AI workflow
I design AI features, and design with AI.
Process
Where AI fits, stage by stage
- 01
Understand
Claude
Synthesise interview notes, survey responses and support tickets into themes and open questions.
My partI check every theme against the source material and talk to users myself — AI summarises, it doesn't replace listening.
- 02
Explore
Claude Design · Figma Make
Generate several visual and structural directions quickly, so the team reacts to options instead of a blank page.
My partI choose and refine directions against the design system, brand and the problem we framed.
- 03
Prototype
Cursor · Claude Code · v0
Build working, coded prototypes with real states — loading, empty, error — in days rather than weeks.
My partPrototypes test behaviour with users before engineering commits; I own the interaction decisions.
- 04
Ship
Claude Code · Figma variables
Draft specs, edge-case lists and token-based components that map cleanly to code.
My partI review accessibility with real assistive tech and pair with engineers on the final build.
Designing AI products
Principles from shipping AI search
Ask for intent before results
At Talstrike, scouts describe the opportunity first. Good AI output starts with a clear brief, so the interface collects one.
Make flexibility explicit
Hard requirements decide eligibility; soft preferences change ranking. People can see — and change — how the system is weighing things.
Keep people responsible for the decision
AI surfaces candidates; the scout evaluates them. The design leads straight from results into the person's own judgement.
Measure relevance, not just speed
Reaching an unsuitable result quickly isn't success. I test no-match states, incomplete data and refinement, not only the happy path.
Guardrails
What I don't hand to AI
- Confidential client data and personal information never go into AI tools without the client's approval.
- Research conclusions are checked against the original notes, not just the summary.
- Accessibility is verified by hand and with assistive technology, not only by automated checks.
Building
I build with Claude
I use Claude and Claude Code to turn designs into working, coded prototypes: real states, real data and real interactions that teams can test before engineering commits.
It lets me hand engineers something closer to production, and spend more of my time on the decisions that need a designer.