Before you hand your design system to an agent. A short playbook for writing down the judgment that never made it into the docs.

Go pattern by pattern. Pick what you rebuild constantly — headers, forms, empty states — and force every implicit rule into words: title limits, how many actions allowed, primary versus secondary.

Write down the why, not just the rule. That's the part teams skip, and it's what lets an agent generalize past the cases you thought of.

Keep this knowledge separate from whatever surfaces it. A skill or prompt should point to the rules, not carry them — otherwise it can't grow without bloating every request.

Then measure it. Run the same task with the knowledge layer off, then on, and check what fraction of components used actually exist in your library. One team watched that number move from 7% to 96%.

Every time an agent still gets something wrong after that, treat it as a missing rule, not a bug — and write it back in.