Y Combinator Startup Podcast
00:18:10 · Aaron Epstein, Stephen Haney
[00:18:10] S1: Yeah. So what is the thing when you look at this that makes you not trust it? [00:18:14] S0: Some of the design elements is, like, the the alignment of things, the contrast choices. They feel maybe vibe coded or, like, somebody didn't spend a lot of attention to this. And then your mind goes to what else don't they spend Right. Attention on. The one other thing is, like, this must be a YC company, but I don't see that here. And I would say that'd be a great thing to call out Yeah. As a little bit of social proof Yeah. Or evidence. But, like, this is nice. I mean, there's a lot to like here. Mhmm. I would just... I would work on the fit and finish a little bit. I'll show you Paper Snapshot. Paper Snapshot is a way to grab content from your live site. This replaces taking screenshots of things. So we'll just grab this section because I think this is a cool one to work on. And we'll paste it into Paper. And these are all live sections. Right? So this is not a screenshot. This is actual stuff that you can come in here and edit, and you can, like, alt drag and get another copy of it, all these things. So let's ask the agent to work on this little bit. And one of the things I like to do is actually just ask it for a bunch of variations. Mhmm. So we'll use con... Let's... Conductor. Hey, Conductor. This is using GPT 5.5. Mhmm. Can you make three new variations of my selection in paper focusing on the craft and alignment and contrast choices to elevate the visuals. You can do a lot of prompting, but I often find, like, simple prompting is is fine, honestly.
00:19:34 · Aaron Epstein, Stephen Haney
[00:19:34] S1: What's really powerful about this... A couple things. One, I love this concept of just spitting out a bunch of different variations. It's the best way to get inspiration. And sometimes, you know, you have a sense of what you want, but you don't really know. And this is a great way to look at a bunch of things and go, I like that from this one. I like this from this one, and combine it and and make so much more progress so much faster rather than being stuck in that creative rut. And then the other thing that's really cool is you're able to just say, hey. The thing that is my selection on the canvas Yeah. And paper, and it knows what that is. Yep. And now it's able to go to work on that, which is which is really powerful. [00:20:09] S0: I think I had the wrong selection, though. Grab the, the YC card. Let me let me restart it there. One more time using my selection. Oh, I see what happens. It it changed it, with the ADHD content. Oh, actually, know what, this is a good way to show another thing we can do. So we can leave a comment on this frame and say, please make three new variations of this content, explore the layout for visual impact. And instead, what I'll do, so I left a comment in Paper. [00:20:40] S1: Uh-huh.
00:20:40 · Aaron Epstein, Stephen Haney
[00:20:40] S1: Uh-huh. [00:20:41] S0: And I'll switch over to the agent now and say, please address my comments in Paper. There's only one. And this will be another way that you can kind of like target. And so you can imagine having a big design and leaving all of your feedback and then just setting your agent on it, it spin. But we see a lot of folks are leaving these running overnight. They're setting up loops of an agent just generating hundreds of variations. And it's actually a challenge for us to be, like, agent scale with our servers and everything. Like, they're they're doing a lot more work than I thought they were gonna be doing. And it's so cool. And they they... Now they come in and they curate. You know, they come in, oh, that's a good idea. That's a good idea. Let me combine these things together. Yep. It's kind of this new curation design process that I think is really interesting. [00:21:20] S1: It's so cool to think about this world where you just go in, you leave a bunch of comments on the work, you go to bed, you wake up in the morning, and they're all either addressed or lots of iterations for you to choose from are just sitting there waiting for you. [00:21:32] S0: Exactly. I think it's [00:21:33] S1: really cool. It's a new world.
00:21:33 · Aaron Epstein, Stephen Haney
[00:21:33] S1: really cool. It's a new world. [00:21:35] S0: I find that it'll do things that I know, but I wouldn't have thought of because I was too busy. Mhmm. And so it kind of, like, it lets me scale myself a little bit and have more time in my day because I can just set it on these tasks that are maybe more boilerplate, or I just don't have time to do four things at once. And so let me set the agents on it to to go address it. So it's... This is 5.5. It is working through ideas now, and you'll see it iterated. It'll... It it can take screenshots, so it'll notice things like that alignment isn't... It's running right into each other. It'll fix itself Mhmm. As it goes through. [00:22:03] S1: And Yeah. That's cool. Okay. So it's actually spit back a variation here, but the alignment's little screwed up there. And it's still working. So it's not done once it gives you something. It's it's an iterative thing that's happening here. Right? [00:22:17] S0: It checks its work. Yeah. And it can do... Agents are getting better at this too all the time. [00:22:21] S1: Mhmm.
00:22:21 · Aaron Epstein, Stephen Haney
[00:22:21] S1: Mhmm. [00:22:21] S0: For instance, like, see, like, the first shot, it ran the text over the cards. It'll usually catch that and correct it. Or if it doesn't, you can just, you know, tell it, hey. Correct the the text overflow. Mhmm. Let me do this. I'm set to fit, and we'll pop back in Conductor. Please address any text overflow issues. The other thing I noticed too is is the models are always getting better, they have different strengths and weaknesses. Mhmm. And so I I often use fast models for demos like this Mhmm. And they tend to make more mistakes. And so, you know, if you're doing this for real, you're probably using something that's right at the cutting edge, and those prove to deliver, you know, better performance. [00:22:57] S1: Yeah. What are your go to models these days?
00:22:57 · Aaron Epstein, Stephen Haney
[00:22:57] S1: Yeah. What are your go to models these days? [00:22:59] S0: So Fable is a wonderful... You know, I think that's probably the best visual taste right now. Yep. It's very slow. That's why I don't use it for demos. Yeah. I use Cursor as a Composer 2.5 Mhmm. Actually a lot. It's a very good model for the speed. Quality is very good, very precise. 5.5 is still usually a pretty good balance. So I jump around. I think if I need pure taste, it's the Fable and the anthropic models in general. If I'm going for precision or I have like, hey, have 100 artboards and I need a task to go down the entire list, I might reach for the composer. It it just does a really precise job. I find cursor is a really good harness. It... It's not too lazy. It'll often spin up sub agents to get jobs done. Yeah. So I jump around a bit. So here, it gave me kinda like three different ideas. You know? And I think from here, what I would usually do, spend a little more time on exactly what I actually want out of out of it and then give it more direction. But I might start branching. I might be like, oh, I like this. I like... You know, I kind of like the list of this. That's actually very easy to read. I have kind of like... The card layout of this actually, like, featuring one card and having some other cards to the side is nice. So I might just start branching. And think the power of this is you can ask five agents to go generate five more of each, you know, and and kind of find one that really fits