AI Does 90% of My Work Now as a Product Leader | Yana Welinder (Kraftful)

At this point I'm not actually doing anything without AI. The first thing that happens is I have social posts going out on like LinkedIn, Twitter thread, Bluesky and all of those are primarily written on Sunday afternoons using AI. And there's some aspects. I usually write them in ChatGPT. Either I have kind of a general idea that I prompt with or I rewrite them, but there's literally no post that goes out that hasn't been somehow written with AI. So that's one. I then had a customer call. I transferred describe all calls actually usually with the Zoom transcription and I sync them to Craftful to get analyzed. Yeah so kind of those are a few. I jumped in and did some emails which I always write with AI as well. I jumped into our support channel for a little bit which actually is not AI generated. Just we usually just chat there because I want to usually catch folks and find out what they're struggling with and ask them follow up questions. But I then what we talked about then syncs with AI to Graphful and gets analyzed. So there's always some component of AI, even if kind of not everything is AI enabled.

Cool. Welcome everyone. My guest today is Yana. Yana is the founder of Craftful, an AI assistant platform that helps product teams build products by extracting science from user reviews, surveys and more. And you know, Yana is also a power user of OpenAI's latest models and features, including Deep Research and Operator. So really excited to have her do a live demo of her top AI workflows and discuss the joys and frustrations of building AI products. Welcome.

Ioana, thanks so much for having me excited to chat about these topics.

So let's actually demo some of the stuff. Right? So like I recently paid the 200 bucks to use ChatGPT Pro and you know, I haven't had a chance to try Operator yet, but I've been playing around with Deep Research. But yeah, do you want to show us Operator first or Deep Research first? I'd love to learn from your so.

Let'S do Operator first just because that's what launched first. So maybe. Yeah, just a. Yeah. So let me show you some ways I use these tools. Yeah, so what I've done and you can kind of see it here, I sent Operator off to Craft, which is my product and then I have it select all the different sources of feedback. This is for Slack. I have it look at the feedback that came in about notifications and then it adds that to the roadmap. So now it's in the I'M going to put it in the notifications projects. It does a little bit of clicking around because that's what Operator does sometimes. And then it goes to the notifications project. It first filters now, next later, just looking at the now item, has all of these items here. Now it's going to create a PRD with all the insights about notifications and then it goes and reads all of those mentions. It then creates a user story with a item here and then it clicks around a little bit to find its way. Then it takes the user story, it syncs it to linear, so it takes the right project to put it into. And then it's gonna, once it syncs it there, it becomes a ticket in linear. You can kind of see this is what Operator does. It kind of tries to find its way around. Sometimes it like clicks around multiple times, which is interesting because it's trying to kind of understand how this works. And then it needs to sign into Linear to sync this ticket. It's now going to try to assign it to an engineer. But this is another one where it's like starting to have a look at these other things. Now it assigns it to an engineer and then it clicks back to see where all of the insights came from. So that's kind of, that's the whole flow where it takes all the feedback from these different sources about Slack, creates a prd, creates a user story. It syncs the user story to linear and then assigns it to imagine you're on linear. So it's sort of like an end to end use case.

This episode is brought to you by Merge. Product leaders cringe when they hear the word integration. They're not fun for you to build, launch or maintain. And they probably aren't what led you to product work in the first place. Luckily, the folks at Merge are obsessed with integrations. They built a single API that helps SaaS companies launch over 200 integrations in weeks, not quarters. Think of Merge like plaid, but for B2B SaaS companies like Ramp, Drata and Electric use Merge to access their customers accounting data for bill reconciliation, file storage data for searchable databases and HRIS data for auto provisioning access for their customers employees. If you need AI ready data for your SaaS product, then merge is also the fastest way to get it. So if you want to solve your company's integration dilemma once and for all, book a meeting and receive a $50Amazon gift card when you attend. That's merch.devpeteryang. now back to the episode.

Wow, what was your prompt to make this work.

Yeah. So my prompt was. There's. I kind of. I actually posted it in Twitter at some point. It has like 10 different steps. It's basically go to craftful, select all the items about notifications or actually select. Select all the different sources of feedback first and select all the items about notific. Then move those notifications items over to your notifications project there. Write the prd, look at the prd because otherwise it wouldn't look at it. And then write a user story about the top notifications item and then sync the user story to linear. So step by step to these things.

That's amazing. And you didn't give it any like websites or anything, right? Like URLs. You just kind of.

