Design after generation
A series on what evals mean for product design, now that agents make most first drafts. Written by Mark Boyes-Smith while building Glitterball.
- Part 1Prototypes got cheap. Judgement didn't.Generating an interface now costs minutes. Deciding whether it's any good still costs a meeting. That gap is the most important problem in product design right now.5 min
- Part 2AI slop is what the average looks likeGenerated interfaces converge on the same purple gradients, glass cards and
Get started
buttons. That isn't a bug in the models. It's what you get when nobody states a point of view, and it's measurable.4 min - Part 3What design should steal from how AI labs test their modelsAI labs spent years learning how to tell whether a new version is actually better. Five of their habits translate directly to product design. One of their failure modes does too.5 min
- Part 4Measure what you can. Judge what you must.My first design eval flagged so much that nobody read it. The fix was to stop asking a model questions a ruler could answer, and to let the judge say
I'm not sure
.5 min - Part 5Design systems were written for people. Agents need rules.The most prolific contributor to your product's UI may soon be an agent that has never read your design system. Here's what changes when the standard has to be machine-readable.4 min
- Part 6The design leader's job is now to write the evalWhen agents produce most first drafts, a design leader's taste can't be applied one screen at a time any more. It has to be written down, versioned and run. That changes the job.4 min