From Prompt to Product: How AI Is Disrupting SaaS

0:00Today is the worst AI will ever be. and AI will only improve from here. And I strongly believe that, right? You probably need somewhat of an engineering background and like somewhat of a prompt mastery background to build stuff. But I think like a year from now, two years from now, five years from now, like maybe you don't need an engineering background. Welcome to the Talking AI Podcast, where we talk AI with both experts in the field and early adopters. I'm your host, Matt Page, and we're here to demystify AI for you so you can get some value from it. Let's talk some AI. AI is going to change the current SaaS landscape in two major ways.

0:39The first one is creating new and novel ways of solving existing problems, existing jobs to be done. And the second one is just building solutions similar to existing products in a way more efficient way, leading to completely different cost structures, potentially entirely new business models. And this episode is really about the latter. And it all started with a tweet from someone that couldn't believe how expensive DocuSign was. And they're desperately looking for alternatives. I'm going to let our guest Michael Lew tell the story to all of you firsthand. But a bit of context, Michael is a smart dude.

1:13He was a previous software engineer at MetaZillow and sold his last company, now working at Stripe on the product side. But welcome to talking to you, Michael. Thank you. Well, I don't know if I would consider myself as smart, but I appreciate the compliment there. But yeah, like you mentioned, I saw a tweet from someone named Andrew Wilkinson that was about DocuSign specifically and how much they pay for that. I previously used DocuSign and, you know, I had a good experience, but I would agree that it's a little bit on the pricier side. And so that kind of inspired me. I was like, you know, DocuSign itself is not super hard to, you know, create a replica of.

1:49These signing tools are not super hard. There are compliance things that you do have to pay attention to. But as long as you pay attention to them and follow the letter of the law, it's not super difficult, at least in the US. Europe is a little bit different. And so that kind of inspired me. I was like, how easy or hard is it to create an e-signing tool with AI? So then I kicked off a very simple prompt in Lovable. I was like, help me build an e-signing tool. And it got in pretty much like two prompts. it got to the bare bones of what an e-signing tool could be it didn't have authentication or anything like that but you know it had upload a pdf and be able to you know sign it and then i basically took that as a very rough skeleton very not production ready right but then i took that into chat gpt and cursor and then added a bunch of features like authentication sending out emails and you know being able to you know get signing notifications things like that and made it production ready and now it's being used by thousands of users and it's all completely free and it cost me about 20 bucks a month to run which is amazing wow that's insane and full disclosure like docusons a great product you mentioned like the security and serving like those very high-end enterprise clients but this is just such a big shift in a democratization effect going on and i didn't realize that that you kind of like lovable a couple prompts and you got like good bare bones right there.

3:19It's funny too. So I did a TikTok video on this that went kind of viral, you know, quarter million views, 7 ,000 likes, a bunch of shares. But the best part was the comments. So this got a little spicy. So what I did, I like, I took all the comments, I dumped them in a chat, GBT. I was like, Hey, give me the highlights. Give me the interesting ones. And it was really interesting. There's like this polarizing effect of one group of people that were super pumped for like the old guard, you know, folks sitting on top of Sass Mountain, you know, to be challenged essentially. And there's phrases like Ren Appesance that somebody coined.

3:56And then there's folks on the other side completely dismissing it, talking about, like you mentioned, the security side, compliance, et cetera. And I don't know if you've heard of this term, but I don't know if this is a new thing, but they called the trend Timuware, essentially like building cheap knockoffs of real things. And then you had folks on the other side saying AI coding is a joke and will lead to a ton of technical debt and you get a bunch of AI slop. I think it's probably true on both sides of the aisle, but I'm curious, what's your take having gone through this experience? And you're obviously unique because you're an engineer by profession at some big time companies.

4:33So what's your thoughts on that? And I'd be curious, what reactions have you gotten and similar on, you know, social and whatnot. Yeah, yeah. I think there's truth on both sides, right? Like on the, let's call it like higher end, like enterprise sales side, like I think DocuSign, like obviously has a very robust product, but if I had to venture, yes, I don't know anything about DocuSign, right? But if I had to venture to guess, they probably have a much better enterprise sales motion, a much better enterprise brand. They probably have a lot of customer success slash customer support people, helping, you know, I don't know, like the big, like Nike's of the world.