I did give it the URL, yeah.

Oh, you didn't give a URL. Okay, I got it.

It can also figure it out. But that is sort of annoying once I start searching for things.

So you just like wake up every morning and run this and be using this repeatedly?

This flow I have been using repeatedly. Not exactly like this, but in different configurations. Because one thing that I discovered after doing this is a few different things. One is this is great for creating product demos. So whenever we launch a new feature, I just have operator recorded. I don't record a feature launch demos anymore. The other reason I do this is to see just from like a usability perspective, if it clicks around too much, I kind of have a sense that, oh, here's something we may want to. And then. And then I sort of, you know, I. There's kind of this other. We did a bug bash earlier today where the whole team sat in a meeting and we were just like looking for bugs and actually had operators sit next to me and also click around in graph and do different things. This is not something I do every day. But like I was like, I'm just going to be looking for bugs. Let me have it look for bugs too.

That's actually a really good point. I didn't think about using this thing for usability testing or user research. I guess most of the cases, like trying to book a restaurant or something. But it sounds like actually the workplace use cases are even more relevant, right?

Yeah, because the more ambiguous the task, the more you get, like you don't want to give it too much instructions and then you'll get a sense for like, could it figure out the onboarding? Right. Like, was it harder or not?

Got it in a workplace. I mean, maybe not your company, but in a lot of companies there's a Lot of like, really mundane tasks that you have to do. Like, I, I actually tweeted this a couple days ago. Like, this product will be perfect for like watching corporate training videos. Right? Like the ones where like you have to actually sit on your screen, like, watch for. Watch the videos and like click something. Click next.

Yeah, yeah, I mean, I use it. I think I shared this with you, which is I. I use it to look for interesting content on social media that I may want to read and engage with. So that's another one where I like, give it a general sense of like, here's the topics that I'm interested in now. Go log into my LinkedIn. Log into. I also use it on YC Bookface. Log into YC Bookface and then look for posts about these general topics. I want to look at those. Tell me when you find interesting ones.

How does it tell you though? Like it just stops or how does it.

Yeah, it stops and it pings me in the browser. You get a notification. Yeah.

So you all sorts the. You've all sorts of dooms going into the computer.

Yes, yes. Which is wonderful.

Yeah, that's great. Yeah. So, yeah, I definitely think I should give this a try, poke around with it. So another thing that I launched is obviously Deep Research and a lot of people online saying, like, this is like the standout feature of the subscription. I'm curious to hear how you use it too.

Yeah, I use it quite a lot. I would say that I actually get much more usage out of the operator that I get out of Deep Research, which is sort of interesting, but that I would kind of disagree with that general opinion online. But I do think it's really useful. So some examples that I pulled up here from the things I used it for recently. So we were looking for this kind of API for. To create an agent in our app. So I described that that use case. I'm looking for an API that I could use to build an agent inside my app. The agents would allow my customers to specify a URL and the login credentials and have the agent collect content from that URL and essentially kind of analyze user feedback and craftful from things that we currently don't have integrations for. And then I asked it to put together a table with the name of the company, link to the website, product launch dates. I just want to kind of know how long has this thing been around for notable customers use cases and then if they are a YC company, we get a YC discount. So I want to know that. And then I came Back and it was yeah, great questions but actually have you thought about like, do you, is it, do you want it to be specifically focused on credential like things that require credentials and like authenticated websites or can it be anything? And then it asked me for like then do I care about any kind of like compliance things and things like that. So then I was like, oh yeah, those are great. Those are great things to include too in pricing and is it self hosted versus SaaS? So let's include that. Then it went off, it researched for eight minutes so not that long. And then it came back with like a really good overview of all the different companies, mostly YC companies that provide something like this. And it gave me all the content not in a table. I think it decided that this is too much information to put in the table, so formatted it this way. I had a look at this which I thought was great and then I was like, well but there's two other companies that I know of that it didn't include here probably because they're too recent. So then I asked it a follow up question which is go and research these two companies in particular. And then it had more follow up questions for me in terms of like what I would be interested in, which I responded to. And then it went off and researched for six minutes and gave me this really good overview of these two agent companies that would have taken me a really long time to research all of this like if I did it myself. Right. But this is like literally in less than 50 minutes. I have everything I need. Just really cool.

