Scaling Design Judgement at Fin

This week on the podcast Jonny and I caught up with Domingo Widen, Staff Product Designer at Fin (formerly known as Intercom), where he works on the front-end infrastructure team with one foot in design and another in engineering.

We got into what Surge Intelligence, Fin’s new system for scaling design judgement, is all about, plus Domingo’s passion for Vitsoe shelving, among other things. A few takeaways:

1. Surge Intelligence helps Fin hold a consistent design bar across every team that ships, even those without designers.

Their front-end infrastructure team distilled the whole front end into markdown, things like when to use a button, how to use tokens, the right way to import, then rewired every AI skill to read it. Skills are only as good as the knowledge base behind them, so that judgement now reaches everyone who ships. Or as Domingo put it: “I’m building the brain, which goes and helps you, without needing me to be there.”

2. “About 85%” is as far as Domingo thinks AI can get a design on its own.

The last 15% stays human for the foreseeable future. AI wants to please you, he says, so it’ll break its own rules to build exactly what you asked. But this is often overkill when all you really need is a slice.

3. Taste is earned through the reps, and AI can’t shortcut that.

Jonny offered the framing, and Domingo liked it enough to say he’d steal it: the hours spent poring over the details, building things that aren’t yet great, and asking why, that’s what trains your eye. And there’s no way to skip that.

4. We can all prototype in real code now, but there’s still no good way to align on it.

Designers screenshot their live builds back into Figma, just to comment on them, because code-to-Figma still isn’t there. Domingo wants an infinite canvas where the thing is real code, but you can stick a comment on it and say “this is version one.” Nobody’s cracked it yet.

5. Paying for his own AI subscriptions for side projects has sharpened his intuition around token use and which model fits which job.

His current stack: plan with Fable, execute with Sonnet or Opus, review afterwards with Fable. Essentially, models as a small team with roles, not one tool you throw everything at.

It really feels like Fin are operating at the forefront of AI-centric ways of working, so there’s a lot here that I’m sure will become standard practice for design teams within the next year.

Take a listen at nearfuture.works/podcast.