Transcript: ‘An Inside Look at Building an Email Client in Three Months’

The transcript of AI & I with Kieran Klaassen and Brandon Gell is below. Watch on X or YouTube, or listen on Spotify or Apple Podcasts.

Timestamps

1. Introduction: 00:01:56
2. How the maker of Cora describes the product: 00:02:33
3. Our first mistake while building Cora: 00:06:31
4. The story of how Kieran shipped the first MVP overnight: 00:09:37
5. Why Dan believes software is becoming content: 00:13:44
6. Products with a point of view will win: 00:16:40
7. How Kieran approaches building a new product: 00:19:16
8. Best practices while using Cursor: 00:31:55
9. Hacking together a copy editor in Cursor live on the show: 00:41:05
10.  The future of Cora, and the hardest challenge we face today: 00:53:58

Transcript

Dan Shipper (00:01:56)

Kieran and Brandon, welcome to the show. 

Brandon Gell (00:01:58)

Thank you.

Kieran Klaassen (00:01:59)

Thank you.

Dan Shipper (00:02:01)

Brandon, you're the top recurring guest on AI & I, so people should be familiar with who you are. But if they're not, you are the head of studio and consulting at Every. And Kieran, this is your first time on AI & I and you are the GM of Cora. Tell us a little bit about who you are and what Cora is.

Kieran Klaassen (00:02:28)

Yeah. Very excited to be here. I'm building Cora. It's a product that solves the pain we're all feeling when we think about the word email, hence my hat here. I wear a hat that says anti-email very largely printed on the front. We’re trying to solve a problem that everyone is feeling when they're thinking and working in email, which is dread, and no one feels like, oh yeah, let's open my email inbox with like 10,000 unread messages and let's go through. Or people that want stuff from me. Or random people that sent me stuff that I have to do now—today. And there has to be a better way. Dan and Brandon and I, when we met, we all kind of felt it. And we were like, why the hell did no one solve this problem yet? Why is Google not doing this? And we were playing with AI and we felt like there was something here. So we started working on Cora and it basically helps you reduce the stress you feel when you open your inbox because we take away 90 percent of your email from your inbox and brief it twice a day. So there are two moments in the day where you can read through everything that you need to know and read the bottom line of that and really focus on the important stuff. That's a very short version of it. We do way more, but I’d love to get into more details and hear other perspectives as well.

Dan Shipper (00:04:14)

Totally. And for context, we built this internally. We launched it as a waitlist product right before the holidays and it went really viral. I'm super proud of that launch. I'm super proud of the product that we have and we're just starting to onboard people right now and try to figure out how to make it work really well. And I just want to give people a sense of, what does the product actually look like? Because I think it's very cool. And I think it's a really good example of the new types of things that you can do with AI that were previously impossible a couple years ago. So this is the product. This is what my email inbox looks like. So basically all of the emails that I get that I don't really have to respond to. So you can think of newsletters or any kind of email that I get where someone is saying something like, here's like a status update or whatever. Instead of going and archiving those and reading them and archiving them one by one, what I can do with Cora is I can just scroll through and be like, okay, there's some partnership stuff going on with Every. I got an agreement from our lawyers, all that kind of stuff. And I can just sort of scroll through. I can see the money that we spent. I can see all my newsletters summarized. I scroll through it. Oh, I can see all the promotions I got. And then when I'm done, I'm done, and I don't have to press “archive.” I don't have to do anything like that. If any of these are things that I want to actually see in my email inbox, I can press “return to inbox.” I can click and read the full email if I want to. But, basically for me, it just saves a ton of time every day, going through my email. I think Kieran, you nailed it. It really deals with that dread feeling and, for me, it's sort of a pleasure to use. But I think what's really important for people to understand is Kieran, you built this end-to-end in three months, which is kind of crazy.

Kieran Klaassen (00:06:11)

Yeah. And it's really fun as well. It's the best thing. And yes, I built this in three months, but also we were feeling this idea already for four months and we were thinking about it. It's in the air. We're not the only email tool launching.