5:12I'm just assuming Nike is DocuSign. I actually don't know. Right. And I don't imagine Nike using, you know, my app, which is like the spring times of the world. Right. Like it's very much like a fun side project that could become a business at some point, but I'm not trying to compete with DocuSign for the same types of customers. Right. And so I think there is like some truth and validity to that. on the like AI slop side, I would probably agree with that today, right? But I think I saw another tweet, which was like, today is the worst AI will ever be. And AI will only improve from here. And I strongly believe that, right?

5:52And I'll be the first to admit that, again, lovable, like what it gave me was like not production ready, but it was a good skeleton, right? Like there were a ton of like exploits. There were like, again, there's no authentication even. So like anyone could log in as anybody and defend like random contracts. You know, there's it didn't really follow like the compliance, like letter of the law, for example. It didn't give me audit logs until I explicitly prompted Cursor. I was like, hey, give me audit logs so that, you know, I am compliant with, you know, some USA laws. Right. And so I think today, like you, you probably need somewhat of an engineering background and like somewhat of a prompt mastery background to build stuff.

6:30but I think like a year from now, two years from now, five years from now, like maybe you don't need an engineering background. Maybe you don't need to be very good at prompting. And I think AI will only improve from here. Quick break in the pod. If you're listening to this podcast, chances are you've been thinking about how to actually use AI inside your business. And that's exactly why we built the AI Opportunity Finder. It's a free tool that helps you uncover high impact tailored AI use cases based on your business, your goals, your pain points, and your industry. No fluff, no generic use cases, just real ideas that fit your business and they're ranked by ROI potential.

7:06It takes about three minutes to run and it's like having your own personal AI strategist for free. If you want to try it for free, check out the link in the show notes or go to hatchworks.com backslash AI dash opportunity dash finder. Yeah, that's the interesting point too. And we kind of have our methodology. We're defining generative driven development and we had another guest on Andrew Miller, super smart guy. And he's kind of like dev turned designer turned product guy. So he's hit this, like the triad of roles. And he has a similar methodology called prompt driven development, but it is, that's the interesting piece.

7:42Like the folks that have background and context, because like going down into the syntax and all these things, but it's more about what, do you know the right questions to ask? Yeah. I think it's critical. and it's like those, I feel like it's that group of product folks that always wanted to build and they worked with development teams. They know like the ins and outs of things, but they never got deep into it. Like, I think that's going to be the most empowered group in this whole like shift because now they can build to their heart's desire and they're actually like really good at connecting like user to the, to what they're building, right?

8:15Yeah. And I think to add onto that, like maybe to be even more concrete, I think what these AI tools have product market fit for is like creating prototypes. And whether you believe or anyone believes that is slop or not slop, that's your opinion. I personally don't believe prototypes are slop because even though they might be like that code or whatever, they're like the first iteration of your product, right? And maybe you're just looking to get feedback and then you have more confidence to go build V2 or maybe you're building a prototype to raise the money so that you can go build V2 or whatever, right?

8:44And so I think like prototypes are valuable. And I do think that prototypes are like how we get to something that is more production ready, let's say two years down the road. Right. And I think just a natural stepping stone. And yes, I would generally agree that there is like some slop today. But I think the slop is it has to be a stepping stone to more production ready stuff down the road. Yeah. So I like I'll give you an example from my own side. I'm doing this workshop. It's actually tomorrow. But you know, when this goes out, it'll be in the past, but it's with a nonprofit local here in Atlanta called open hand.

9:19So they do meals for, you know, less fortunate or elderly things like that. So I saw Weight Watchers, they created this thing where you could take a picture of like your food, your meal, and it, you know, tell you how many points it is using AI. And I'm like, oh, that's pretty cool. And I started with chat GBT, where I just gave it a photo and said, Hey, give me the breakdown of the nutrition in this. And it was able to do it. And then I kept going a bit deeper. And then I found this API called is the nutrient IX or something completely free. And it actually has all this nutrition details. So it's like, okay, well maybe just for a demo in the workshop, let me just go through this in lovable and I did a very similar to you, like just a few lines and a, it came up with something beautiful to start.

