Prompt, Build, Repeat: The New Software Workflow

0:00The prompt-driven development framework is basically me trying to take all of the best practices that I've observed and help build out with my teams and squashed that down into what would be a teaming that I'm doing. any wheels. I'm really just trying to kind of distill all the best practices down into a framework so that people can build software with LLMs without having to code in a more kind of reliable, methodical manner. 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.

0:47Let's talk some AI.

0:51AI is completely changing how software is built. And it isn't just how it's built. It's the whole day process, soup to nuts, start to finish, all the things. And here to discuss that with me is Andrew Miller, engineer turned designer, turned product manager, who has developed this methodology on how to build with AI called prompt driven development, PDD for short. But Andrew, welcome to the show. Hey, thanks for having me. Glad to be here. Yeah. And those for that are not watching on YouTube, I swear, Andrew, you are a Christian Bale doppelganger. I don't know if you've ever gotten that before, but you're his double.

1:27So this AI stuff doesn't work out. You got a career in it as being his own. I'm actually a Christian Bale deepfake. Don't tell anyone, but you know, I knew it. So we're actually talking to the AI version of Christian Bale. He's developing ways to build more Christian Bales. It's so funny. It's a matrix part five, I think. But kick us off here. I think the context is super important of how you got here because your journey, like I mentioned, developer, engineer to product kind of gives you this unique lens on this like massive change that AI is driving. So talk us through that. And like, where was that aha moment where you were like, holy crap, this thing can do some really powerful stuff.

2:10Yeah. Yeah. I mean, that whole arc that, that you're referring to really just my own career arc started in the development seat. I've always kind of had one foot in development, one foot in design. Even back in college, I was taking interaction design courses and software development courses at the same time. And I was just really drawn to the idea of building things that could be used on the internet because the scale was blowing my mind. And right out of college, I started a startup. Yep. So kind of right off the bat, kind of threw myself into the deep end of the whole sort of entrepreneurial software startup world.

2:52Didn't know I was sort of getting some product experience at the time. That startup failed in a few years. It was a cool ride and then landed a job as a dev, small team, small startup of 10 people. And it was really in that context that I was able to transitioned into design because we had zero designers. I was there for a year and just, I noticed over and over again, design was a big bottleneck. So I kind of petitioned to the founders like, Hey, you know, I can take this on. And so I became the first designer there and built out a design team. It was a really great experience. That's cool. And then the transition to product was at the same company actually with the same people, the same founders, great guys.

3:33I'm still working with one of them actually. And moving from designer to PM was basically this sort of organic process of going from the idea of a feature all the way to taking it to market as a designer. And as the company grew, the need for that specified role of product manager started to grow. And so that's how I transitioned into PM and ended up being really happy there. That's where I've been since and I've bounced around a couple different companies and stuff. But yeah, I like it in product. I think with, along with many other people, I got really excited when ChatGPT came out and it's just, wow, this is total magic.

4:14And so I've been absolutely obsessed for a few years since, since that launch, you know? And that's what it was like, complete crap too. Cause like I was just talking, doing a podcast with Nate Jones. I don't know which one will come out first, but it was the same kind of thing. he mentioned like in the beginning it was really cool but it also kind of sucked at the same time and there was a lot of people i think that either had your reaction like oh my god this is magic they saw the potential for it and i think there was a group of other other people that saw it as oh this is kind of a neat thing and then they just forgot about yeah yeah i mean that's that's fair it's i was on the holy crap this is amazing side of things so i i didn't really look the other direction too much.

4:59But yeah, there's definitely an adoption curve there, you know. 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 the ranked by ROI potential. It takes about three minutes to run and it's like having your own personal AI strategist for free.

5:31If 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. And at this point it's not building. It's not really helping with code. I mean, I'm assuming the first time you used it, But did your mind immediately go to, this can be used to write code? Or at what point did it get to there? No, it did not. I think it was capable then, but I wasn't really using it for that reason at the time. I think the big unlock for me in terms of using LLMs to generate code is I started tinkering around with Crew AI, which is a framework for developing agents.

