While the "5 Whys" technique is a terrible interrogation technique for UX research, it seems surprisingly effective with LLMs. It made the model realize what it had messed up and correct itself. The 5 Whys were developed to determine the root cause of production line defects at Toyota factories. Fitting for what we do with models today.

As long as design represents quality, how fast we ship cannot be its measure. Design also means creative experimentation. You try something, see what happens, change your mind and try again. We have more time to explore further, instead of expecting an earlier finish every time. People shipped "fast" before LLMs came and sometimes they shipped without any user research or a dedicated designer: optimizing for quality is not new. Faster tools are a chance to use more creativity and make better things if the designer has time to do so.

"Report" is my all-time favorite comic strip. I have it framed on my wall. I just went by it and thought how relevant this one has became lately... (Poorly Drawn Lines, 2018)

One-shot animation test again of a low-poly German Shepherd. This time, zero manual iterations, just checking the first result:

Fable 5 vs. Fable 5.1 for AI 3D generation. Same task: one-shot animation of a low-poly German Shepherd. Fable 5 took me many iterations to get right. 5.1 has a serious head start. w/ Claude Design.

With LLMs, designers started to ship front-end. That is where much of accessibility actually needs to be handled, not in Figma prototypes where color, contrast and size mostly cover it on a static canvas. Keyboard use, focus, semantics, screen readers, motion, zoom and actual interface behaviour are now part of our daily practice. Surprisingly, I don't see LLMs generating accessible interfaces by default. You would think standards are the kind of things machines should be good at following. Relying on 3rd-party Skills feels like patchwork. At least, LLMs are good at pointing out what is missing, so you can do the job for them.

Let's see how this evolved lately. My last test was in May.

Been working on a Design Systems LLM logic for existing products for the last few weeks. It takes a lot of hands on work. One of the TLDR results is: the "design.md" magic does NOT work. No one shot Skill is feasible. Don't trust them, use as inspiration only.

Watching @Claude making annotations on my web app in real time is rewarding and a little deceptive. Opens browser, scrolls through everything, clicks on elements, and writes design critique annotations directly on the UI. Stack: Agentation (by @benjitaylor) + agent-browser (headed browser CLI) + Claude Code skill. Made 15 annotations on my new university tool: broken visual hierarchy, misleading empty states, competing with CTAs... and fixed all 11 in one pass. Super useful, but Claude still misses deeper issues around value proposition and user flow (this was made with gstack, btw). Senior UX designers are still very much needed to orchestrate this.