Brandon Gell (00:06:30)

Yeah, I think it's worth talking about the history of Cora for a second too because we actually didn't start here. We didn't start with briefs. We kind of started where everybody else is right now, which is an email assistant, which there are hundreds of products out there at this point that do that for you with all drafts of your emails. And we did that at first and we got really good drafts, but I think what we realized is it doesn't matter how good an LLM can mimic your voice, if it's not in your brain. So it doesn't have the context that you have, and it just can't write a good email for you—period—for 50 percent of your emails that you need to respond to. Because 50 percent of your emails that you need to respond to, you need additional context that the LLM doesn't have.

And so Cora does draft emails for you, but it only drafts emails that it thinks that it can do a great job drafting, where you're going to have to edit a very small amount. And I feel like we actually, throughout this process, realized that the actual cognitive load is the stress that you feel about managing your inbox, not necessarily responding to emails because responding to emails is kind of the pleasurable part of emailing because it means you're maybe progressing something forward. It's like managing and cleaning your inbox, which is actually the majority of what we all do that is really stressful. So I feel like, yeah, there's been like an interesting arc getting to where Cora is today.

Dan Shipper (00:08:05)

It has been an interesting arc and I think that we're starting to hit on something that I think is important for anyone who's thinking about building in AI right now, which is for a long time— We've always talked about solving problems in software, right? But for a long time, the most expensive thing was actually building the software. And that's changing dramatically. It's much, much cheaper now. And what that does is it makes the question of what you're building way more important because you can build anything in a couple days if you want, at least a really rough version. And so going on that journey of, okay, we have an idea, let's do drafting and then, oh, wait, drafting works, but, we're actually spending a lot of time reading and archiving emails that we don't need to respond to and maybe AI is better for that. That arc, in some ways, is getting more and more valuable to be able to do that than it is to write the underlying code. Because a lot of their underlying code can be written for you. But I am curious, Kieran, how were you able to make this so quickly? And take us into a little bit of your process. How much of it is just, you've founded companies before. You're a fantastic programmer. How much of it is that you're just a fantastic programmer and how much of it is you're using AI to make you faster at this?

Kieran Klaassen (00:09:30)

Yeah, it's a great question. I think it's everything together. it's really about really drilling down to understand what the problem is you're trying to solve. And at the same time being super free and just doing shit. We started with drafts and Brandon and I were on the call and we were like, oh yeah hey, hey, I have an idea. And we started suddenly the spark lit and we were like, oh yeah, we should do this and this and this. And I was buzzing in my head and I think I was walking back home and opened up my phone, just a memo recorder, and started speaking a large prompt. I was already like, okay, it's in my head. I have some idea of what we need to do. And from my experience as a programmer before I already was thinking about how I would implement this with one technology that I use. I would just talk it all out for probably 5–8 minutes, and convert it to text, and yeah, go from there. And that night I think I built an MVP and sent an email and said, yo, Brandon, here it is.

Dan Shipper (00:10:50)

We were both like, holy shit.

Brandon Gell (00:10:55)

I remember that. So Kieran, remember we got a phone call and I was catching up with Dan and he was like, how long do you think it's going to take him to do a V1? And I was, I think he is going to have something by the morning. And that's just not something that we could ever imagine a year ago, two years ago that you could possibly do that. And you did. And we woke up and we had a functioning V1 of Cora in a day. Kieran, I don't even know how you'll answer this, but how much of the code behind Cora, which is a huge product now—it does a lot—was written by AI, do you think? What percentage?

Kieran Klaassen (00:11:38)

It's a good question. I think everything is written by AI or maybe 80–90 percent, but 100 percent has been thought of by me. So I think that's how I think about it. It's like AI helps me, but it's really a collaborator just enabling me to do things faster. The tedious things are just faster. So I probably didn't think about everything, but who does most of it?

Brandon Gell (00:12:10)

Well, it's amazing. You're almost more of a writer now that knows that programming vs. a programmer.