6:13It's become really popular. That founder is doing amazing work. The whole team's doing amazing work. And so there was like a little meetup of some friends here in town in Asheville where I live. And a friend of mine was showing everyone how to use Crew AI. And there was like seven of us in the room, right? And so we're going through the process that he set up and I'm installing Crew AI. And then I get to the part where I define what the agent is supposed to do. And I just write it out in English. And then I hit run. And then it did it. And it was such a funny moment because it's me, the dev gone PM.

6:53I'm the least technical person in the room, right? Everyone else is like super senior devs. And I literally stand up out of my chair and put my hands up. I'm like, guys, do you realize how freaking crazy this is? And they're like, yeah, you just write English and it does what you say. And I'm like, no, you're not appreciating this. Yeah. It just blew my mind. And then ever since that moment, which was probably like six, seven months ago, I've been pretty obsessed with building with LLMs. Yeah. And it's, it's, it's crazy to think about that. I kind of remember back to the same moment. I feel like it was, maybe it was Replit or some tool like that.

7:36And it's just like this mind blowing thing. And even the stuff that, that, that it did early on days was not great, but you can kind of sense the capability of it and where it could go. But what I love about your framework is it's, it's both like simple and very logical. And it's just, after you read it, you're just like, well, duh, obviously we should work in this manner. So like take, take us through kind of high level prompt driven development. How do you think about it? And what are these like patterns that you've seen start to emerge with how you should interact with AI because it really is like a collaborative experience.

8:18Yeah, that's a huge compliment that it's easy to understand and feels intuitive. I appreciate that. I think, you know, I've been building software for at this point, you know, a decade and a half. I've worked with a bunch of teams, been in a few different roles, kind of seen the whole every major side of software development. And for me, the prompt-driven development framework is basically me trying to take all of the best practices that I've observed and help build out with my teams and squash that down into what would it look like if one person, i.e. me, is trying to create that entire process just by working with an LLM.

8:59So I think I'm really not trying to reinvent any wheels. I'm really just trying to kind of distill all the best practices down into a framework so that people can build software with LLMs without having to code in a more kind of reliable, methodical manner. And I want to contrast it with a lot of examples or demos that I see online, which is I lovingly call like zero shot app development. It's kind of like your big prompt that you put it in, click go, and it builds the whole product for you. And there are tools that can do that. And I think that's really interesting. But I like to stress that it's not a great way to learn, right?

9:42And it's not a great way to build software in a manner that allows you to continue maintaining and developing that code in a reliable way. And so PDD is really intended to, you can build sophisticated applications without having to write the code, but you're keeping all the good stuff, right? The code review, the breaking work down into small scopes, small increments that you can understand one at a time and so on. And there's a lot more I can dig into. Yeah, and maybe we talk through the workflow because the example you gave is perfect. There are tons of cool tools out there that can like zero shot and do something really cool.

10:25But what I've experienced is if you don't start in the beginning with building the foundation and aligning, literally the way I liken it is like you're working with another human. You need to align with the human on, hey, here's the thing I want to build. Okay, let's define the requirements together. And then the back and forth there, I find to be really interesting. And then also like AI, I don't know if you noticed this, but AI will bring up points and be like, oh, did you think of this? And it's like making light of these unknown unknowns that we didn't know we should be thinking about and then defining a plan with AI and then, you know, the architecture and all these things.

11:05And you have this foundation. And another part I love about PDD is working in small increments. But to your point, like first principles, that's, that's the principle of agile at the end of the day. Right. Yep. Yeah. Take us through the actual flow, though. I love the visuals you have where it's the old school receive requirements. You kind of iterate locally, submit for review, merge. What's the element that's different here when you start adding LLOs into the mix? Let me actually pull that up real quick. Yeah, let's do it. Because I do have that. And if you're just listening on Apple, Spotify will be very descriptive with how we talk about this.

11:46there, but it's also on YouTube if you want to go actually watch. Yeah. So I think the easiest way to quickly understand PDD is just contrasting it with traditional software development. So this visual is a very high level workflow for what that looks like, traditional software development. Typically developer receives some requirements that they hopefully were involved in generating as well. And then they go off and they iterate within their IDE locally, write some code, test it out, see how it's working, tinkering around with it, get the code to a spot where they're happy with it, and then they submit it for a review.

