User (founder of Magic Patterns) highlights the challenge of AI models not being able to effectively ingest and output custom component libraries (e.g., from Storybook, Figma, GitHub) due to their highly custom nature and lack of presence in training data. This implies a need for this capability to be improved for AI design tools.
A guide to creating AI prototypes that match your brand from Lenny Rachitsky, featuring Magic Patterns and other amazing toolsπππ In a past life, my co-founder Teddy Ni and I worked very closely with design systems teams at big tech companies, so we have always approached building Magic Patterns from the perspective of "how can I use my existing styles?" Even when we first launched in 2023, we viewed ingesting your component library as a core feature. We connected to Storybook / Figma / GitHub. (Shoutout to our early customers, you know who you are!) But we learned this is a *very challenging* problem because component libraries are highly custom. LLMs don't know how to output your custom library. It's not in the training data. So were a bit too early: the models weren't good enough and it didn't scale. We also didn't understand the problem. We thought engineers would use the code created by Magic Patterns. Instead, we learned noticed it was PMs and designers. π Fast forward 2 years, it is super cool to see Magic Patterns in a great write-up from Colin Matthews on how to leverage your existing styles in our AI prototyping tool. It's a problem we care deeply about. There is so much to build!