Faster and deeper with AI

The way I work has changed every few years with the technology, and nothing has changed it as much as AI. What it changed is easier to see against how I worked before.

I always prototyped

I came to interaction design with a technical background and a passion for motion graphics, so trying an idea was never a separate step from having it. I would turn the idea into a prototype, feel it, and let that decide what came next. I built tooling for it. For Instants I made Swipe-Through, a prototyping platform on top of Flash: craft the animations on a timeline, then drive the timeline with touch, so an interaction can be felt on the device before any of it is coded. In 2015 I released the custom components I had extended Flash with as open source. At Canonical a working prototype and a behaviour spec were what carried the Launcher through a large team of engineers. At Spatch the calendar was prototyped and tested in the wild before it was built.

I did it to explore the solution space: prototype, feel the idea, test it, and iterate toward the best outcome. Exploring properly is harder than it sounds. I have seen many designers explore two or three variations at most, and not radically different ones.

Now anyone can

That is the part AI changed first. It is now easy to feed a prompt and a few screenshots to an agent and ask it for alternatives. Earlier this year I gave the same brief for a small modal to three tools at once and got three directions back, and on another day a couple of ideas I would never have come up with on my own. A new flyer interaction for Beatflyer went from paper, to a carousel vibe-coded in v0, to the full layout in Figma, to the real thing built with Claude Code, in a weekend. You no longer need a technical background to prototype, only the willingness to explore.

Polishing the best candidate is another matter. The agents still need quite a bit of guidance there, and the code they write still adds up to buggy software if nobody reads it. I review AI’s changes in pull requests and have AI review them too, and I published Product Map, a skill that keeps a map of what the software actually does, because as more of it is written by AI you need a human-friendly way to stay on top of its behaviour.

Choosing the problem

Exploring starts before the screens, with choosing the right problem, and this is where I think AI is most useful. Research, support tickets, reviews, surveys, product data: the material that used to take a week to read now takes an afternoon to interrogate, and it points at the highest-value problem rather than the most visible one. On Beatflyer the design partners group and the surveys decide what to remove next; AI is what lets me actually read all of it. GenTable, a research project of mine, applies the same idea to the models themselves: one column per model, one row per input, a judge scoring every cell, cost and latency beside the score, so choosing a model or a prompt is something you look at rather than guess.

The trap

In startups being fast is often celebrated more than being right, and innovation and speed don’t always go well together. When speed becomes the only goal, the productivity gains from AI don’t get reinvested where they matter most, exploring alternatives and validating choices. We just ship the first decent output, faster. The new speed of execution pushes us to think faster, not deeper.

So I am sceptical of AI design processes that skip the uncomfortable parts: research, exploring several directions, testing with real people, revisiting assumptions, killing good-looking ideas that don’t solve the problem. Low-fidelity thinking is not obsolete because everything can be rendered instantly. There is a reason many of us still start by dumping our brain on paper.

I have been on the wrong side of this. In some startups I have worked with, there was a need to show the pace of new features, to investors as much as to customers, and the team was under real urgency to deliver it. Features went out quickly, and many of them didn’t move the needle, because they hadn’t been thought through.

The motto I prefer is the one Jeff Bezos gave Blue Origin, gradatim ferociter, step by step, ferociously. In his words: you can’t skip steps, you have to put one foot in front of the other, there are no shortcuts, but you want to do those steps with passion and ferocity. AI lets us do every step faster. It doesn’t let us skip any.

The same principle became a product decision at Glide. The AI builder could generate a whole app from one prompt, and I put a mockup in front of it: the agent asks a few questions, shows something the builder can look over and change, and only then builds the real thing. It is the prototype step again, this time inside the product.

What stayed

I still want something I can feel before it is built, whether a Figma prototype or an AI mockup, and several directions on the table before one is chosen. The machine does more of the drawing now. The case studies on this site are written the same way, with each decision next to the options it beat.

October 2026