Skip to main content
James Madson

AI Practice

I design and build with AI every day. I prototype in code, not static screens, so what I hand an engineer already runs.

Good design still starts with the problem. AI changed how fast I get from there to something real. It didn't change what makes the result worth building.

AI in practice

Cursor for code, Claude for thinking through design problems, Figma Make for quick concepts. The honest answer is that it changed how fast I move. A design system that used to take weeks takes a day. This site would have taken a month and it took a week.

I prototype in code instead of static screens, so the thing I hand an engineer already runs. That's the part I won't give up. Deciding what's worth building, and whether it actually holds up, is still mine.

The part I haven't worked out is the cost. These models take real energy and water to run, and that gets skipped over in most conversations about how fast we can all go now. I notice the gap. I've spent years designing things meant to get people outside, including a product built to measure what nature is worth to human health, and now I use a tool with a footprint I can't see and can't fully account for.

I don't have a clean answer, and I'd rather say that than pretend it's settled. What I do is use it deliberately instead of idly: think first, then run, rather than spraying prompts at a problem I haven't sat with yet. That's a small correction and I know it. I'm still working out what a responsible amount looks like, and what I owe in the other direction, in the work I take and the things I build.

I think about the people entering this field, too. I learned by doing the tedious parts badly for years. I'm not sure what replaces that.

What I have built

  • Internal AI tools at Pfizer

    Designing and building internal software that helps product teams move faster. I take the lead on prototyping and concepting new tools, moving between design and production code with Claude, Cursor, Figma, and Figma Make, alongside custom tooling I am helping build in house. That includes agent-driven interaction patterns and design system components shipped as code, not just screens.

  • Juniper design testing, NatureQuant

    Automating design quality checks against Juniper, the design system behind NatureDose, so drift gets caught by the system instead of by whoever happens to notice.

  • Canon

    A curated gallery of design skills and design skill kits: reusable instructions that give an AI the context to do real design work, rather than starting from nothing every session.

  • This site

    Designed and built with Claude Code as a working partner. I made the design calls and reviewed every change; it did the heavy lifting on SCSS architecture, cross-document view transitions, and debugging. Running accessibility skills against it caught a real contrast failure on the homepage, at 3.18:1, that I had walked past for weeks. The colophon documents how the whole thing is put together.

  • A brand skill

    A prompt-based brand skill teams can add to their AI of choice, so people can write and design on-brand for prototypes, decks, and scripts without reinterpreting the guidelines every time.