Kieran Klaassen (00:12:16)

Absolutely For me, this is very natural. A little bit of background. I'm a musician. I studied composition. I did film music for eight years, conducting orchestras and scoring. The whole thing. And for me, software is very similar. What you try to do is tell a story like with music, you do that. You score a film or like a short with music to make you feel a certain way. And I think of it in a similar way with software, where you create an experience where it makes you feel or experience something or tell a story. And that's kind of different than maybe a problem solver, an engineer-y kind of perspective, which is another way to look at it, which is like, hey, I have, I'm going to program this thing. I always think big to small. And then I don't really care how it works. Clearly I care about beautiful codes and things that look good, but if it works, that's more important. And if it's solving the problem or telling a story or making you feel something, that's the most amazing thing. If you build something that excites you as a user, that's positive energy.

Dan Shipper (00:13:41)

That's been something we've been talking about a lot here at Every. In the Q1 kickoff that I did, we were talking about software becoming a little bit more like content. Because it's just much easier to write. And it also goes viral. A lot of our growth at Every has been us releasing software products like Cora over the last couple of months. And also writing is becoming a little bit more like coding because it's because you can actually build stuff with your writing. So there's this merging. And I think what you're pointing to is for a long time, we thought a lot about software engineering as about problem solving — solving really, really well defined problems, which is, I think, a great way to look at it. And we think about that all the time. Also, today we were debating what problem you're solving with a particular onboarding feature. And I think that's a really important lens. But I think being able to really crisply define and resolve a problem is in a lot of ways, a function of the fact that software has been really expensive to build and so you don't want to take a risk to build something that doesn't actually solve a problem for someone. And I think you're coming at it from a different direction, which I think is in general how we try to think about it at Every because it's really compatible with writing and just creativity in general, which is making something that makes someone feel a thing, which is a lot harder to define. And a lot of it has to come from sort of taste and vision and experience and creativity and all that kind of stuff. And I think software is moving in that direction more, which is another way in which it's a lot more like content now. Yeah, if anyone can build an email summarizing tool or like an email response tool or whatever, the one that makes you feel great is going to be the one that wins because everyone can solve the problem. And I think that's like a really interesting thing. And it's one of the most special things about Every. I think we're all trying to do that.

Kieran Klaassen (00:15:36)

Yeah, absolutely. And what is very interesting about that concept is it's more about good taste and style and just experience and what you take with you—your perspective. And especially with Cora, you see that everyone on the team likes to bring their vision and it's a different kind of solution to email. We hear people say, there's no one else like this. It's different and that can be good. That can be bad, but we're very proud of it. That and sometimes we have discussions like, oh, should we make it a little bit easier on the user here? Or make it a little bit smoother. And I'm like, eh, maybe it's okay to make a statement or do something a certain way, especially to stand out and that comes with that taste.

Brandon Gell (00:16:40)

Yeah, I feel like a lot of what we build—and a lot of what the most successful businesses are that I'm seeing right now that are being built using AI or leveraging AI—they come from people or organizations that have really strong perspectives. Sparkle is the same thing, and Cora—they're actually very similar. We're applying a perspective that we have, and it's not going to work for some people and it worked great for some other people. And it's a perspective that can only be applied by hand in a previous world. How I organize my files, that's something that takes a certain level of intelligence to be able to do. Only looking at my emails at 8 a.m. and 3 p.m.—that's something that you need to force yourself to do. So having a really, really strong perspective or methodology feels like for the first time you can actually force people to follow that methodology. Superhuman for email, kind of did this a little bit. I remember when I used it for the first time and instead of marking emails as read, you were marking them as done. And that felt like the big unlock where you're actually saying, I'm archiving this email, I don't want to see it in my inbox whatsoever. And I think it's just amazing how much more you can do now that you can apply an intelligent perspective through LLMs.

Dan Shipper (00:18:10)

Yeah, and I think that's also another overlap with writing. It's a writer's job to have a perspective. And that's another way in which I think product building is starting to overlap. I do want to go back though, Kieran, you said you had this great anecdote of building that first Cora MVP in one night. And you said your process is basically go take a walk, talk out a prompt, get home, transcribe the prompt and then get started. But, give us more details into how that works. Because selfishly, I'm like, how can I leverage this in my own process? Because, for example, one of my big questions is what is that prompt? Especially for a new project, what are you trying to get into that initial prompt as you're walking and talking that sets you up well?