10:01Yeah. And then I was able to just through prompting able to hook up with open AI on the API side. And then I integrated with that nutrition API. So it like you could do drop in a Pinterest recipe or take a picture of your food. AI would do this basically identify what it is. Then it would feed it to the nutrition IX thing. And, you know, just kind of a random thing. And I built it in, you know, a couple hours. And then I came across this tweet here for those on YouTube. There's this company called Cal AI. It's essentially the same concept. It's four people. It's now doing 12 million in ARR, right?

10:38So it's almost this shift, I think that's going to happen. And this person here, Ben Lang says, we've entered the era of tiny teams. The list keeps going. First time coming across some of these in, you know, it just gives you this list of several companies that are doing pretty insane numbers with, you know, anywhere from four to what's the biggest one on there? 16 people. Yeah. Right. What's your thoughts on that side? This concept of the tiny team. And then I think that's what gets interesting with like incumbents too, because the cost structure is just completely different, which could be a completely different business model that leads to as well.

11:12Yeah. I fully buy into tiny teams, right? Like maybe for some context that I build springtime, my free e-sign tool. It's just me. And I probably put in two to maybe four hours a week on building out, like specifically using prompts and building out features, like pretty much 99 % of my code is AI generating. Right. And I only put two to four hours and I can easily turn out anywhere from two to five features like this past week i only i was only able to turn out one feature but it was a pretty big feature right that was like a really big new feature that a bunch of users have been asking for and so i definitely believe that there is like tiny teams right because i'm almost like the product manager right talking to users figuring out what to build then i can directly translate that into like requirements aka a prompt and then so i'm now the software engineer right And then I'm also kind of like designer because now I have the solution in my head.

12:07Like I just need to go prompt it out, so to speak, what the user experience might look like. And so I think I do agree with Tiny Teams. And I think to piggyback on that, I think there will be a lot more people who don't just take on a predefined software engineering role or a predefined designer role. And I think roles will be like talking to customers and then designing it and then implementing it. And I think that will be like all one person. Right. And then I have some thoughts on the cost structure, but maybe I'll let you comment on that. Yeah. And we'll definitely get to the cost structure.

12:40I'm curious there, but like you're talking about the roles. What do you see the team of the future being? And hinted on it a little bit, but you know, it's almost don't need massive teams of engineers with specific specializations in every case where instead you may just have a really good generalist. paired with AI, they can do a lot more. I don't know. What do you think happens? Because there's a lot of people saying, oh, software engineering is dead. I don't believe that. I think there's a role for people to build things. I just think like the way you approach doing it is going to like foundationally change.

13:17Yeah, I think it's a good question. I do think there's like a subset of software engineers, you know, the ones working on groundbreaking stuff or, you know, NASA type of software that I think will be hard for AI to replace just because it does not have much training data on like NASA's code or like the future of frontier AI models. Right. But I do think that's like a subset, you know, maybe we'll count that in 10 % of all software engineers. And I think there's like, I don't know, like you look at it too. And this may be a side transition, but like AIDR, I don't know if you know that AI coding tool.

13:47I saw a stat today. It's like some large percent of the code being made for AIDR is actually being generated with AI. What was it? Gumroad, same thing. All their codes generated with AI. You had Claude, the CEO coming out saying, oh, and you know, it was six or 12 months. Everything's going to be generated by AI. But I guess that's more the function and less of the role though. Yeah. Yeah. And so I think it's, I think that's like the other 90 % of code. And I think, what's the right way to describe this? I think AI is a tool, but it doesn't directly replace like like a software engineer, right?

14:22And maybe I'll give you some like good examples. I think software engineers make business trade-off decisions, right? A good example is, hey, do we want to, like as we scale, do we want to, like are we okay spending X amount of time refactoring this code, rewriting it to be more performance, or should we just spend a bunch of extra money and just pay for a bunch of extra servers, right? And that's like very much a business decision that software engineers have like a lot of input into because they're the ones who understand the most to make a decision, right? And I feel AI as it is today, again, I think it could change like down the road, but I think AI has a harder time because it doesn't have as much business context, right?

15:03But I do think based off of the decision, let's assume that engineers, let's go down path A, then you can prompt the AI to be like, help me execute or write code or do whatever path A is, right? But I think software engineers still understand a lot of the business and technical trade-offs that it's hard today for a for an ai to like deeply understand and have all the context of yeah um with that being said i do think you know there as my code base gets a little bit bigger there start to become a little bit more bugs and i think i tweeted the other day there was a bug that like was really hard for my ai to so the ai wrote the code which had a bug and then i used ai to debug it and it was just like it didn't know how to debug it.

