Your team probably already uses AI. Someone asks Claude or ChatGPT for a first draft of a deck, a developer has Cursor build a component, marketing creates images in Canva. The result is often usable, but rarely recognisable as your brand. That is usually not the model’s fault. It comes down to what the model knows about your brand.
An agent only knows what you give it
An AI agent does not see your brand the way you do. It does not know your logo, does not know why your headings never run longer than two lines, and has no idea that light green is reserved for the next step. A PDF brand book helps a little, but it is written for people who read between the lines. Agents need explicit, well-organised information.
What an agent needs
Brand files and intent
A short description of character, audience and design principles, in plain language and as usable rules. Not “we are accessible”, but “technical terms are explained and every page opens with the conclusion”.
Tokens
Colours, typography, spacing and motion as named roles, preferably in the DTCG format. Then the agent does not have to guess which colour is meant for a button.
Components
For digital work: your real components in code, with properties and examples. A coding agent that uses your own Button component produces fewer deviations than one that invents its own. Figma provides design context to agents through MCP, and Code Connect links that context to your real components. v0 can build with a registry of your own design system.
Recipes
Task-based instructions for recurring work, such as a service page, a LinkedIn post or an advice slide. A recipe describes the goal, the content needed, the right template, text limits and the checks afterwards. That way nobody has to read the whole system before they can start.
Good and bad examples
Examples show how rules come together. Counter-examples are at least as valuable, as long as they explain what goes wrong. “Too busy” helps nobody. “The hierarchy disappears because there are three headings of the same size” does.
Instructions where the agent looks
Coding agents read files in your repository. An AGENTS.md or similar instruction file tells them where the tokens live, which components exist and which recipes apply. For a Claude project, a ChatGPT project, Cursor rules or a Canva Brand Kit, the same principle applies in a different form.
What agents do not (yet) do well
To be honest: even with a well-prepared brand, the result depends on the tool, the version, the brief and the context you provide. A few limits to keep in mind:
- Do not have logos and small text generated. Always place them from approved source files.
- Composition and taste remain human work. An agent follows rules, but does not decide what deserves attention.
- Integrations and features change quickly. Some are still in preview or beta.
- Updating your system does not automatically rewrite existing material.
Testing with real briefs
The only way to know whether your brand works with an agent is to try it. Pick a few briefs your team actually handles and have them made with the tools you actually use. For example:
- A service page with an introduction, three sections and a contact action.
- A social post with a long title and an unusual image format.
- Five slides for a proposal, including one dense content slide.
- A form with an error message, a loading state and a confirmation.
For each brief, record which tool and version you used, what the agent received and which corrections were needed. Whatever rubs, you fix in the system: sharpen a rule, add a recipe, explain a counter-example. That way the system gets better with every round, for agents and for people.
A brand that is ready for agents is usually easier for people to use as well. The difference lies in how explicitly you record your decisions.
At undesign we set up brands so teams, agencies and AI tools can keep building with them. We do not promise that every model will always produce perfect work. We do deliver a system that has been tested with the tools and briefs that matter to your team.
