What Claude Code and Figma MCP actually do in my week
The Tuesday-afternoon version of AI in design, and the parts I don’t hand over.
There’s a version of the AI-in-design conversation that is all demos and no Tuesday afternoon. Here is the Tuesday afternoon version.
Research synthesis
After a round of interviews I hand the transcripts to Claude Code along with my own coding frame and ask it to tag, cluster and pull quotes. I still read every transcript. What I get back is a first pass I can argue with, which is faster than a first pass I have to build from a blank board.
Specs and test data
Most handoff friction isn’t about pixels. It’s about the edge cases nobody wrote down. I draft specs with Claude, then generate realistic test data for the states that matter: the account with four thousand transactions, the prescription with three refills pending, the subscriber who paused twice. Those states find layout problems before engineering does.
Figma to code
With Figma MCP, Claude can read the actual frames and tokens and produce HTML and CSS that uses them. I’ve written CSS by hand for twelve years, so I can tell when it’s cutting corners, and I fix it. The point isn’t that the code is finished. It’s that the gap between prototype and implementation is now an afternoon rather than a sprint.
What I don’t hand over
Which problem we’re solving. What the research question is. What we trade off when two good options conflict. And the writing people read, including this. Judgment is the job. The tools give it more room.
Keep reading
More from the same notebook.
Which workflow is costing you the most right now?
Onboarding that leaks customers. An operations queue that still runs on phone calls. A design team that needs a lead. Tell me which one, and I’ll bring design, AI and twelve years of taking flows like that end to end.
Thirty minutes to start. You describe the product and what is slowing it down. I tell you honestly whether I’m the right person for it, and if not, who might be.