Figma's AI Prototyping Tool is Here and...It's a MUST TRY | David Kossnick

Figma make is a prototyping tool that can start with a simple prompt, with a screenshot or with mock ups. Seen a bunch of really cool cases like teams building out their specific workflows for their marketing idea campaigns, using public APIs to figure out their next train near them in a form factor that they like most. And so we think we're going to be the best place for starting from designs to get something really high fidelity that is usable and functional out of the game. Alpha testers for figma make is my 6 year old son. He's made like 10 games in figma Make. He comes home from school with sketches of ideas and he's like, can I give this picture to make and start making the game?

What was the most challenging part about building this product? Okay, welcome everyone. My guest today is David, the head of product for Figma AI. Figma just launched make, its brand new AI prototyping tool. So really excited to get David's demo how to use Make. Discuss what it took to build the product and his top lessons after building AI products at figma Coda and Google for a decade. Welcome David.

Thanks Peter. So great to see you.

All right man, so super exciting. I've been playing with Make. It's a really amazing product for prototyping and can you just start by showing us how it works and showing you some examples?

Sounds great, yeah. Figma make is a prototyping tool that can start with a simple prompt, with a screenshot or with mockups. It's designed initially for designers and PMs and engineers to bring ideas to life and to start with concepts that come from figma and more and get those into something you can play with and feel. And so just to give you a quick run through, we'll start with just a text based prompt here, Modern image gallery where you can tap around on different photos and transition. And this will take a couple minutes and you'll see it's starting to reason about what the scope here might include. So here's how I'm going to plan to do it. Create a gallery component that displays images in a grid. These are the types of things I'll need. Then it'll describe the work it's about to go do. As it starts generating code under the hood, you'll see that stream in here as part of this figma built for the first time our own native code composer, complete with syntax highlighting, code auto complete error console and more. You'll be able to go in at the end and actually edit that code directly. Here we go, that one's pretty fast. Let's see how it did. Pretty decent. You have a sense of a starting point for this.

That's awesome.

Hop on over, look at the code and actually one cool feature. I'll call this the pointer here. You could go ahead and click on a component and then hop over to the corresponding part in code as well to go ahead and look back and forth on how it works.

I wonder where it's getting the images from. Is this getting somewhere?

Yeah. So there's a number of ways it can generate image. The most common one is via unsplashed. And so we found for placeholder imagery especially to get an idea across. That's a really helpful and very fast starting point.

Awesome. Show us some more demos, man. What else you got?

Cool. So, yeah, let's do another one. So just starting from an image here, I'm going to take a picture of a solar system and say, make this interactive. This one's a bit more complex. I'm just going to switch over because it might take a little bit longer to do this one to show what happened. When I finished running it, you can see here again, it went ahead and reasoned through it started generating, wrote the code, told me what it did. You have this inline, immediately playable implementation. I can try clicking around here. I can see that the planets are rotating. This is amazing. Especially for someone who may not know any engineering skills or even who does. But wiring this all up could take a lot of time. I suddenly have a much better sense of how this is going to look and feel based on just the pure visual simple prompt.

Yeah. This is amazing. I can use this to teach my kids how Solar System works. That's great. Yeah.

And of course. Oh, sorry.

Oh, yeah, I was just saying. So that just came from a static image, right?

That's right, yeah. So that was just uploading a PNG here.

Okay.

Yeah. Of course. We are figma. We're for product designers of digital products as our primary Persona and scenario. Of course, we've done a lot of work to make that really shine. If I hop over to a more normal FIGMA file, you can see here, I have a whole bunch of components and layers. I'm just going to copy this thing and come back to a new make and I'm going to paste this in. I'll say, make this interactive, including the dashboards. And again, I'm just going to hop over to what happened after it finished running. And what you can see is like, again, I have a pretty functional prototype. Let's Try clicking around. It says achievements show and hide all these Edit profile. Yeah, that's actually pretty solid.

That's amazing. I think you're understanding it, man. I think this is actually really amazing. You actually copy the Figma frame or the artboard.