15:47And it kept writing code that was, that did not fix the issue. And as a software engineer, I was like, oh, I know, like, this is my process. I literally told it console.log and output everything that you would need to debug this. And then I, you know, updated the code. And then I, you know, manually ran through it, copied and pasted all the logs and then be like, now help me debug it. Right. And that was an example where the AI actually needed guidance from, you know, I like to think of myself as a more advanced software engineer than just an AI. And I think it needed some guidance, right? And so sometimes it can be really smart and be like a really senior engineer.

16:23And sometimes it can be really dumb and be, you know, it can be laser focused on what it thinks is the right solution. And sometimes it's actually wrong and it needs help like getting guidance. And so I think that was a great example. Now, obviously I think in the next six to 12 months, that'll improve a lot, but yeah. I think it's the example you said earlier, it's this is the worst it's going to be at this point. It's, you know, I tell my wife that all the time, this is as good to looking as I'm going to be, it's only going down from here. Right. But it's that same kind of thing. And you know, I want to get into this, like your current call it approach methodology for how you're interacting with AI and you hinted on it a little bit where you were kind of like prototyping with lovable.

17:06You mentioned some requirements with chat GBT, but I don't know if you've thought about this kind of thoughtfully, what are some of the key approaches, tactics, or things you do when you're working with AI, say the different phases? Were it early like ideation concepting to, hey, I got something that's real. I want to build new features. And you kind of gave the, you know, bug fixing example, but talk me through that, especially, you know, when you're going from a lovable to a cursor. And I think some people get confused there. It's, why am I using two different ones? You know, they kind of have their own unique benefits.

17:43Yeah, so I would say lovable is like really good for like front end web app stuff. And it's good to get started. Again, I don't think, at least from my experience, it wasn't great at like turning something like super production ready. And I had, I would then prompt it. I would then take that code and then prompt it in chat GPT or cursor to get it like super production ready. And then walk you through some example, right? Right. My second tip is I think O1 is like the best. So as from my experience and cursor doesn't support O1 with their$20 a month plan. And so a lot of times I would literally just copy and paste my code from cursor into chat GPT, which does have O1 with the$20 a month plan.

18:23So I pay for all of them. Right. And maybe it's a concrete example. I try to prompt and give as much detail with a very focused feature set. right so so and i try to give it as much context as i possibly can so as an example as an example there was a feature where p like today you can add maybe i can just share my screen and show yeah go for it let's get the for those on youtube will get a chance to check it out those on audio will kind of yeah verbalize so yeah so today you can upload a pdf like i'm doing like an important contract and so yeah before you couldn't drag and drop and reposition these fields right so yeah so obviously today you can but you couldn't drag and drop and reposition you'd have to click x and then you know re-add it and so obviously a feature request that i got was like hey i want to be able to drag and drop and reposition like i can here right and with my prompts to check gpt i would be like okay some context of what i'm building i am building an e-sign tool where someone can upload a PDF, yada, yada, add some fields.

19:27Then the second part of my prompt, I would say, this is the functionality today. So functionality today, people can add a signature field. They can also delete it. And then the third part, I would add the user problem. So the problem is people want to reorder the field. Maybe they accidentally place the signature field, or maybe they want to move it like five pixels down and deleting it and re-adding it isn't as good. And so, and then the fourth piece of the prompt was like the solution, which is help me build a drag and drop, you know, way to reposition these fields. And it's like very detailed, right?

20:04And I think the biggest call out here is I like sit down. This is like the human part, right? I sit down, I think through what the right user experience is. And I give the AI all the context of what I'm building, what the problem is, what the code looks like today and what the functionality is and what I want it to do and update the code to do. Right. And I think that is like the most helpful for the AI. And that's such a foundational thing. Like for folks listening, you need to internalize this. It's something you said a little bit ago was you go very deep on context, but very focused in terms of what you're building.

20:45Like you don't want to go say, build this entire big thing. It's very like focused in terms of the scope you're building. And context can be exactly what you just talked about, like problem solution, examples, screenshots, you know, code, all kinds of different things. And something I've done too, that I've found to be helpful is a lot of people will jump directly to, you know, go build the thing, even if they're providing a ton of contact. But what I'll do when I'm building something a lot of the times, especially if it's something new and not just a small iteration on something. I'll actually explicitly tell whatever ID or tool I'm in, do not write any code.