Yeah, yeah, it's really good for, yeah, it's obviously very good for market and competitive research or like you know, customer research. It's pretty brilliant. Yeah. And I've also found it pretty useful to like understand some technical concepts. Like I was trying to research different databases and how they work and it was pretty good. How about, have you had any other like personal professional use cases with it?

Yeah, yeah, yeah. So I've actually used this for a ton of different things. One, one really interesting use case, I got some medication recently and it came with this like giant label on from Walgreens saying you're going to get cancer from this. I don't think I've ever seen anything like this where it was like you didn't even have to read like all the like the small print. It was like really big. And so I was sort of like this is a little bit concerning. Let me research this. Had a, like a prompt asking about this versus I did a Little bit of background research to understand like what's some other similar type of medication. Then I had to go and do research and it turned out to be actually exactly opposite. This medication does not. This academic research showing that this medication does not cause cancer. There's another one that does, which I'd be curious if they have a similar modus online label. But it was kind of comical that it was just like such a big deal. And then I went in and I got it went in and summarized all of the kind of like academic research came back with like actual citations to papers and what they concluded in a couple of minutes. And I could go to that feeling okay about myself.

I haven't used O3 for this stuff, but like more deep research. But yeah, I asked like ChatGPT medical questions all the time. Look, for better, for worse.

Yeah, I mean for better, right? Because like as long as you get links back to academic research, then you can kind of know. Right. I think the for. For anyone who's had kids and is female, they may have read Expecting Better, which is kind of this book that goes through like all of the academic research on. On. On pregnancy and tells you like, like half the things that doctors tell you are like actually unnecessary or contradicted by research. And then it's kind of like that. But like for any use case that you may want. Right. Like it's wonderful. Like you can get all the information and you can kind of decide for yourself which is great.

Like yeah, it's.

Doctors don't have to be the gatekeepers.

I mean it's also like so hard to get in touch with a doctor. Like you have to make an appointment, you gotta wait around.

Exactly, exactly.

Yeah. I kind of like avoid a lot of like when I had my kids, like I used to argue with my wife about like, you know, you know, breastfeeding verheveress is just like feeding from using supplements and stuff. Yeah. And like, you know, if I had this tools I could just do my own research or you know what, I probably just listens to my wife.

But yeah, you could show the research to her and then you don't have to do anything. Right. Like she probably also be convinced.

And I also found it super useful for I'm trying to replace my windows in my house and like I got a bunch of quotes right. And sometimes like long PDFs with like a bunch of details. I don't know. It's kind of bad actually because it made me like really lazy. So like instead of just reading through all the PDFs I just like upload all them to ChatGPT and be like, hey, which one's giving me the best price? What was the best value? Which window is the best?

I do that with any kind of like contract or anything. But for quotes now you can use Operator to go and get those quotes for you. You can have it reach out to all these different like company providers. Yeah, get that.

Yeah, yeah, yeah.

Take even more time.

Yeah, I should try some of these use cases where like. Yeah. Should I just spend money or make a reservation? Let's see what kind of dinner reservation it makes for me.

Yeah, yeah. Apparently I've been having a lot of date night reservations that were all booked by operator in like the past year already. Something like that. Yeah.

Wow.

That my husband made. Yeah. So. So I can tell you they're good. We've had some nice meals.

He shouldn't have told you that. Obviously take credit for everything.

Yeah, I kind of guessed that's what happened. The moment I got, I started playing with Operator. I was like, I wonder we've had some really good date nights recently. I wonder if this. I asked him and he was like, yeah, that was the operator.

That's funny. That's funny. Actually, let's talk about your husband real quick. So, like, for people who don't know, like, who is your husband? Like, you know, and where does it work?

Yeah, yeah, yeah. So my husband is Peter Wellinger. He's VP of Product at OpenAI. So he's been working on the operator and ChatGPT and all the other cool products that I get to play with now and build, on top of which is even more fun.

So Isaac, do you guys talk about AI all the time? Because you're an AI founder and like you saw AI all the time at home.

I mean, definitely, like for a lot of folks that have woken up to AI and AGI now, and for us that's been like our dinner conversation for like 20 plus years. You know, it's like, and still continuous day. But it's a lot of fun to be kind of interested in the same stuff.

And I think we talk about this personally, but how are you, is your kid using AI or how are you getting a kid to use AI?