That's right under the hood. What happens when you bring things in from figma which you can also do over here, pulling things in from your team or from the community. We're not just bringing imagery, but we're bringing all the rich structured data that Figma has. So it's like all the layers, all the metadata, all the exact details of styling in every configuration. And this really allows the translation from design to code to shine. And so we think we're going to be the best place for starting from designs to get something really high fidelity that is usable and functional out of the gate.

Can you pull up the image again?

This episode is brought to you by Merge. Product leaders cringe when they hear the word integration. They're not fun for you to scope, build, launch or maintain. And integrations probably aren't what led you to product work in the first place. Luckily, the folks at Merge are obsessed with integrations. Their Single API helps SaaS companies launch over 200 product integrations in weeks, not quarters. Think of Merge like plaid, but for everything. B2B SaaS organizations like RAMP, Drata and Electric use Merge to access their customers accounting data, to reconcile bill payments, file storage data to create searchable databases in your product, or HRIS data to auto provision access for their customers employees. And yes, if you need AI ready.

Data for your SaaS product, then merge.

Is the fastest way to get it. So if you want to solve your company's integration dilemma once and for all, book a meeting@merge.dev peteryang and receive a $50Amazon gift card when you attend.

Now back to the episode real quick. Yeah, dude, it looks exactly like the image, man.

Yeah, I mean look, to set expectations, you know, it's AI. These are non deterministic systems, you know, you'll definitely have cases where it'll get confused or fall over and we are really sweating the details on quality. But also we have a lot of context to bring to this problem. A lot of context about the data set, about the data sources. And so we've really excited to make it much easier to start from a design and bring something to life.

So I think this is kind of partially the answer already. But I have to ask you, there's a lot of prototyping tools out there now, right, with AI. So what Are the top three benefits of top three unique value props for make?

Yeah, a couple. I mean, first off, we're just getting started here. So this is an early beta. We're really excited to get a lot of feedback and we're really excited about this space. But for where we've started, the first is being the best place to start from design. Again, we know a lot from the Figma platform from our other products on the context that you're bringing in, and we think it'll be really, really great to start from a mockup or a sketch and go to something that's playable here. I think the second one is Figma is a platform where all of our products are multiplayer first. And this is part of why figma's been so successful. And so I'd actually love to show this to you now. If you open up this file, I share this with you. You can come on in here and make something with me. So you'll enter the same file. My avatar is here in the top right. I had a chat and yeah, you'll be able to see the changes updated in real time as I do.

Let me. Do you share with me?

Yeah, here, hold on one sec, let me switch windows. Is there a chat in Riverside as well? Yes, here, I'll put it in the Riverside chat. Also Studio chat. Okay.

Let me find. I never use the Riverside chat. Oh, there it is. Okay, let me open this. All right. I'm eagerly waiting for this to load.

Okay, so I see your face up there. So one thing you can do is open up this pointing tool and you'll notice we are figma. We're known for direct visual control. You'll be able to go ahead and make changes. So we'd love for you to make some visual changes. And I should see it update real time as well.

Okay, let me. Okay, I made it red.

Okay, there you go. Nice.

And I can prompt the AI to on the left and ruin.

Yes. So you can start chatting. You'll see your face show up there. And you know, another cool thing is. So let's go down to the dashboard here. You'll notice the hover state's a little bit weird. Like, why is there that kind of gray background? We can go ahead and select this element. And you'll see this toolbar here. And there's a part of this toolbar that lets you basically do a contextual prompt. You'll see this says div, and DIV shows up in the prompt box. So I could say make the hover state for this not change the background color and instead animate the bars. You'll see that that prompt show up in your screen as well. Then the other really cool multiplayer feature is if you select an element, you can also hop on over to the code view to see exactly that piece of code. Looks like this one's getting rewritten right now. But once you're in the code view, it's also a multiplayer code editor. You can go ahead and update that code live and I'll see that updated on my end as well.

Wow, this should be a V1. You're making us look bad, man. This is supposed to be a V1 product. This is pretty. It's pretty polished and pretty feature rich already.

