My AI workflow, with the specifics left in.
Plenty of designers have AI opinions. This page is the actual setup: what runs where, what gets checked, and what never leaves my hands.
The pipeline
Figma to production, no handoff
My design system lives in Figma as tokens. Figma MCP reads them, and Claude Code builds against a CLAUDE.md file with my spacing scale, typography, and implementation rules. Change a token in Figma and the change shows up in code.
Synthesis at depth
Interview transcripts, analytics exports, survey answers: AI does the clustering, I do the interpretation. A model will happily sort twenty interviews into a confident wrong answer, so the reading of what matters stays my job.
Prototypes in code
An idea becomes a clickable prototype within hours, which moves testing earlier: real interactions get judged before anyone falls in love with a direction. Onboarding flows and 0 to 1 concepts benefit the most.
The verification bar
Every merge gets read
AI writes faster than I can review, which makes review the bottleneck. I accept that bottleneck on purpose. Nothing ships without me reading it, and the day that stops being true is the day the quality bar becomes fiction.
Hard states get designed first
Probabilistic output raises the stakes on the old discipline: empty, loading, error, and recovery designed before the happy path gets its polish. In AI products, how it fails is most of how it feels.
Agent work is a skill
Multi-step agent sessions, MCP integrations, versioned changes. I do this daily, and it turns out to be a different craft than prompting a chatbot. Designing agentic products comes easier when you already work this way yourself.
Proof
This site is the working example. Designed in Figma, built in Astro through the exact pipeline above, every change reviewed before merge.