I moved my design practice onto Claude Code and off Figma. It was a slow shift, lots of trial and error working out which parts of the work AI could take and which needed to stay with me. I struggled to trust it at first, and it only got there as I built the systems around it and matured how I use it. The speed came without giving up the decisions.
New leadership arrived in early 2026 with a push to use AI everywhere. There was a lot of pressure to show off new findings, be more efficient or faster or better than without AI. The most motivating part of this was actually understanding what AI could and couldn't do. I wasn't going to use it if I didn't find legitimate improvements either in time, quality, or tedium. With a UX team of two, every hour saved mattered, so we leaned into Claude Code on two fronts, research and prototyping.
Written under a non-disclosure agreement, so the company and its product aren't named. “The product” here is its cloud case-management software for law firms, a B2B SaaS tool, and “the company” is the legal-tech company that makes it. The work, decisions and figures are as they happened.
Research was the easy start, since I had already built a manual version of the repository and trusting AI with the slow parts of analysis was low risk. Prototyping was where I struggled, and mostly because I thought too small. I used AI for isolated tasks and kept hitting rework, where a prototype came out roughly 50 percent right and almost impossible to edit, or an Outlook proof of concept came out uncanny. It started working when I stopped treating it as a shortcut on single tasks and gave it a system to work inside. That is when the rework dropped and I committed.
The systems, not the prompts
What fixed it was a system. Claude Code prototypes come out with much more consistency and accuracy when they're built from a bona fide design system, so I built one. It's a real design system in Claude-native HTML and CSS, built on Tailwind, and I owned all of it, the components, patterns, and templates. I kept it separate from the product's legacy code, so designers got control without touching production, and a component map linked each piece to its developer twin so engineers rebuilt with little rework. Once the frame is a system, the tooling only has to reason about the content, which is where the accuracy and the speed came from.
The research side ran on its own system. Our repository followed an atomic model, turning session transcripts into observations and then into insights scored by how many users raise an issue and how severe it is. Nothing that mattered was committed without a person reviewing the proposal.
Fig 1 — The AI systems I work in: an atomic research repo, a design system tuned for AI, and prototyping through to handoff. Each is a small suite of skills with a human gate at the decisions that matter.
What changed
Off Figma since April 2026.
Complex prototypes went from about a week to about a day, and edits from weeks to an afternoon. These are my estimates, not instrumented figures.
Research got deeper. By the time I left we'd logged over 200 interviews connected across projects, and I would put the quality of our validation sessions up by about half.
The design system was in real use when I left. Developers built from it, and there was buy-in to base the product's refactor on it. It was still maturing, so the biggest gains were still ahead.
It shipped the Outlook rebuild, which ran fully through this workflow.
What I found is that AI gets more capable and more stable the more I build the systems around it, and I got there through a series of iterations rather than one call. Speed lets me explore more and move faster, but it doesn't make the decisions any easier for me, and it's not much use if I can't do the work without it. Knowing when to reach for it and when to leave it is the part that took the longest to learn.
Rebuilt a legal case-management product's Outlook add-in from the ground up. The legacy tool was unreliable and full of gaps, and firms logged the pain loudly in feedback portals. I led the research and owned the design of a replacement to win back their trust.