You're very kind, Peter. Yeah, we're really excited. I think we're just getting started here. Let's see how it did on this. Updated the static bars. The hover bars grow slightly on hover state. Interesting. So didn't quite get it. You'll see. The interesting thing about when it does fail is you can often work with it to get it to a place where it's better or closer. And of course you can always go directly into there. Could have changed as well. But these things are not foolproof. And part of why we're in beta too is we're really eager for people's feedback. So you'll see thumbs up, thumbs down. And we're looking for lots more feedback on which specific example cases work and don't. Which we can talk about more our eval loop. But it's a really important part of our process on quality as well.

Yeah, let's talk a little bit about building an AI prototyping product. Like maybe you don't have to review a secret sauce, but how does it actually work behind the scenes at a high level? When I submit a prompt, make this interactive, it sounds like just like a planning step. You prompt it to plan out. Make a spec first.

Yeah, yeah. So we're trying to help explain some of what happens in the response from the AI itself. So you'll see in reasoning, the first thing it does is it sort of works through for itself. What scope should it be? Then as it comes to a decision there, it starts getting into execution and gives you a readout of like, all right, I'm going to get started on these things now and then. We've also invested a lot in code streaming. These things can take a while. It can take a couple minutes for it to generate all the code. And so being able to understand what's happening even While it's going is really important as part of the user experience. And so we invest a lot in the latency and user experience around code streaming to understand like, okay, it's working on one file and it's going ahead and generating it in full or. No, actually the file exists and it seems like it's just editing one part of it. And so it's kind of making some decisions based on your starting point, which could be an existing code base, could be structured data of a figma mockup you've brought into that prompt, or an image or whatnot to decide what to do.

Got it? Yeah. Because it could take a few minutes. So showing that it's doing the work is good for the user to understand.

Yeah, for sure.

And do you have any. There's like a bunch of Vibe coding tips out there. Like, do you try to bake this stuff into the prompt itself? Well, I guess one of them is like to build a plan first. But like, you know, do you have like some tips around like, hey, use this framework or like, you know, you use the stuff.

Yeah, it's funny, you know, I feel like Vibe coding has a version of that now and like prompting has had one for a while as well. Like, like, you know, in the more general prompting version there's been memes like offer to tip the AI or like tell it to do it faster or like threaten it that if it doesn't do a good job, you know something bad's going to happen. Like there's a mini version of that happening in Vibe coding as well. You know, I think the reality on which those are like actually meaningful or helpful is moving really quickly and models are constantly changing. What's baked into at the product layer is changing a lot. You know, we're doing a ton of iteration on our side with the goal of users not really having to think about it. And like, what we find most users is they want to come in with a high level idea and get some sketches and get some stuff to play with and get a sense of if it's what they want and then refine from there. Or they come in with a very specific idea, like a mockup and they want to feel it and see what corners they haven't thought around and how that should influence the approach. And so we don't want users to have to carry the burden of like knowing what the latest best practices are, how to exactly phrase things. Like we want a system that's a bit more flexible and sort of simple to understand.

Okay. So we've Been talking about building AI products before. In practice, a lot of it is just like you said, a lot of it is iteration, maybe some notebook, just trying to edit the prompt. And a large part of that is obviously evals.

Right.

And maybe you can give us some idea of the types of evals you have for this product.

Yeah, Evals has been a really interesting one for this project. We realized pretty early on there were actually a couple different axes for quality for users. We have two scores today. We have a design score and a functionality score. They're somewhat orthogonal. You can actually have cases where it gets the functionality completely right and it looks totally wrong relative to what the user wanted. There is still legitimate user value in that. I think the other thing that we learned is even though we're targeting designers as the hero Persona, there are different sub Personas and different use cases and scenarios where the design score versus the functionality score matters really differently. We've had a testing group which you've been a part of. Thank you for all your feedback. It's been really interesting to see how many internal tools cases there are where the functionality really matters, or how many games there are where the feel matters much more than necessarily all the nuances. And so I think that was one early insight, which is like, there's two scores, and the bars for those scores from users vary a lot by Persona and scenario. And so it's not like there's one number you have to hit everywhere. It's like there's different numbers you have to hit on different axes for different demographics and targets.

Okay. So you basically just have real people actually give the scores. Like real figma employees are like the beta head testers.