Brandon Gell (00:19:02)

Can you find this prompt, Kieran?

Dan Shipper (00:19:05)

That would be great. I love that idea.

Kieran Klaassen (00:19:08)

So this is an example of something I was thinking of at some point walking somewhere. So normally what I do is when there's a new model coming out—I think this was o1 coming out—I want to build an app and I'll see how far I can push it, how far it will go, and where it breaks. It's just a good way to test it out. So I was a composer and I wanted to make a music app that had a specific thing where you click somewhere and there's a synthesizer and a sound generator and a mirror node, a whole complex compositional system. But really what I do in a moment like this is I just go walk. Walking enables me to just keep going, not thinking too much about it, getting more into a flow state, and it just starts building a prompt in my head. It's kind of weird. You're basically talking like it will become a prompt. And this is something I actually learned from you, Dan, in earlier episodes where you said you need to ground it and how to push the model into a certain direction.

So I start dropping like, oh yeah, you are a very good iOS engineer and you like Swift 18, you're amazing. Or just start with something like that. And it’s a start. And you start grounding it. And here it's like, I'm going to describe an app and its elements. This app is called “Theese,” if you open it, there's only one screen. Well, I just visualize the app here in my head and just do— I'm opening my phone and just imagining it and just talking through what happens. So maybe every corner has a different color, four-way gradients, tones of green and blue, a little bit of texture on it—like a grainy, noisy texture. So it looks a little bit fancy. Using words like “fancy” and maybe “Apple design,” just describing the feel. That's the background. Then in the middle, there's a separation. There's a line in the middle—somewhere in the middle. Actually, it's not a line, but it's like two parts of gradients. So you see, I was thinking about something. It was, oh, no, no, no. Scrap that. Yeah, let's go. So I just jam like that. Here you can see everything and until I don't know, and I try to just add details everywhere I can until my brain is empty. And then I stopped the recording and I don't take it anywhere directly from that. That's just step one.

Dan Shipper (00:21:59)

That's really interesting. So for people who are listening, basically what you did is you went and took a walk. You just rambled off the top of your head, but the way that you're rambling is— Actually, it's kind of wild how much detail you're able to give about what you want, which I think is really actually important and interesting. There's a lot of talk about how prompt engineering is dead or will die. And I think what's important to know is maybe I'll pay you 10,000 if you do give me a good answer and I'm going to die because the model should just give you a good answer no matter what. But being able to say, hey, I want an app that does this and I want the background to look like this and I want the style of the button on the left hand side to look like this is actually that's not going away. And that is prompt engineering. And that is really, really important being able to know that. And so it seems like that's what you're doing is you have a vision in your head of what you want, you're walking and talking in as much detail as possible, and you're not really worrying about whether or not you're getting it right as you say it because you just like to walk back and say, actually I don't want that, it should be like this. And then it sounds like you're transcribing. What was that screen you were showing me? What did you transcribe it in?

Kieran Klaassen (00:23:28)

So what I do is I use Voice Memos and then I put it into MacWhisper, which is a free Whisper converter. But actually now Voice Memos do transcription as well with the new iOS 18. So I use that once in a while as well. So yeah, I'll just use that. And then from there I go into my choice of LLM for the moment and start working on it more and converting this into a PRD. Most of the time I say, hey, okay, I have this idea. In Cursor, sometimes I add a notebook. And I create an outline of files or, depending on what it is, if it's simple and Claude is great for artifacts, but whatever I want to do after that, I take it from there, but it's a separate thing. That initial burst of inspiration, it doesn't matter where it goes.

Dan Shipper (00:24:30)

Well, let me get into detail on that. So basically you have the initial burst of inspiration. You write down the prompt, you transcribe it. What is your model of choice right now to make it into a PRD? And for people who are listening, what is PRD? What are the elements that make that effective for moving into coding?

