Michael Riddering’s Dive Club interview with Loredana Crisan, Figma’s chief design officer, is about how Figma wants a bigger, faster, more explicit design loop.

Riddering asks Crisan whether taste gets less defensible when everyone can make more things faster:

I actually think that what happens when everyone has the power to build is actually it elevates the need to stand out. You could buy those pretty funny or soulful congratulatory cards, birthday cards, whatever cards, and they have the message written inside. And it’s a great message. Do you just give somebody that card? You could, but that’s not what you do if you’re a good friend. That’s not how you show up. And so I think that the more software is getting created, the more articles are being written. Basically, we’re in a place where everything is getting produced at such high, high rate. And we already had trouble consuming what was already there. And so what we’re going to try to do is look for even more authenticity. […] So this is why I think it’s not taste. It’s just your point of view, your human intuition, what you want for that experience needs to come through. And then the tools that they build are very much important to help you drive that.

I think Crisan is right about the greeting card metaphor. When production gets cheap, the work that still feels made for someone will matter more, not less.

That matters for designers because AI abundance changes the filter. The scarce thing is knowing why this draft, this interaction, this bit of motion belongs in this product.

Riddering then pushes on control, which is where Crisan gets concrete about Figma’s tool bet:

The premise is that AI could get you to 70%. And this is true with motion. You could prompt your frame and then you will get the first animation set up. But then from there, you need to make it yours. So it’s almost like AI sets up the workspace. It gives you the parameters, but then you are the one that makes that thing. […] If there is a color that I want to put in my design, a color wheel will never be replaced by a prompt. So there are moments where you just want to take control yourself. […] And it’s an interesting way of working because increasingly as designers, I believe we’re creating systems, not just screens. […] Design is never done alone. Design is evolved through critique. It’s a place where one of the tools of design is other people. And so it’s really fun to be able to have AI, have agents, have the direct manipulation and direct control and your team in one space.

This is the version of AI design tooling that makes sense to me: the tool accelerates the setup, the first draft, but the product still gives designers enough control to make taste visible in the final result.

When working with AI design tools, we tend to think we need to one-shot a prompt to get everything. That, of course, is far from true. We iterate—a lot—by describing the changes we want, then waiting, and inevitably, iterating again. Crisan is describing a tool that still leaves room for the hand: sliders, variables, shared artifacts, and other people.

And she points to evals as part of the designer’s new job:

Increasingly, we’re working with non-deterministic systems, right? So a lot of what the experience is, is produced by an LLM. Previously, it could be produced by an algorithm. I worked at Meta and can tell you, we’ve had a lot of conversations about the buttons that would exist on a feed story. Those buttons are important, but when you think about what people really experience when they open one of those products, it’s the content. And so all of the conversations that we had about buttons as designers were far, far less important than the conversations that we should have had about the algorithm itself, right? And when it comes to these non-deterministic systems, how do you ensure that they meet your quality bar? There’s this thing called the evals is, on the one hand, the least sexy thing in the world, because you are now in a place where you’re trying to specify what good looks like and how systems that could either recognize that themselves or be augmented with humans that could do that. And it’s fascinating to try to do this on an unverifiable domain like design, because design is not, there’s not a correct answer to design, but there are many incorrect answers.