Yeah. So we've done a couple ways to get. To get the scores themselves. So, you know, those. Those we started initially with the project started initially with like a prototype, and it was fairly basic, and we did a bunch of internal testing and the question was like, how good, how. How far away is quality from a point of viability with this? And so we had like a giant figjam board. In the very first eval, we had the whole AI design team split. Spend hours just going through and trying prompts and recording, like, input prompt. Here's my file, here's the screenshot, here's the design score, functionality score in my notes. And, you know, it was a score of one out of four for design and for functionality. Four being like perfect. Three being like good. Two being like not that good, one being like absolutely terrible. And from there, you know, we started building it we felt like, okay, there's a path to making, making this viable. We expanded that same process actually to the whole company. And it was pretty wild. I think it was the most aggressive sort of widespread, pervasive dogfooding effort at Figma. We called it the Great Bake Off. And we had in person sessions, like 12 of them across time zones for all of our global employees, where people would meet up and like jam live, sort of like group vibe coding sessions and like use this same kind of thing. Jams again fairly early in the project in a somewhat structured way. It's like, here's the exact things, here's what works, what didn't. Like we didn't even have thumbs up and thumbs down built into the product at that point. And so we're kind of scrappily using the FigJam to get feedback very quickly. And that was so, so helpful. Like that became like the first thousand items in our eval set. We're like, these were the great cases people love. We should not lose these. Here's the most frustrating cases people want improvement on. We took those verbatims and it was like a very scrappy and very, very helpful process. And then from there, you know, we do have a team of contractors we work with in order to scale up a bunch of the human judgment piece to have a faster feedback loop. And so that was really helpful sort of canonicalizing a bunch of these data sets before we even kind of were rolled out to more users.

What are some examples of stuff in the eval set? Did it run? That's one of them.

What are examples of.

Sorry, what are some examples of stuff that's in the eval set? Like the eval criteria? Like, you know. Yeah, yeah, yeah.

I mean there's, there's, there's interesting cases. And so like we've had cases where, you know, make me this dashboard. Like I showed you, something would be visually right and like nothing would work at all. And you'd be like, make it interactive, like add the hover states. Like literally nothing happens. You're like, it's nailing the design. But like something is really broken in functionality and like, we're not sure why. Got it. So you'd have cases like that and then you'd also have cases that were super interesting where you give it a very ambiguous prompt and there are many different right answers that have trade offs to them. I think that was another interesting piece is having designers and PMs very involved in the eval set is really important. It is product design process. Of when you have an ambiguous prompt like this, these are reasonable answers and these are not. And I think there is a human judgment involved in like, yeah, make me a micro site for my team meeting up. And it's like, well there's a lot of ways that could go.

Yeah, yeah. Unfortunately people don't necessarily give AI specific instructions all the time.

So for sure. And I mean there's an interesting question on like, you know, how what is the back and forth to understand user intent better as well?

And do you have, so you have like, basically you got the entire company rally toggle for this, but do you also have like some sort of AI as a judge, like synthetic eval or is it too complicated for this product?

Yeah, great question. So there's actually four types of evals. Probably should have started with this. There's actually four types of evals. One is deterministic evaluations. And so you know, we also have an AI feature that's been live for a while called AI Text Tools inside of Figma Design. So if you have a text node, you can ask AI to write stuff, you can ask it to shorten text. And like asking AI to shorten text is a deterministic capability. Like you can look at the output and see did it shorten it. And like we've had bugs for a while. Like if you run it on like a sentence, it'll turn it into a paragraph. Like that is a, you know, very measurable, very repeatable logical error.

Similarly snow thing, right?