12:22That either comes back with some feedback or it gets accepted by the team and it gets merged. Pretty straightforward. That's how most software development works. If we were to contrast that with PDD, there's a couple different steps. So we start the same way, receive some requirements, and then the next thing we would do is new. We would break those down into prompts. So instead of going to code, we're thinking in the context of prompts. So I'm trying to build a feature that does X. How would I go about building that feature? What are the small incremental steps that I would take to scaffold that feature up and to make it reality?

13:02That's the type of thinking you're doing when you're breaking it down into prompts. And then from there, when you're happy with your prompts, you would literally paste those into an LLM one at a time to get the necessary code written. What I like to use is an LLM native IDE. My preference is cursor. There's a bunch out there, Windsurf, Lovable, Bolt, V0, and so on. I'm a cursor person, so I'm pasting the prompts into cursor's composer feature. And then I'm iterating with cursor, right? It writes some code. I test it out locally. I read the code, review it, make sure I generally understand what's going on.

13:44And then when I get it to a spot where I'm happy with it, I would send it off for review, commit to the repository and so on and move on. One thing that's really important I want to... I think that's a... Yeah, go ahead. That's such an important piece to boss over it there. But the human in the loop piece of that feedback loop is so critical, right? Cause it gets very easy to, when you're working with cursor would name your ID to be like, except, except, except, but you do kind of got to at least be aware of what it's changing. Does it make sense? And then you also have context for the path that's gone down and you're walking it together versus oh shit, something breaks.

14:25And then you're like feverishly trying to go through the chat log, trying to say, okay, what did you do? I think, I think the human in the loop part is, is exactly right. Okay. I was, I was trying to go to this slide. So this is one of the mental models that I tell people. I think AI should only write as much code as you can understand at one time. And so part of the challenge is we have these tools that are actually capable of going off on their own and generating a lot of code. They can write multiple files at once. They can build entire products on their own and then come back and say, hey, the whole job is done.

15:03but does that mean that you understand it? And so keeping the distance between the changes that were made in your own understanding as narrow as possible is a big part of the challenge. And so I don't think the goal should be get the AI to do as much work as possible. I think the goal should really be get my brain and the AI to move together as fast as possible. And it shouldn't get too far ahead of me because if it does, then I'm, I'm toast. Right. Yeah. Yeah, exactly. Exactly. What about two of the pieces too? Cause this is like the perfect, like steady state. I'm in the rhythm. I'm like working with AI, all that.

15:45I'd actually, by the way, if y 'all don't follow Andrew, go find Andrew ships on TikTok, great, great TikTok and a sub stack. I saw your video right before we jumped on and you were talking about. Like the design aspect of things. It also makes me think of discovery in a traditional sense. So it's like that upfront work that you do before you're like in the heart of building something. Where do you see that changing? Because we have thinkings on this of how we operate and where we think team of the future is going to be. We have a similar framework, GenDD, that we practice. And we're kind of getting to this.

16:21Like this is my vision for it. I'm curious your thing. I almost see in the future where you have two core functions. It's like the agentic engineer and the agentic call them product strategist, product manager. And those are like the two core roles. The product person like is representing kind of the user, the agentic engineer as a great foundation and architecture, all these different things. Right. But then they're leveraging AI to build from there. So what kind of goes away? BA, I think like AI can supplement that and the product person is kind of guiding it. the architect, the AA person, all these different roles.

16:59And, you know, for large scale projects, probably different, but I think like in the future, it's like those two humans moving to orchestrators, leveraging AI can do some really powerful stuff. So I'm curious, like, A, how do you think of what the team model of the future is going to look like? And then this like really important stuff, discovery design, all those things. How do those change? Yeah, really, really interesting topics. So maybe we'll start with the team one. There's generally two, the two ends of the spectrum when you talk to people is on one side, it's that AI is going to replace a bunch of jobs or maybe all jobs, right?

