax on an incomplete brief.

The brand-as-a-file approach

The most practical way to fix this is to keep your brand in a single file an AI can read, and attach it every time you start a job. Not a PDF. Not a set of screenshots. A plain text or Markdown file that contains the actual information the tool needs: the voice, the rules, the vocabulary, the things you do and the things you never do.

When the brand is a readable file, the workflow becomes consistent. You paste it in. You ask for the work. The output is already pulling from your brand rather than from the average, so the first draft is close, not a blank canvas you have to rebuild from scratch.

Step 1: build the brand file

You have two routes.

By hand. Write a Markdown document that covers your brand in plain, direct terms. Include: the voice (not adjectives, actual rules, see the post on AI brand guidelines), the colour and type values in their exact codes, the key messages and their wording, the words you use and the ones you avoid, and a short description of what the brand is for and who it is talking to. This takes a few hours, and it is worth every minute. The risk is that it rots quietly as the brand evolves and no one updates the file.

From Brand Atlas. If your brand is in Brand Atlas, ATLAS.md already exists. It is your whole brand exported as a single Markdown file, structured for exactly this use. When the brand changes, the file regenerates, so you are never pasting a stale brief.

Either way, the goal is the same: one file, complete, readable by any tool.

Step 2: attach it at the start, every time

The file does nothing sitting in a folder. It has to go in with the task.

For Claude: paste it at the start of a new conversation, or store it in a Claude Project so every chat in that project already has the brand in context. You ask once and every conversation in the project inherits it.

For ChatGPT: paste it at the top of the conversation, or keep a Custom GPT that has the file baked in.

For Cursor or other code-adjacent tools: add the file to the project context or reference it in your system prompt. The brand travels into the build, not just the chat.

For Figma AI and similar design tools: paste the relevant sections, voice rules, colour values, type choices, into the prompt before you ask for anything.

The rule is the same across all of them: brand first, task second.

Step 3: ask for the work normally

Once the brand is in context, you do not need to over-engineer the prompt. Ask for the work the way you would ask a briefed collaborator: write three captions for this launch, build a pricing page for this audience, draft a welcome email in our voice. Because the brand is already there, what comes back reflects it.

This is where the time saving appears. Not in the prompt, but in the editing. You are not correcting voice or fixing colour or rewriting off-brand phrasing. You are reviewing something that is already close and making it better.

Step 4: check the output before it ships

A good brief reduces drift. It does not eliminate it. AI tools follow instructions well and occasionally follow them loosely. Before anything goes out, check the output against the brand. Is the voice right? Are the rules followed? Is anything in there that should not be?

If your brand is in Brand Atlas, Oswald (the on-brand check in Guardian) does this automatically. You check the work, not from memory, but against the brand as it is actually documented.

What this is not

This is not about restricting what AI can do. It is the opposite. A well-briefed tool has more to work with, not less. It can be more specific, more distinctive, more useful, because it knows what it is making something for. Generic output is the result of a missing brief. A complete brief is what makes the tool genuinely useful.

This applies to every hand the brand passes through. The Claude-specific version of this workflow is here if that is your primary tool. The underlying principle is the same regardless of which tool you use.

Your brand is worth protecting, including when the tool doing the work is an AI. Give it the brief, every time, and it will stop guessing. See how it works.

Frequently asked questions