My (evolving) AI UX Approach

My process and principles for designing trustworthy AI experiences.

Image

My AI-augmented process

AI compresses the pass. The judgment, the taste, and the read on what the user actually needs — that part's still mine.

Understand
Is this actually the right problem to solve? I pressure-test the brief, surface assumptions, dig into research, study what competitors get right and wrong, so the problem is sharper than how it arrived.
–Receive PRD
–Pressure-test the brief
–Synthesize research
–Competitive analysis
–Define problem, goals, success metrics, hypotheses
Map
I map the territory — JTBD, user stories, the workflows that connect them, the blockflows underneath — to decide what the product needs to do, in what order, and what good looks like when we ship it.
–JTBD, user stories, tasks
–Workflows
–Blockflows
–Design principles
–Product requirements / priorities
Iterate
This is where the work becomes visible: low-fi, hi-fi mocks, and prototypes. The goal isn't the first idea, it's exploring enough directions that the right one becomes obvious.
–Low-fidelity wireframes
–Hi-fi mocks against the design system
–Prototypes & motion
–Critiques & iterate
–Usability testing
Build
Designs aren't done when the file is pretty — they're done when they're shipped and working. I write specs, hand off to engineering, and increasingly produce real product code myself.
–Spec & engineering handoff
–Ship-ready design code
–QA against mocks
–Launch
–Measure & learn
AI & Tools
Claude Custom Research Agents
AI & Tools
Claude FigJam
AI & Tools
Figma / Figma AI Claude
AI & Tools
Cursor + Figma MCP Claude Code
Where AI actually helps
Surfacing gaps, edge cases, and assumptions in the PRD before a pixel is drawn. Generate prioritized hypotheses.

Synthesizing research notes into themes.
Where AI actually helps
Helps craft first pass user stories, flows, and edge-case maps.

Drafting blockflows I refine rather than starting from a blank canvas.
Where AI actually helps
Exploring ten directions in the time it used to take for one.

Generating variants I react to and refine — often surfacing options I wouldn't have considered.
Where AI actually helps
Shipping designs as real code, not just images.

The handoff gap between design and engineering shrinks, and I stay closer to how the product actually behaves in users' hands.

Designing for Trustworthy AI

Lately, I have been focused on how best to design for AI. Thinking about it, I believe the friction people feel with AI isn't primarily a capability problem, it's a shared-understanding and alignment problem (and ultimately, a trust issue).

As humans and AI work together, an effective working relationship depends on knowing what's being handed over, why decisions are made, and how to correct course when something goes wrong. Most AI products haven't been designed around those mechanisms yet.

Below are some emerging principles for trustworthy AI that I am noodling on:

The AI needs to demonstrate understanding

Good delegation isn't built in a single handoff. You establish how you work, what's important, where the boundaries are, and what success looks like. Trust builds through demonstrated understanding.

Most AI products skip this step. They execute immediately and hope the output matches what you intended.

The AI needs to earn trust through transparency

People ask AI to triage email, book travel, or draft reports. Then they hover over every action and double-check every output.

That's not user resistance. It's appropriate caution for a relationship that hasn't earned trust.

The AI should evaluate how best to balance understanding and speed

AI products are still optimized for speed to output rather than speed to understanding. Demo culture rewards instant answers, while small cues like "thought for 5 seconds" reinforce the idea that faster means smarter.

But speed doesn't create the shared understanding required for delegation.