Yeah, so we're both builders and so I think we used AI very much to kind of teach our kids to think about how he could build and how he could build things in the future. And so kind of from the kind of, the earliest memory I have of this is back when they were first building Dolly and I had kind of like early access to Dolly and like, our kid was, I think he was like two or three at the time. We sat down and illustrated a book together. And like, by then he had, you know, he's listened to like, AI, generated bedtime stories and things like that. But there's like this moment when it was like very eye opening to him that I give the computer a prompt and it starts creating things for me. And I can now suddenly, like, make things. I can make a book, I can make the thing that, like, I would have seen, but I can make it with AI. And I think that was like, that was a big moment for him. And then a year later when they launched, or maybe it was, I don't know, two years later when they launched the GPT store, he saw me build a GPT and he was like, cool, I want to do that too. So he built this Pokemon GPT which now is like, I think the most used Pokemon GPT in the GPT store. And that's cool for him because it's sort of, you know, it's this moment of just like creating something for other people and then seeing other people interact with it. Like, just this thrill of knowing, okay, a thousand people have used the app I built. Right. Like, it's very easy to build a GPT, but still there's this like, thrill that he. I didn't get that at that age. Right. Like, but I think it is really powerful as a kid.

Yeah, I mean, making. My kid loves using Dall E to make images. She keeps making, like, rainbow unicorns and stuff. And then we've been making. We've been making, like, videos now. We've been using Google's VO to make videos. That was pretty fun.

Yeah, yeah, we've done a few of those too. That's. Yeah, that's so much fun.

Yeah, it's like. And then she brags to her friends that she makes all these videos and then they're like, what?

Yeah. Yes. Yeah, yeah. My son does that too. Last night we were looking at. So Infinity AI, there's video generation tool that you can then turn yourself into cartoon character and make videos with that character. And so we're gonna, we're gonna. We. We were looking at it together last night and he was like, can we make a video? I'm like, I need to, like, prep for launch. But, like, this weekend we will.

Infinity AI is how it's called.

Yeah, Infinity AI, it's a YC company.

You probably have, like, what are some, like, really underrated tools that you use AI tools other than like obviously ChatGPT or even Craft. What is Monterey Tools?

I don't know if they're underrated. I mean I think lovable, I love that. And I'm sort of curious if it's going to replace Figma for me now that they've added these like visual editing tools. I am a huge Figma user. Otherwise these kind of video generation tools, the ones that you can create like actual product demos with. So I mean heygen has gotten much better but also like Infinity and Oregon otherwise, you know, I've played with pretty much every, whenever like any tool comes out around like image generation or video generation I use in music generation. I will always play with them but I don't feel like they're lasting for me. I don't have necessarily like a good use case to use them on a daily basis. And so then they usually just like whenever something new comes out, I try it out just to like have it in like my mental database. Okay. I have a sense of like state of the art or kind of or how it compares to the state of the art and then I just move on to the next tool. So I don't necessarily kind of use them repeatedly, but I use them all.

That is actually the key problem I think with AI startups, right? Because like a lot of them have a novelty effect of like this is super cool and then you don't come back. Like you probably have a big retention problem.

Yes, I think so. I think so. But the thing is, I think for some of like for example, Heejin, when I went to play with it today to create this video, what I realized was that I haven't used it since early 2022 when they launched. So that, so I had a video sitting in there that I had recorded back then, which is when I tried it out the first time. And then I was like, this is not good enough, I can't actually do anything with it. So. But now it came back, right? So I think for some of them they do have immediate retention problems. But a lot of these folks then hopefully eventually come back and have more.

Well, I don't know if coming back after a year is enough for three years.

Coming back after three years.

Yeah, it's. Glad you came back. Yeah. Let's talk about craftful. Let's talk about, maybe we can talk a little about what graphl is and you know, what it does.

Yeah, yeah. So craftful takes easier feedback from all the different sources. So that can be internal sources like support tickets, call transcripts, Slack or external data like Reddit comments or posts on X Twitter or Amazon reviews, app reviews. It then crunches all that data, extracts feature requests, complaints, other kinds of product insights, and then turns those into PRDs, user stories that you can then sync to linear, like in the demo or to Jira to become a ticket. So it really kind of takes all the tedious tasks out of product work. So you can really just see what's important and focus on product craft and product strategy, which I think is really exciting.

