Post

Developing products with AI can make the UI look inconsistent. To fix that, we changed the way they build and use design systems. We turned those changes into a framework that enforces rules on the AI agent when it checks and chooses from the design system. Here are three of those changes: 🧵

1. Separate product building from design system maintenance so the AI agent can focus on one task at a time. When the product team builds screens, it logs the design system limitations it runs into, and passes it over to the design team for them to improve the system.

2. Turn design intent into component contracts. The code of a component doesn't explain how it's supposed to be used. When an agent doesn't have exact answers, it makes assumptions. To reduce those moments, we added component contracts and descriptions to help agents make the right decisions.

3. Move deterministic checks out of AI skills. AI is surprisingly good at fuzzy, judgment-based tasks and surprisingly bad at deterministic ones. To fix this, we wrote scripts for everything that can be described algorithmically and simply tell the agent which script to call and when.