Yeah, yeah, yeah, exactly. And similarly for code generation, like if you generate code that does not compile, that is like a logical error. So that's the easiest form of eval is like, is it correct or not? I think the second type is qualitative. I think there's both like humans on your team as well as like a scaled version of that with contractors. And like, you know, before Genai, I used to work at Google. Google search has like 100 page public document describing how search raters should evaluate a search results page. Like what is good or bad quality? And you should like click on each page and look at different things about it. They have like thousands, tens of thousands of people trained on this hundred page book on how to evaluate things with good judgment. And so I think like every product team right now is kind of developing their own book of like, okay, if I'm going to evaluate stuff like only have so much time, it's like how do we make that go faster? And I think on the one hand you have humans do it on the other hand, you could have AI do it. And so is that 100 page book basically a really long prompt now to AI as a judge. I think for some problems that's more viable and yes, you can actually get AI as a judge to work pretty consistently and be fairly aligned with human judgment. I think for others it's harder. And I think in our domain of design, there are really interesting cases where even designers will disagree on what the right output is. And, and it's like, how can you expect an AI to consistently do a good job if a bunch of humans have debate what is a reasonable behavior in that case? And so I think it can be helpful for order of magnitude changes. Something really broke this time, but I think it's often pretty nuanced for is this actually really right or going in exactly the right direction?

It can't do tests. Right. You have to get humans to do that.

That's right, yeah. And then the final form of eval is like usage. And so we're not live, we're now live with figma make. But A B tests are really helpful to understand, you know, different approaches like how are users behaving differently?

Got it. And what was the most challenging part about building this product? I mean, you've probably been working on it for a while, given how polishing it is.

So yeah, yeah, I mean, there's been layers of challenge, I'd say. So the first has been infrastructure. So figma make came out of a project called Code Layers. And so we have a whole project called Sites. We're shipping a new Sites product at Config, I guess we just shipped a new Sites product at Config. And Sites has been a long, big effort. And it involves totally different rendering technology than figma design and involves converting figma nodes into renderable, publishable things. There's a ton of latency scalability work. And so that's been like a very long arc on bringing Figma primitives and code closer together. And out of that foundation came Code Layers, which is like, well, what if we let you write React on Canvas and publish it inside of Sites? What if you could chat with that and what if you could convert designs into code in that context? Then as we were working on that in a Sites context, we were like, you know, this is really good. And there's a lot of cases where you don't want to just do this on one part of a site. You'd like to do it on the whole site. And that became figma Make. And so we sort of pulled out that part of sites turned it into its own form factor and shape. And it really was able to benefit from a lot of long infrastructure investments on sites, on code layers, on agentic interaction with code, but in a very different form factor. And so that was the first bucket is like there's a lot of underlying infrastructure that this is built on top of. I think the second challenging area is quality, speed and cost. And like the model, quality is a really hard hill climbing grind and there's a lot of interesting trade offs, business trade offs and product experience trade offs like where, where do you want to be in the quality speed cost world and how do you move in a direction that you know, meets user intent or expectations in the right way? And it's like doing good evals, hill climbing well, working through all of those details, having an experiment log, like all that is really detailed work that takes time and a lot of nuance. And then I'd say the third challenge has been this is the level of non determinism in AI. And so we had interesting cases where you would give a design, an image of a design to the agent and ask it to do something and it would sometimes do a pretty good job and it sometimes fails spectacularly and it was hard to figure out why. And we had one of these cases and it turned out the image format we were passing to the model was wrong and the model just couldn't read the image at all. And we were like shocked that when it literally couldn't read the image much of the time just based on the prompts, it did a pretty good job guessing it, but sometimes it spit out something completely unrelated. And it was like, how could we know where the problem was? How do you diagnose that? And so being able to have full visibility into all the steps of like, oh, we accidentally truncated message history and so it doesn't know about something earlier when you thought it would. Like there's so many bugs and edge cases where it's like really hard to tell. Like wait, why is it now not quite right?

Yeah, that's why I think it's also like you try to keep the system as simple as possible so that if something breaks, you can either fix the prompt or eval or otherwise you have too many steps. It's very hard to find out what's going on.

Yeah, for sure.

You know, people keep talking about like, you know, all the PMs got to become AI power now and so on and so forth. And I think probably your team is most giving your building AI product is the most AI powered team. So like, how, how do you guys, how do you guys actually build this product? Like it's, it's not, I'm assuming it's not like a waterfall development process.