17:41That's the extreme side of this, of the spectrum here. the other side of the spectrum is AI is going to elevate everyone and, and everyone's going to be more powerful and we're going to need more people to manage more AI. And it just means more, more people need it. Right. And more people do more sophisticated things, more of everything. I think I tend to lean more to this side, the more of everything side, but it's probably somewhere in the middle. Right. And there's an arc there too. There's an adoption curve because in early days, especially new companies starting out they're just not going to hire as many people because we're still figuring out as as an industry as individuals as a society like what's the how far can we push this tool in like what's the what's the real limit of me as a solopreneur when i'm surrounded by ai i don't really know because i haven't even built out all the systems to fully leverage all these tools, right?

18:37And the tools keep getting better. So I think early days, and especially with small companies, you're just not going to get all those specialized roles hired. But I think eventually the pendulum will kind of shift over to the other side. And my vision is actually that if we talk about entry-level developers, mid-level developers, which has been a particular role that's been called out as being particularly vulnerable, I think those roles are not going away I think those roles are going to be made to be more sophisticated right and an entry level dev will have to do more architectural thinking, they'll have to do more sophisticated work because it's not going to fly anymore to go in and solve a couple logic problems in a code base which is the type of things that I was doing as an entry dev because AI is actually better at that than I was back then, right it's like way better yeah so now i love that take because home humans are still important in this in this future which is a great thing but yeah i mean we've seen this pattern before too right where a new technology comes out we're like oh crap we're you know all our jobs are growing but then new things emerge and humans are very adaptable in that set so i love that perspective what about like on the let's go to discovery and design pick which one you want to kind of hit first but what the point you were making earlier, I think was you see a lot of folks going straight to high fidelity versus like the, okay, let's not worry about like the colors and like all the components.

20:08Let's like just get wireframes, lo-fi done to kind of figure out the flow. But a lot of people, because you can, it's easier now, are going straight to a high fidelity. What, what do you think that's a good thing? What are the implications? How does that start to shift and manifest as we go? And it's funny because the post you're referring to, I was reading the comments a little earlier and several of the comments are like, I don't even go to HiFi, I go straight to code by using lovable.dev or something like that. And so, you know, they've leapfrogged my whole point of no more low and medium fidelity mockups and they're just going straight to the proof of concept.

20:47So I think that's really cool and I think we're going to see a lot more of that, especially for early stage products. But I mean, the fact of the matter is as you get customers, as the product grows, as the organization grows, all of this stuff becomes more complicated. Every little change matters a lot more. Let's say you have one customer, okay, and you make a change in the product. Well, you know exactly who's going to be affected and how. If you have 10 ,000 customers and you make three changes in the design, you either need humans to be holding that knowledge in their head so they can actually make coordinated design decisions, or you need to have orchestrated some highly sophisticated AI system that can actually hold all that knowledge in its system.

21:34which yeah if if you can do that cool but it's even designing and developing a system like that's really hard so it's like there's there's these thresholds where it's like how much at some point it's probably actually cheaper to hire a human than design an ai system that's as sophisticated as a team of three designers um yeah so i don't So that threshold's interesting to me. I think people that go into the press and say that all the jobs are going to be wiped out are not thinking about that reality. I think the counter argument is that the AI will become so sophisticated it can kind of do that itself.

22:17Maybe. Yeah. Maybe. But we still have to guide it. Like we're still training these things, right? They're trained on our language. So I just don't know. I don't know. Um, I think too, the other piece that's interesting, cause I like, I always like a lot of things right now, I keep going back and forth on both sides. This is another one where I'm like, oh, it's cool. You can skip steps and go straight to high fidelity. But then again, like I noticed this pattern with anything with AI, you can become over relying on it. And if you're not careful, it can kind of take you down a road that maybe you wouldn't have gone if you kind of were more thoughtful about it, especially in the very upfront phases, you know, where you are kind of really intentionally thinking through the workflow and the experience and all of these things.

23:05So it's like another one of those like nuanced things that I think over time we'll just have to like be how it all plays out in a sense. Yeah. Yeah. And I think it comes back to, is your brain keeping up with the AI as it's making changes or not. And if it's not, then you wind up in these weird spots where you're like, oh, this isn't what I wanted and I don't know how to get out of it. Yeah, exactly. Exactly. Yeah. And then because you were so dependent on AI to get there, you're like, oh, I can't get out of it myself. Yeah. Another interesting pattern I've kind of noticed, I'm curious you take here, is engineers adopting these AI IDEs and tools and all the different things out there versus someone like you that's not sitting there pumping out code every day and you're not like deep in the syntax.