Yeah. So what I've been doing with my PRDs, I mean, not that anyone reads PRDs anymore, but what I've been doing is that instead of like writing it by myself, I basically post a. Some sort of a prompt, basically in my community, like my customer community, I'm like, hey, you know, I'm thinking about building this. What do you think about it? Or like, you know, do you have a problem? Like what problem? We're trying to solve around this and there's like a bunch of feedback that comes in.

Yes.

And then, and then, you know, don't tell my customers. But I don't really. A lot of times I don't actually read every piece of feedback. I just copy it into AI and try to get like, hey, totally. Can you summarize it? Or like, can you actually make a PRD based on this, like template that I have?

So I mean, we kind of do that, but by default, right? Because we are taking all of the, we take all of the insights, we collect all those insights for you by default. But it's much broader, right, because it's going to be when people are actually using your product and when they've had a pain point about a similar thing. We will have taken all of those instances and then our PRD reader goes out and reads like hundreds of comments from different, like call transcripts, support tickets, all of these survey data. And then you can, yes, you can put in them your general, like, here, write the PRD in this manner. You can do that in the prompt piece, but the collection piece happens automatically. So you kind of don't have to, you don't have to ask for new input, which you could use our survey product to go and collect more data.

I don't have to keep copy and pasting everything, basically.

No, you can't. No, you kind of have to use, you have to use Graph as the conduit for data collection.

And it's more so. So that's more of an agent than a copilot. Right. Because it's like collecting all the data for you automatically.

Totally. Exactly. Got it.

Got it.

And it automatically analyzes it for you.

Got it.

And it automatically turns it into PRDs and user stories. Right. So there's like all of these different steps. You can come in and press the button and say this one, turn this one into a user story and then Raffle does all of that.

Wait, but Yan, so then what does the PM actually do?

Like relax on the sofa? No, I'm just joking. I do think that, you know what I. This is my own personal pain point, right? Like I used to be a pm before starting Craftle, I've been at various tech companies. Everything from like a PM number two at a unit, like a fast growing unicorn to actually leading products, millions of users and having to being kind of way past like product market fit. And I think that the thing that I always wished that like me and my PM could do is to just sense what we know about the world and our usage. Right. Because you can also use Krafl for competitive research and then be able to like think about it deeply. Right? Because then you can do things with that data. Like if you don't have to spend the time figuring out the data and the data is like there in front of you. Because I think one piece, one thing about a lot of PMs is that they don't spend the time thinking about or trying to find it because they just don't. So they just don't have access to it at all. Right. Like this by default makes sure that all the insights are at the core of the development process and then you can spend the time figuring out what, like what do we know based on this? What's the best way to solve the problem and is there maybe something else that we don't know that we should go out and try to like collect more data on them, then create a survey and gather more insights. So there's sort of, you know, it just makes sure that you spend more time thinking deeply about what the problem is as opposed to just thinking deeply.

I mean, I mean, yeah, there's like a lot of busy work to gathering and synthesize the data that Craftman helps with. And also a lot of busy work to produce these like internal artifacts. Right? Like, like these tickets and all stuff. Yeah, yeah.

And the other thing I would say is that there's also a lot of busy work in do having a bunch of back and forth communication. Right. If you can take all the insights, have it, write a user story about all those insights, think that as a ticket and have all of the insights attached to the ticket because that's what they are. Then the engineers have access to the why and then they have access to all of the related feedback so that if they have some optionality that came up during the process, they don't have to hunt down the pm, have a bunch of back and forth. Maybe the PM needs to go out and do more research or find the researcher and do more research or whatever. Just make sure that engineers have access to everything. They technically don't have to bug the PM either.

Yeah, got it, got it. Yeah, that makes sense. That makes sense. Like, you've been at it for a couple years, right? So like, can you share some stats about, you know, how many customers you have or how fast it's growing now?

Yeah. Yes. Now we serve over 50,000 product teams and product builders. We've analyzed many millions of pieces of feedback for folks which we translate into many hundreds of thousands of hours saved. But yeah, it's been a wild ride, which we are about to hit two years in, in about, in. In about a week.

And what have been some kind of like really great moments. And also, you know, like, like a founder journey has a lot of ups and downs. Right. Can you share some stories?