Yeah, I think, you know, you and I have chatted about this in the past, but I think like our prototype is worth a thousand marks. Um, and the great news for industry is like the cost of making prototypes is going down so much. Like the speed, the expertise, the team size. And so, you know, it's been fun working on a prototyping tool, figma make, to be dogfooding it itself. Like there's been a whole bunch of ideas for Figma make that we've just PMs and designers have gone ahead and prototyped in Figma make with no engineer in the loop and like gotten a sense of how it would play and feel. It's like, oh wow. Yeah, that's like really helpful. Yeah, I think, you know, at the design level I'd say prototypes have been a big thing for a while. Like figma's had a concept of noodles and being able to link things, but it is more time consuming, it's not quite as high fidelity. And I think we've seen, you know, in the last year prototypes move from like a very helpful thing in a lot of projects to like kind of the new gold standard as the design artifact of choice. And I think in the next year we're going to see that change to being kind of the new gold standard for the PM artifact of choice. And like, you know, your PRD is a good starting point, but why didn't you just put your whole PRD in a prompt box and get out something you could play with that made you realize a bunch of your PRD actually would feel kind of awkward and not right and actually give a directional sense to a designer and engineer of like, this is kind of the shape of the thing and like these parts we know are problem. I need your help figuring out what we should do there. Like these parts definitely won't scale and like how would we actually build it. But getting further into the problem, getting further into the questions, I think that's a huge industry trend and is really exciting and also personally empowering for as a product person myself. And then I'd say the other reason for AI teams in particular, prototypes are so helpful is you don't know if it's going to work until you kind of get it hooked up. And I think it's like you can do evals in isolation, but you Also don't really understand like what the expectations are in a flow until you feel it or users can feel it in some way. And so, you know, I talk sometimes to, to my team about amaze the maze of choice in AI and like, you know, you have an idea for a feature and it might be you can go straight through the maze and ship and like the whole thing just works and like the model is supported.

Yeah, maybe 1% of the time and.

The prototype will improve that. But it also might be the case that like you can't even get into the maze and, and like it's too slow, it's too expensive, it just is. Like it's too hard a problem. Your idea won't work on today's models and you gotta like leave it on the shelf and come back to it. And then there's also cases where like you get into the maze and it sort of works in some ways, not in others. And this is a fairly common case and you have a choice to make. I'm like, do I go out of the maze and wait? Do I try to add product constraints or narrow the scope and make it easier so it can work? Or do I take out a sledgehammer in the middle of the maze and try to bash through these walls and fine tune and build custom models and just solve the hard data problem and make quality get to a point where it's actually good. I think a lot of the art of being a modern day AIPM is figuring out when to take what turns into maze and is like, is this actually solvable? Should we actually change the constraints in some way that simplify the products, deliver most of the value to users?

Yeah, you have to be comfortable with getting all the way to dog food stage. And then the model doesn't work properly, so you have to hold off on the launch or that feature for a quarter or something until a new model comes out.

Holy. And so how can you do that before you staff engineers on the project? How do you do that before you make it a full thing is you move prototyping earlier, you get feedback from users on what is the quality bar in this experience. Maybe it works out of the gate, maybe it doesn't, and maybe that frames what's left what the big hard parts of the project are.

So let's say you want to add a new feature to make. Do the designers just prototype two or they go to think about design first? That's the key question. Which one do they go to?

Yeah, I've seen all of the above. I Think it sort of depends. We definitely had cases where you can, starting with a good existing make, you can kind of prompt your way to feeling out how a few changes would feel and be like, okay, yeah, this would, this would feel good. There's also cases where it's like, it's hard to explain and like, you know, it's fun. It's funny to your, your point earlier in this conversation, like sending a one sentence description of an ambiguous thing to an AI is kind of incredible that it can come up with anything good. Like imagine sending a software engineer a slack DM being like, hey, build me this thing. They'd be like, what are you talking about? Like, why are we doing this? Like, what's the goal? Like, what are the constraints? Like, you know, they'd be asked for a whole bunch of thoughts and like, what are we actually trying to do here? And I think it is really powerful that you can quickly get to a directional starting point. But also specs exist for a reason, mocks exist for a reason. Being able to be really specific is what figma is great at. And so being able to go and say, no, this is exactly how I want it to look and ladder that into this framework. Let me try it there is also really powerful.