23:54I've noticed this pattern where a lot of engineers will try the tool quickly. It won't give them what they want and they revert back to their old habits. And I get a lot of this in, you know, the comment section and things, just engineers just going like deep into, oh, it'll just do front-end code. It sucks and all this stuff. But like you've witnessed it right when you're actually building intentionally in small increments like you can build some powerful stuff and somebody like you or me who's kind of more of a product based person not deep in the code we're forced to keep working with ai because we don't have another option i'm curious your take there what have you witnessed from engineers specifically and their adoption of it it's like the ones that do adopt it do some crazy stuff because they have deep, deep knowledge.

24:41Yeah, I think I agree with all that, by the way. I think myself and yourself and people of our type of profile, our type of background, we're like the perfect early adopters for these tools because like you said earlier, I'm not deep in the syntax. That's the main reason why I picked up Cursor because it solved the syntax problem for me. At the same time, it's also true that for probably most developers who are writing React every day for four to six hours a day, you know, they're just more efficient writing it themselves. They've got a particular way they want it to be written. Maybe there's like code standards and they're on their team.

25:25They know exactly what they need. They know what files to go to. It's just in their brains up front. And so it probably is more efficient for them just to write it. And when it introduces bugs, which it does, and you have to fix them. I bet that's really frustrating. However, really, really technical folks are in the best position to get the most value out of these tools. Some of my friends who are very, very technical and have like fully gone all in on PDD or whatever their flavor is of LLN-generated code are doing incredible things. Like I was just on a call with a good friend of mine who created his own programming language.

26:09And I don't think he's done that before. I certainly haven't. And he didn't write a line of code, right? GPT wrote all the code. He prompted his way into a new programming language that solves a very particular technical problem that he needs solved. And it's just fascinating. It's like, that's a really sophisticated, advanced thing to do. So if you are a really technical person, you can take your skills a lot further by using this stuff. So yeah, no, that's great. Wait, so just to kind of wrap it up one thing we didn't get to earlier on, like your experience of building, you've actually built like a real product that you brought to market is generating revenue.

26:51And you did it a, in a very short period of time and you did it as you know, let's we'll, we'll, we'll put you in the non-developer camp for now, but distill talk us through a, like what it is. What was that experience of doing that in that kind of like moment of I have something that's real and lives out in the world? Yeah, absolutely. It's been awesome. It's been an awesome wave to ride. The V1 I built over Thanksgiving over a weekend, full PDD, didn't write any code. And shopped that around to some friends, got some feedback, quickly did a V2, V3, made a TikTok post about it. You make TikTok posts like I do.

27:31sometimes they catch on sometimes they don't it feels like you're playing a game of roulette and roulette yeah and and that particular post caught on really nicely and so i got tons of traffic out of nowhere i didn't really i wasn't really prepared for it my servers crashed that night which is hilarious i had to like oh let's go into heroku upgrade the servers to like a higher tier and i wasn't ready to be monetized didn't have any payments processing built out built that out as quick as i could did a bunch of analysis to figure out what proper unit economics might look like for the product and by christmas i was monetized so it was a it was a really fast like one month sprint basically from idea to monetization first time i've ever experienced that it didn't write any code and it's completely changed my entire world view of software development of product development really so and it's it's really cool probably so essentially like it's leveraging ai you kind of you can describe it better than i can but it's basically you're setting up these different areas of interest and it's like bringing you back these news stories and just relevant information so you're not having to go through a million different newsletters or you know scrolling tiktok aimlessly yeah yeah if you want it to it can be the last newsletter you ever have it's it's cut that's that you're you're like a marketer too now that's great that's your hero copyright that that was on the home page at one point i think i took it off okay cool i still like that line but it's a little bit like google alerts but way more powerful because because llms exist and it's like you put in any topic you want it searches the web every day for top for new web pages that are related semantically to your topic and then it summarizes them and kicks it over to you in an email and this kind of like executive summary style output.