Building an LLM safe design system (polar.sh)

Polar is developing a new design system called Orbit, aimed at ensuring consistency in UI components generated with large language models (LLMs). The challenge addressed is that while LLMs can fluently write CSS or Tailwind classes, they lack awareness of specific design decisions, leading to a hodgepodge of variations in styles across components. Orbit aims to lock in design decisions by using named tokens that represent intent, rather than raw values like traditional CSS classes, reducing the risk of inconsistency. The significance of Orbit for the AI/ML community lies in its innovative approach to integrating LLMs into the design process while mitigating their potential pitfalls. By constraining styling options and enforcing these constraints through ESLint rules in continuous integration (CI), Orbit ensures that only compliant code is approved for use. This method not only enhances consistency but also simplifies the design decision-making process by allowing developers and AI to work from a shared vocabulary of design tokens, thus preventing common issues related to styling drift, like incorrect shades or assignments in various contexts.