AI is useful when it accelerates a clear process. It's noisy when it replaces product thinking. I use it every week in Figma and Cursor. The days that go well share the same shape. The days that don't usually skipped step one.
A delivery workflow that holds up
- Define goals and constraints. Who is this for, what success looks like, what we refuse to break.
- Build or extend a reusable design system. Tokens and patterns before one-off screens.
- Prototype key interactions. Focus on the risky flows, not every screen at full fidelity.
- Ship in small, testable increments. Vertical slices beat "design everything, then build everything."
- Iterate with real usage. Support tickets, session recordings, beta feedback.
That loop is what I run on products like Hermeneia (closed beta across languages and insight modes) and Spontrec (Flutter recording flow that has to feel instant). AI drafts variants and glue code. I still own whether the flow earns trust.
How I use AI without lowering the bar
- Drafts for copy options and layout alternatives, then I cut hard
- Scaffolding for components that already have tokens and rules
- Refactors and exploratory PRs, reviewed against accessibility and real content
- Never "ship whatever the model produced" as the definition of done
Senior design work in an AI-heavy stack is taste, prioritization, and knowing when to stop generating and start deciding.
One check before you open the tool
Write the goal and the constraint in one sentence. If you can't, the model will happily generate confident junk. If you can, AI becomes a multiplier for craft, consistency, and speed instead of a substitute for them.