Yeah, totally. So today we actually had a high points for graphful. We had, we've had, for a few weeks we've had exceptionally high MRR growth, which is like, even by like yc. We're a YC company. And even by like YC company standards, we're in what YC considers to be exceptionally high growth, which has been really cool to see. And then this week we broke a record in our own exceptionally high growth and have even more exceptionally high MRR growth. So that was just like a very cool moment for the team. Cool, cool. Way to celebrate Valentine's Day. But you know, I think we've had a bunch of. We have had a bunch of low moments. One low moment was when we first launched and that was really exciting, kind of early 2023. And then we had so much adoption that it broke our products. And we kind of had this happen repeatedly. So it would kind of break. We would break past every API rate limit and things like that. And so for most part, this is like exciting or just like we break through an API rate limit. I would have to like hunt down the companies like they're actual. Usually try to get the founders to get like really big rate limit extensions that I needed like in like half an hour. And that was all cool until suddenly we broke the absolute rate limit extend like rate limits of GPT4. And we could not like the amount of data we were sending in for our customers to analyze was more than you could analyze with GPT4 at the time. And we could not get any more extensions. That was like one of those moments where I was just like. I was just like losing so much sleep. It was actually really painful. I could see these big brand customers sign up and just. I could see how long they were waiting for analysis. Right. Because we just had too much and we could not. And some of them would wait for days, which is normally to take minutes. And it was just really painful process. We had to rebuild our entire architecture at that time using kind of less performant models and do lots of fine tuning, rethink our entire pipeline, do all kinds of different things. And then eventually COMPUTE became more available and we could go back to using GPT4 for all the things and kind of things changed rapidly. But there was like this moment in time that was incredibly painful, which you would think that it's actually like all great things, but actually. But it on the inside it was just terrible.

Because you want to let your customers down, right?

Yeah. And they were just like so disappointed. They were like craftful. Just like it's really slow. I don't understand why it's so slow. Right. Kind of kept debating, do we tell them that there's too many of them? Like, you know, are they like.

Yeah, I think, I mean people say like the model is becoming a commodity, but like the model does really matter for like the quality of the product that's built on top of it. So yeah, if the downgrade from like 4 to 3.5, like that probably impacts the quality by a lot.

It was huge. It was really like, it was really impact. We had to kind of change. We had to create lots of additional steps in the process. Well, our analysis is like 100 different steps to get to from like one support tickets to a feature request already as it is. But we had to add more steps and we had to fine tune the model to do things to be as close to GPT4 as we could get and things like that. It's just really complicated.

How do you think about. Let's talk about prompting reg and fine tuning. I'm just curious, as someone's building a successful AI product, how much time do you spend on this stuff? In some parts, sometimes I'm just updating the prompt again and it kind of feels dumb, but it does have a big impact. Right. So I don't know, how do you balance? And also running evals, which is also really painful.

Yeah, it's all really painful. Particularly when you have something like that pipeline that I described, right? There's like a hundred steps, which. Not 100 steps, but there's a lot of steps. And each step has a different prompt and each step uses a different model. And then we're running evals whenever, like a new model consultation, we're running evals to see can this model, can the new model do any of our steps better than the existing model we're using for that particular step? Can it do it better across few different things, right? Like, can it do better analysis? Can it do it faster? Can it do it cheaper? Right? So always looking at like different things and some things are like, oh, it's actually not better, but it's cheaper or it's faster, right? So then you're, oh, maybe then maybe we should invest in like doing the upgrade for this, this one.

So like, do you have like a list of tips or like hard lessons that you learned building AI products? Like, is it like trying to keep as simple as possible? Or like start the prompt first before it getting more complicated? Like, do you have any tips as well? Who wants to build this stuff?

Yeah, I mean, a ton, right? Like, definitely starts with the prompt first. I was recently thinking about this in particular, that like, when we design something, we first start by designing it in like OpenAI playground, right? Like, it's not designed in Sigma, it's designed. We're designing what is going to be the experience, thinking through all the different steps and then the output of that we then know, okay, this is best represented as a word, or this is best represented as a list, or this is best represented as a longer sentence or a paragraph. Then the UI ends up being very different based off of that. So always starting with, from what can the AI do to best solve the problem? And then like, how do we then design the experience around that in the most compelling way? So that's definitely it. And then the other piece is you're always kind of looking at your options. You're like, okay, if this, if the AI can solve the problem, what can I do? Can I like, fine tune, can I wait for the next model? Can I change the user experience? Right? That's always the option, right? Like, we have different types of things. Like, here we are waiting for the next model. That's what we're doing, we're waiting for the next model. And you can change the experience, right? Because there's certain things where you can kind of make it clear that this is, this can be edited or this can be changed. Right. So that even if like the AI isn't perfect, then you're giving the human control to make it perfect in their flow. As long as. Because some things are like if you're saving someone a hundred hours, if they have to invest a minute and making it perfect, that's great. Right? And so just thinking through what's the whole workflow and then how do I create a really good product around that workflow based on what models can do today?