Kieran Klaassen (00:24:50)

Yeah. I just always try to use the best model there is, the model that thinks the longest and the deepest. And currently that's o1 pro. but I have to say others do it great as well. So Claude Sonnet is great. There are no bad models anymore.

Dan Shipper (00:25:11)

Are you using o1 pro a lot? Do you like it?

Kieran Klaassen (00:25:13)

Oh yeah. I think of Cursor most of the time for coding, but o1 pro for everything else. It’s good. It’s slow.

Dan Shipper (00:25:28)

Like, planning coding? Or what else?

Kieran Klaassen (00:25:30)

Yeah. It's planning coding. But also thinking through strategies or just doing some collaboration that's not coding. Just a partner, bouncing ideas or interesting laying out options.

Dan Shipper (00:25:45)

I find myself using o1 all the time. That's my main model right now. And I have o1 pro, but it just takes so long. And I think sometimes it feels like it's a little galaxy brain. It overthinks and then maybe the quicker response from o1 is actually better. And so in certain cases, I haven't figured out— Basically, where I use o1 pro is a lot of the coding that I'm doing is sort of like you were, it's 90 percent AI-generated, but it's different from you in that I'm too lazy often to think about what the architecture should be, because I don't think that most of the stuff that I'm doing is not going to last beyond the couple hours when I'm doing it. And sometimes right now, a lot of the coding assistants—I use either Windsurf or Cursor Composer—they get stuck in loops and do weird stuff. And so when it's getting tangled up, I throw it into o1 pro and it usually figures it out, or I'll send the code to o1 pro and be like, is there anything in here that I should be aware of? Is it actually doing what I think it's doing, which is helpful? But other than that, I tend to just use o1. So it's interesting that you're using it for almost everything other than coding.

Kieran Klaassen (00:26:55)

Yeah, I use it for coding too, but very specifically, if I know what I want and just say do it and then it does it perfectly. But the interesting part about using it is actually that it takes so long. Yeah, you can say it's a bad thing, but what I use it for, if I use it, also makes me chill a little bit because working with AI can be very energizing, but also a little bit stressful because it goes so quickly and your brain, it really goes hard. And it's kind of refreshing to have to just sit on a chair, do nothing for three to five minutes once in a while. So I kind of embrace that.

Brandon Gell (00:27:40)

How often do you actually look at its thought process? 

Kieran Klaassen (00:27:42)

Never. Never.

Brandon Gell (00:27:44)

Because I would think that would also be helpful to look at that and be like, do I agree with the way that it's thinking and do I think the output will represent this. But I guess you can just go for it.

Kieran Klaassen (00:27:55)

So I never look at it because it's like a summary either way. Also, why I use it is because it is the newest model. I want to learn the most with it. And the only one to really learn about a model is to use it. That's how you get better is to use the model. So it might also be that where it's still a new model in my mind where I don't really understand or have a good feel for it. So that's why I use it as well—that's another reason.

Dan Shipper (00:28:24)

Brandon, what are you using these days?

Brandon Gell (00:28:26)

I use primarily o1. I actually can't use Claude right now because it's not available in Nicaragua. I have no idea why. I just can't wrap my head around why. So I'm like full OpenAI right now. And I just, honestly, use whatever it defaults to, which right now is o1. I don't do anywhere near as much coding as you guys, but I did use it to make a sort of a luxury personal app the other day and was blown away at how much better it is since the last time.

Dan Shipper (00:29:02)

What was the app?

Brandon Gell (00:29:04)

I'm still building products to help me learn Spanish better and I just can't seem to find a flashcard product that works well because I really feel like my writing and my reading is far surpassing my listening and my speaking. And I just want verbal flashcards. 

Dan Shipper (00:29:25)

What's a verbal flashcard? It speaks to you?

Brandon Gell (00:29:29)

I want it to say the word and I wanted to respond with Spanish or English. 

Dan Shipper (00:29:35)

Cool! Did you get that to work?