Tom Johnson outlines a nine-step AI-heavy design workflow where he starts with messy voice transcripts, uses Claude and tools like Willow, Notion, or Granola to structure the problem, then lets AI generate a deliberately bad but functional app as a scaffold. This matters because it reframes AI’s weakness at UX as a feature: a cheap way to explore directions, expose edge cases, and pressure-test scope before committing to real craft in Figma and a proper engineering handoff.

Speaking of Tom, in this interview with Jay, he shows the DM he sent to Linear’s CEO to get an interview — and later the portfolio and case studies he used to land the job at Vercel.

Tom Johnson: ​“It’s literally faster for me to build a concept inside of the actual codebase than it is to work in Figma. But the amount of versions and breadth of the final result is not up to the quality bar that I usually hold. I’m loving the speed, but the output is sloppy at times. So I’ve got this weird flow of Code → Code → Figma → Code → Figma. Repeat, reorder, etc. The issue is that the transition from Figma → code with their MCP is solid. […] But code → Figma… is a terrible flow.”

I often spend more time recreating a particular screen in Figma than actually designing the change. I’ve been beating on the Code Connect drum for a while as it has huge potential for connecting design and code components. Even though it was built to translate components from Figma to code, I don’t see why it wouldn’t work in reverse. Code Connect was announced almost two years ago (way before LLMs and MCPs got powerful), but because it’s locked behind Organization and Enterprise plans, it’s rarely discussed in the community or considered by plugin and tool makers.