That's pretty funny. Yeah. It's not like the old days where you have a multi quarter roadmap. Now maybe you have a bunch of initiatives and some are waiting for the model to be better. Some are like, yeah, I'm actually curious, how do you even do roadmaps? All right.

No, I mean we do, we do, we use the like our projects page, right? The one that you kind of saw in the operator and that does have like initiatives that are filtered by like now, next, later, but it is. And then we kind of have some that are like generated by what we know that's outside of what we've learned from Insights. And then we like run surveys to fill out those. But yeah, I mean that's our roadmap. But I think I like not in a dock somewhere.

Like the time horizon is like maybe like three months or something like a quarter.

Like the planning, it's probably, I mean in our, it's probably six months this time horizon. There are some things that, where people really want this and there's enough people who really want this that are impactful enough that we want to build this. But like it doesn't also doesn't make sense for us and like the, in the next few months or we know like certainly, you know, like certain new models have launched and they can, they suddenly can unlock really exciting things. We want to do these things first before we get to those type of things, you know?

Yeah, yeah. I mentioned to another founder, it's great to have like, you know, like OpenAI and these companies spending billions of dollars basically making your product better. Right?

Yes, it's wonderful.

It's kind of nice.

You know, it is incredible. Like you can sort of constantly, I don't know, I'm like people telling me that they're still like they're so daunted by the fact that there's so many like new launches all the time. Like, are you kidding me? Like, I'm just like whenever something new launches I'm like, what is it going to do for my product today? You know?

Yeah, it's just like a one line change. Right. It's like not too hard to change the model. Yeah.

So yeah, I mean it's not because we have to run emails for all of it the way you feel like. So it's actually very complicated, but still, it's always worth it.

Yeah. Part of me, I've been trying to build Air products too and like, part of me is kind of annoyed at how manual or iterative some of this stuff is. And you can never feel as sure as you can with a deterministic product determines that you make the UI pixel perfect and you're like, okay, this is a good experience. And with this stuff, very hard to say for certain. Right. That everyone's going to get a good experience. Do you have any closing words of advice after 10 about this stuff?

Yeah, I think actually to your point about was it a good experience or not, I would say that the thing that I found the most kind of interesting is that a lot of experiences, AI experiences today are built like a chat with a chat interface. And with a chat interface it's really hard to know is it a good experience or not, because it really all depends on whatever the person ends up putting into the chat. We have an Ask AI part of our product, so it's a very small part of our product. We do have that. And then most of the product is, as you mentioned, kind of more of an agent that just goes out and does these things for you. And those pieces are actually incredibly like they, while they're not built on a deterministic model, they're still, because of the way they're built, you have a very good sense of how they're going to operate. They feed in a bunch of insights, but they, you've chunked it into small pieces to the point where you have a really good control over. And then you run emails and you have a really good control over it. And so I think that my biggest would be that it's, I understand why folks are building like chat interfaces because they look at ChatGPT as like the fastest growing product in the history of products. They want to have that success. But there is a whole different other way to build AI products in like new and different ways that I think if, you know, as a, as a person building with AI, I would really encourage folks to look into more.

That's a really good point. Yeah, because like ChatGPT is like a very horizontal product. Like, like you said it depends on what the prompt is. But like, you want to be 10x better at ChatGPT on a specific use case, right? That's like less horizontal.

Yeah, yeah, yeah. And you can do that. And it makes sense. Like for general purpose product, it makes sense to have a chat interface. But for a very specific product, that may not be the best solution. But lots and lots. Most AI products out there are kind of like a chat interface, which is not what they should be, in my opinion.

Okay, that's a really good point. Yeah. So where can people find out about Craftal or give it a try or fall follow you?

Yeah, so you can kind of follow me. I'm most most active on Twitter, so Yana tweets and craftful is just craftful.com with a K. So feel free to check it out. We do have a free plan that folks can explore with.

Awesome. Yana. Yeah, thanks so much for your time. It's been great to chat with you through the past year and see you grow the company. It's been awesome.

Thank you. Thanks so much for having me. I really enjoyed it.

All right.

Sa.