AI Practice

Building ideas, together with AI

AI is driving the cost of execution toward zero — an idea can become a working product in hours. But once execution stops being the bottleneck, judgment becomes the scarce thing: AI can generate endless options, yet can't tell which one is right.

This is exactly where I come in. Years of interaction design have given me a clear sense of what "good" means, letting me cut the mediocre and keep what's right out of everything AI produces — and refine it to a quality I'm satisfied with.

Tools keep changing; methods endure: breaking needs into clear tasks, building stable workflows for each stage, holding output to my own standards. That process from concept to launch is my real, stable capability.

And so the ideas that once stayed in my head, or stuck in a prototype, can now be built one by one. That's what AI has given me — and my answer to it.

How I work

Discover

My role: Tell which problems are worth solving.

Before building, I use AI to map a field fast — market, competitors, real pain points. Research shrinks from days to hours, but telling real problems from noise still takes a designer's eye.

Ideate

My role: Decide what to build, and why.

Diverge widely with AI, open a vague notion into multiple directions, gauge risk and feasibility, then converge on the one worth doing.

Define

My role: Turn a fuzzy idea into a clear spec.

With AI I produce two key documents: a product design doc that pins down features and boundaries, and a development plan that breaks the design into executable steps.

Design

My role: Own the quality of every screen.

Interfaces and interactions designed primarily in Figma, with Claude and Midjourney woven in to iterate fast. Tools combine as needed; quality stays my call.

Build

My role: Review every detail like a delivery.

Claude Code turns the designs into real code. I test screen by screen, log issues, give precise directions, until what's built matches what was designed.

Refine

My role: Grow a product from usable to mature.

Ship a minimum usable version first, then refine through real use — filling in details, growing "usable" into "mature."

A few things I've learned working with AI

“When AI says ‘done,’ often nothing has happened.”

It reports “fixed” with complete confidence while not a single pixel has changed. I never trust the report — only what actually renders, screen by screen. That's rule one of working together.

“One correct example beats ten adjectives.”

Ask it to “make the spacing feel better” and it misses four times in a row; tell it to “copy this page that's already right” and it lands at once. AI doesn't lack ability — it lacks reference.

“Don't expect it to be smart. Break the task down until it doesn't need to be.”

One thing at a time, clear boundaries, clear acceptance criteria — and it rarely lets me down. Most “AI can't do it” is really “the task wasn't broken down.”

“It can generate ten options. Choosing is forever my job.”

Generation is cheap; choice is expensive. The more critical the judgment, the less it can be outsourced.

“What changed isn't just speed — it's ownership.”

At a big company, I was one link in the production line; now AI hands me the whole line. It didn't take my work away. It gave a whole product back to one person.