Gideon AwolesiWhat your agent needs to find
A developer who joins your team learns the system by asking questions and reading other people's code. An agent learns it from whatever's in the repository when you give it the task. So everything it needs has to be written down and easy to find: the components, what each one is for, the tokens, and the rules about when to use what.
Write docs the agent can actually use
Documentation that lives on a separate website, or in screenshots and videos, doesn't help an agent much. Put a short text file next to each component that says what it's for, when not to use it, and shows an example. Those same files help the people on your team too.
Give it instructions
Coding agents read an instruction file in the repository before they start, like CLAUDE.md or AGENTS.md. Use it to tell the agent where the components and tokens are, and what to do when nothing fits. "Stop and ask" is the most important line, because without it, the agent fills the gap itself.
Connect the design to the code
If your agent reads designs from Figma, Code Connect is what tells it which code component each Figma component maps to. At Syndigo, I wrote the Code Connect and component files for all 200 components, so a design and the code that ships it point at each other.
Test it like you'd test any user
Ask the agent to build three real screens from your product, and then review what it wrote. Count the components it reused, the ones it wrote from scratch, and the values it hard-coded. Every new component and every hard-coded value points to a gap in what the agent could find. Fix the gap, and run the test again.