Drink me, eat me, improve me

Designer Justyna Tarkowska's design system team makes AI agents log each corrected mistake and turn it into a build check, and reports that the mistakes now recur less.

Justyna Tarkowska led a team's move from a Figma component library to a design system kept in one GitHub repository, with tokens, coded components, contracts and documentation. AI agents document, audit and port components through that system, and they kept making confident mistakes even when they had the context. Tarkowska describes the protocol the team built so that each correction leaves a lasting fix.

• A person, a failing automated check or the agent itself can catch a mistake. The agent then fixes it, writes a lesson in a file next to its instructions, turns the lesson into an enforcement and reads its lessons before every run.
• The team ranks enforcements: an automated check that fails the build first, a step in the agent's instructions second, and a written rule only as a last resort.
• A rule about disabled-state colour tokens kept failing as a step in the instructions and stopped only when it became a check that fails the build.
• A register of deliberate exceptions, each with its reason, stops agents from reporting intended design choices, such as a smaller icon in the icon button, as bugs.

A design system team can treat every correction of an AI agent as a change to the system, reviewed through a pull request, instead of a one-off fix. Tarkowska's evidence is a few months of use and a couple of dozen lessons, which Tarkowska calls one story and not a study.

Derived by Working Surface from the article; more in the Playbook. Source line: You can fix mistakes, but if the fix is forgotten between runs, the mistake comes back tomorrow, just as confident.

22 October 2026: as AI agents make building cheap, teams are moving effort to deciding what to build and writing down the reasons where people and AI agents can reuse them.