A walk through how I actually work, from research through a shipped, reviewed design, and exactly where AI fits along the way.
AI is genuinely useful for synthesis and speed. It doesn't decide what's right for the person on the other end of a workflow, that's still my job.
This is the principle behind everything on this page.
Four stages, the same ones I've used on regulated, high-stakes products for over a decade. AI shows up at specific points, never all of them.
Talk to the people who'll actually use this. Interviews, usability testing, support feedback, heuristic audits, understand what's actually breaking down before proposing anything.
Clustering findings across interviews faster, so patterns surface sooner. It doesn't decide what the pattern means, that's still mine to interpret.
Turn what I learned into a clear problem statement, often different from the original ask. This is where I push back if the real problem is bigger than the requested feature.
Sitemap, prototype, and design in Figma using our design system. Test early rather than waiting for a polished version.
Fast first-pass prototypes to react to sooner, and structured documentation drafted from a plain description, which I still review and correct.
Usability testing before shipping, then watching real adoption after launch. Stakeholder check-ins happen throughout this whole process, not just at the end.
The step most people don't see: how a finished design actually becomes a working feature, and where a human has to be the one who signs off.
Structure, real tokens, documented components, same as always.
Actual color tokens and component names, not a screenshot guess, plus links to design system documentation.
Claude Code, Codex, whichever a team has standardized on, generates a starting implementation grounded in that real data.
Every time. The AI's confidence is not the same thing as correctness, especially in a regulated workflow.
The best design disappears.
What remains is trust.