21:25Let's just have a conversation about this. And there's that conversation that you have is just beautiful context for AI to go do what you want it to do. And it can provide you feedback and thoughts. And that's the thing I like to tell people. It's like, there's this concept of known knowns, known unknowns, unknown unknowns. it can surface some of those unknown unknowns you're not even thinking about. But that's another way to iterate. Yeah. I a hundred percent. I a hundred percent agree with that. And I think two of my favorite questions to ask the AI before it writes any code. Yeah. Point is ask any clarifying questions to help you.

22:02And the second thing is state any assumptions that you're making. Right. And then I clearly can understand, okay, you are making this assumption about the code or the product or the user experience. And then, you know, maybe we have different assumptions and then maybe I can correct it on the next prompt. Yeah. Cause you know what AI is, I can't remember who came up with this analogy, but it's like the smartest intern you ever came across in the sense that like, they're very eager to do the thing and they don't want to ask you, Hey, but like, where is this? I'm just gonna go try and figure it out.

22:32It's very much like that. Like you almost have to give it permission to ask you this clarifying questions in a sense. Yeah. Let's go a bit deeper now. Do you, it's maybe kind of tangential, but do you play with like cursor rules at all or have you gotten into the mcp stuff anything like that or i guess just i'll pause there and then i got a couple other just curious questions yeah i mean i've played around with mcps a little bit i think my hot take is i think mcps are overhyped in the short term yeah by short term i mean like the next one to six months but i think they're underhyped long term like in the next two plus years.

23:11And I think the main reason is because I think like MCPs are just like a standard. And I think why they are overhyped is it's just like a very simple API. Like it's barely a tool, right? And like, you still have to tell it how to use the tool and like what to use the tool for. And that's the context that you give the prompt. And so I think like it will be powerful once AI or obviously once we have more MCPs and more tools to use, right? But once also AI and we get comfortable with telling the AI when to use the tool and how to use the tool. And then I think the third powerful unlock will be like once it just knows how to do everything automatically.

23:52And I think that will come once the context is really large and it can fit all of these MCP tools. You know, Stripe has an MCP tool, you know, shameless plug, you know, I don't know, like other MCP tools. What does the Stripe one do? I believe, it's a good question. I actually haven't played with the Stripe one, but I believe you can like do some payments. And I, at the very minimum, I know we have like docs. We've MCP'd our docs, I believe. And so you can kind of have access to that via like cursor or something. But I believe we're doing something around like identic payments too. Yeah. I'm looking it up right now.

24:25So yeah. Yeah. So what does it interact? Yeah. Interact with the Stripe APS. Yeah. Looks like there's some different tools. I'll give you a good example of one that I like. And it's a very, to your point, like it's very basic and simple. And it's, I think it's called browser tools, but to your point earlier with taking the console logs and whatnot, copy and pasting them into the chat, it basically is able to see your console log. So you're not having to do that very just minor copy paste step. So it's like a minor thing, but it's nice. It's nice to have in a sense. I think it'll evolve a lot, a lot there.

25:01Yeah. What, so I think the other piece too, I've noticed like when chat conversations get very long, like it'll start to, you know, hallucinate a little bit more. The context window is getting too long for it to work. So a lot of times, you know, you open a new chat. So any thoughts on like, when do you determine when to open a new chat? And I think the other piece too is like, we've talked a lot about context. Yeah. When do you actually explicitly say, okay, here are X files that relate to the thing we're going to do versus just allowing, you know, whatever ID you're using to just kind of know what files it needs to update and interact with.

25:44Yeah. So I, I scope my chat windows to basically be about a specific feature. Right. And I think the upside is the chats don't get liked. Honestly, most of my features can be done anywhere from one to five prompts. Right. I think that's like the upside where it's like very focused. The AI only has the context that it needs and nothing else. The downside is every time you open up the chat window, it like loses a lot of context. So I actually have a bank of prompts and like notion that like I can easily reference and like copy and paste based off of whatever feature. So if I need to reference, you know, this drag or drop feature, like I have a prompt bank, so to speak, kind of saved.