Yeah, I think, I imagine you guys just probably just have meetings, have big jams and look at the makes and talk about them instead of like looking at PRD or email, you know, like it just makes the things more fun, right? Like who wants to look at a prd? No one wants to read my prs. Yeah, so look at the, look at something as a user from a user facing point of view is like way more fun and also probably better and looking at a doc, you know.

Yeah, for sure. And I think it also, you know, it's a bunch of, it forces more product thinking too where it's like, you know, how many PMs have been in a situation where you just didn't think about like a bunch of edge cases or other scenarios until like it's staffed and engineers working on it and they're like, but how should this work on mobile? You're like, oh shit, I forgot something about mobile. And it's like, you know, you get to a prototype, like you got to see the things that you weren't thinking about that like, should have, you should have been thinking about a little bit earlier.

Yeah, there's no such thing as like coming up. Like I feel like very rarely I have a perfect PRD or you know, even a perfect design. Like in one One go. You have to play with it and like make change changes. So it's. Yeah, yeah.

Oh, sorry. Go for it.

No, go ahead, go ahead.

Yeah, I was gonna say, yeah, I was curious, like, you know, you've been playing with the products a bit. Like what, what your reactions were to it and like what you found it useful for or less useful for.

I was really. So I think I've been playing with all the prototyping tools. Right. And, and what I was really impressed by make is it did take a little bit longer than some other tools to generate the prototype, but the prototype was like much more beautiful and feature rich than I expected. So like once it generates, it was a very, it's a very positive magic moment. Yeah. And I haven't, you know, because we're talking about right now before the launch, so I haven't actually shared my designer and other people, but I think the collaboration features that you show me are like pretty, pretty amazing. Like, I joke that like, you know, when my designer shares design with me, like, you know, I'm like, hey, but what about this variation? What about that variation? And then now instead of having to ask her to make a bunch of designs around this stuff, I can just make it right for better or for worse and then show it to her. And then we can show some users, we can have a debate and then we can figure out what to do. So it just speeds up the iteration flow a lot faster.

Yeah, that totally resonates.

Yeah. I mean, I joke that I'm like a IC 0.1 designer now, but let's kind of wrap up by talking about figma's journey into AI. Right. So let's go back to like a year. I think you joined Figma like two years ago or something, right?

About a year ago, actually.

Okay, so. Right, okay, so let's go back to a year ago. I think in the previous Convig, Figma launched some AI tools and then one of the tools got like a mixed reaction which is like this like text to design tool. Right. So I guess, like, what did you learn from that experience that you kind of apply to building make?

Yeah, yeah, for sure.

Yeah.

We launched a feature called Make Design and it had a pretty mixed reaction for a bunch of reasons. You know, I think for me there were two big learnings. One is be clear on how your system works. I think for make designs. We also at the same time released a new data training policy where we could train on user data with their consent. We're very transparent about it. Give users lots of control. And I think people got a little bit confused sometimes and thought like we had actually trained on their data already because we didn't say how the feature worked, we didn't say what model was powering it or anything. And so we've been much more direct with make like you'll See powered by Claude 3.7 Sonnet in the prompt box. And we think it's totally understandable to want to know how that model was trained and what the biases are in it. And I think, you know, Anthropic's done a fantastic job publishing a lot of detail on those types of things that customers reasonably want to know. I think another big learning was, you know, we being clear on what something is for. And so with Make Designs, some designers perceived it that our intent was to replace their job. And I think even the name of it, Make Designs, it's like we will make designs instead of a designer making designs. And it was funny because our internal name for it was First Draft, which we ended up calling it later publicly, which was very much the intent is just like, save you some time in getting to the starting line. I think this was also a message that we're was confusing in the market and people didn't quite get. It was like came across as a we're going to do everything designers do. It's like, no, no, no. That's like, we love designers, we're a tool for designers. We want to save designers time, we want to give you superpowers, we want to empower you. And so I think with this we've really focused on prototyping and starting with marks, like extending what you're able to do beyond your current set and trying to be really clear in that in our message to designers, it's like, we love you, we are you, we want to give you more powers and make you able to take your vision much further and not be limited by your current tools.

