All startups are an experiment, and experiments begin with a hypothesis. Our hypothesis with Hermae: combining the power of design systems with AI could help teams ship products faster.
At the time, this idea was novel.
Here is what we learned from the experiment:
First, the bet was right. Today, Vercel v0, shadcn/ui, Cursor, and Figma MCP are examples of tightly-coupled design systems with code generation. None of this was mainstream when Hermae started.
The product was not. The root of the question was this: how can we provide LLMs with the context of a design system? Hermae's approach was to convert Storybook docs to Markdown, then use that to create the vector embeddings for RAG.
While a good approach, today's best design system context solutions take one of these approaches:
1. Local context within the IDE
Allow a tool like Cursor to access the design system by exposing a large set of usage examples and/or exposing the source code of the design system. This way the vector embeddings can be created, and the AI within the IDE will have the context it needs to generate production frontend code.
2. Public examples, public context
Use a public design system library like shadcn where there is an abundant amount of public code that uses the design system, which guarantees its usage examples are baked into the training data of the underlying LLM. Generation accuracy may not be 100% but can be supplemented with AI-native documentation.
3. Design tool MCP
Connecting to Figma MCP allows an agent to "see" your designs by pulling out screenshots and token values. This won't generate production-ready code without other code examples, but it does a great job at pulling out values that you'll need to match the designs.
The best solutions don't require a third-party tool — they're integrated into the tools that we use everyday to make shipping as fast as possible.
Combining some or all of these approaches is best, especially combined with advancing models and self-verifying systems.
If you haven't tried one of the tools mentioned here, I highly recommend doing so.
Thanks again for joining on the journey.