I love that. Yeah. I think even with the prototyping stuff. Right. Like this stuff looks beautiful but like is my engineering manager. Let me check this stuff into code. I don't know, it's more to like just get feedback from users. Right. And then maybe you can build it properly afterwards.

Yeah, for sure.

Yeah. It's like a. It's like a co pilot. More. More than just like, you know. Yeah, cool. And any kind of closing, words of advice now that you've been on this journey, I'm sure you feel very proud for team and yourself and you know, I'm very genuine when I Say it's a great product, it's awesome. What are some closing words of advice for people trying to build similar things? Yeah.

Words of advice on people trying to build other AI products. Yeah.

Other AI products or trying to get into the stuff. You're going to have a lot of patience with the stuff. There's some closing words of.

Yeah, I mean it's an incredible time to work in software and to be a maker. And I think as a company, our mission is to make it much easier to go from idea to product. And I think it's been personally inspiring to see just how wide a range of things people are making in figma make as well as in the market that people are making with new AI tools. And as an example, one of the alpha testers for figma make is my 6 year old son and he started playing iPad games on the weekend and so he started vibe coding with me and he's made like 10 games in Figma make and he's like a 6 year old. I'm like, oh my God. Like I didn't get to playing video games till I was 12 and making them till I was like 18. Like it's such a different generation now and it's so perspective changing for him too. It's like he comes home from school with sketches of ideas and he's like, can I give this picture to make and start making the game? Yeah. And so I think my biggest advice to makers in general, both AI teams, professionally, hobbyists, students, is like, just, just start making. Like you're going to learn by doing. You're going to build empathy. You are, you know, build for yourself and your friends. Like you are the user, you have empathy. It's easy to understand if it's like working or not, start there and get to something that's playable and like follow, follow that direction.

Yeah. And I think, and I think one of the most exciting parts as a builder is putting your product in the market. And like, you know, make has to be best in class at figma to prototype. Right. We talked about. But like, it'll be interesting to see how people are using this stuff completely outside of that use case. Like your kid making games. Who knows, man? Maybe that'd be the primary use case.

Yeah. If we have a minute, actually I collected a few of my favorite examples that I've seen people make.

Yeah.

So yeah, very much to your point. You know, we optimize for this in designers, professional designers, starting from designs. But it has been amazing to see just how many different Personas and Scenarios this works for. And so here was a case where someone made a very simple spray painting tool in like five minutes. And it's like, you know, imagine making software to make software, you know, making a creative tool yourself. It's pretty incredible. Another one, this is one of my teammates actually is a gamer and made, you know, a whole 3D first person shooter. Let me refresh it real quick. The arcane realm awakes.

Oh, man. Wow. This is like a. Wow, this is like a 3D first person shooter.

Yeah, yeah, this is like a full on 3D first person shooter. And like we did not optimize for this use case. Like we did not, we did not go out of our way to make like game creation and 3D awesome. But it's like pretty amazing how much you can get similarly like internal tools. Like we've seen a bunch of really cool cases like teams building out their specific workflows for their marketing idea campaigns, personal journaling tools that have it configured exactly the way you want. Some artists just playing with vector fields and graphics. What is this vibe like? People trying to figure out they're using public APIs to figure out their next train near them in a form factor that they like most. And then. Yeah, actually this was my son's game. Is like feeding cats brunch. Nice. Yeah, There you go. Yeah.

Wow. This is great stuff, man. This is awesome. It must bring so much joy as the PM on this to see all the different use cases.

Yeah. I'm standing on the shoulder of giants with an awesome team I feel incredibly privileged to work with, you know, across figma itself, leveraging infrastructure from figma to sites to CO layers and then incredible group of makers who've been extremely passionate and hardworking and, you know, very creative on this project.

So where can people find this feature? Let's travel forward in time. This thing is launched. Where can people find this feature?

Yeah. So Figma make is rolling out over the coming weeks to paying customers and you'll find it as a new file type in your Figma file browser.

Awesome. David. Well, thanks so much for your time. Let's all go out there and make stuff. Thanks to your products. Very exciting.

Thank you, Peter.