Claude Code for Product Manager

Thomas: Until a month ago, I never did write a single line of code. Since then, I built a game, a browser, plugin, a app, an entire app. Is that crazy for you as a product manager? It changes the game totally. You have a data analyst available for you, uh, 24 7. Right. You can prototype things really, really fast. Um, you can analyze huge amounts of data and synthesize, um, you know, customer profiles and stuff like that. Right. And it, for me, it all started with one conversation. The one conversation I gonna show you right now. Enjoy. Welcome everybody. I'm super excited to have Mikhail with me today. We met in Timi Ishuara at pro conference, right?

Mikhail: Yeah, exactly. Glad to be there. Thank you, Thomas, for invitation. So currently I'm a chief product officer for the largest E commerce marketplace in Azerbaijan. And that gives me the luxury to implement those tools at scale. Um, and um, I'll also share some things that have worked, uh, quite well and some things that didn't really work out that well. Uh, so it will kind of give you like a better practical overview. But let's start, let's jump into the actual setup. Uh, I'll share my screen. Nice. Uh, do you see my screen?

Thomas: One second.

Mikhail: Yes, lovely. Um, maybe to make this, uh, more um, effective, I would like to structure it into multiple, uh, use cases. So the first, uh, use case is let's just figure out how to build, um, an entire, uh, web app from scratch from basically nothing. Uh, in a matter of, I don't know, 10, 15 minutes. Um, that would be decent enough in terms of architecture, scalable and secure. Second thing, I'll also kind of give an overview of how to implement this at scale in your organization because there are definite use cases. Uh, in our case, uh, the problem was the teams were spending a lot of time on building landing pages and all. And this could be fully automated with AI, so there is no need to involve actual engineers. The third, uh, example, which is also quite cool, is outside of coding and um, it's in relation to product discovery. How, how the same setup could be used to figure out with new insights and estimate the potential impact of whatever hypothesis that you might have. So, uh, one question, because I see

Thomas: you are already in cloud code, right?

Mikhail: Well, I'm in courser. I haven't yet.

Thomas: Uh, but, uh, it's connected to cloud code because, you know, like, I would assume that many people don't have the basic setup yet. So what do you need as a prerequisite?

Mikhail: So as a prerequisite you basically need to, um, I mean, open your terminal

Thomas: yeah,

Mikhail: and install Claude code. And um, for the installation there is a simple command that you can use, uh, from Cloud Code's website. I think it's just a one liner of code to be honest. Uh, let me see what it is. Um, brew install. I think that should do the trick. Yeah, that's doing the trick. Um, I already have it pre installed, so I'm going just for demonstrational purposes.

Thomas: Yeah, so it downloads now the files and installs it and then you need to connect it with your account, I guess.

Mikhail: Yeah. Uh, it's also quite simple. Uh, it will kind of interact with the web. So once you install it, uh, it will ask you to log in, you'll do it from the web, it will ask you to input um, a, uh, separate, uh, it will ask you to choose a separate tier. Right. And you can also do it from a web and pay and all. Um, so kind of the most difficult part is done. And the way you launch Claude is like with a simple command like this and yep, proceed, that's it, you're in cloud code. But the only limitation with this is that um, uh, it works in terminal and you don't have the luxury of a full ide. Uh, which is just. It makes it more complex for you to interact with the files because you'll need some kind of an editor on the site. Uh, of course you can also fit folders into it. And that's the beauty of Cloth Quad because it kind of spans out multiple agents. Uh, but I think the best convenience that you can get is if you combine the cursor id, which is kind of a background here with uh, a cloth and they work quite nicely together. So what I did, I just created a separate folder and uh, then you just press this tab, this tab opens a terminal and then you can launch code code right inside your cursor, which is kind of like next layer of depth. Right. Um, and uh, so um, on your left I prepared just for demonstrational purposes some specifications here, so I'll quickly guide you through it. So the first one is specification, uh, md. So this is the project that we want to build, uh, from scratch. And this is uh, just like a basic Trello style clone. Um, and um, outside of uh, the actual functionality, like boards, list cards and uh, more granular details. Uh, we also have uh, design requirements here which is like ui, uh, visual themes, like how the motion should look like interactivity, the toggles and all. Uh, so take the basic stuff. And the beauty of Cloud Code is that it will use all of those specifications as A context for any query and uh, your project will stay consistent so uh, the model will not hallucinate like random stuff from scratch. That's one angle. Another angle is that you need to also load skills into cloud code. And skills. I mean it's a big word. Uh, in essence it is just a text file. So yeah, here I've already preloaded multiple text files, uh, with specific job requirements. So for instance we have an architect. And an architect is um, like a full stack senior uh, engineer that knows a good deal about backend architecture, database front end integration, system design and all. And the thing is, um, every single query to Claude will kind of require Claude to adhere to those limitations that you've outlined here for each of the roles. Um, then the backend engineer thing. Right.

Thomas: Um, and uh, how did you create these specifications?

Mikhail: That's the fun part. I didn't uh, and you don't have to, um, uh, actually there are repositories, uh, online. Let me find one so I can show it to you. I think it's called subagents, uh, cc. Okay, let me see if it works. Yep, let me just Google it. And you have this project here, uh, cloud code agents. I have like some security issues with the firewall, but I mean you can find them online. It's very simple. Uh, these are basically, basically text files that are like grouped by different categories.

Thomas: Perfect. So we hired, now an architect. We hired a backend engineer. A designer.

Mikhail: Yes, yes. For a fraction of a cost of a real, uh, real specialist. Um, so yeah, we have all of this loaded. Um, we have the specification. So basically what we need to do, we need to break it down, break this large ambiguous project into smaller chunks. And I think this is what we're going to do. So uh, just talk to Claude directly, rely on specific, and using the specific requirements from architect D, break down um, project into the granular tasks. Yep. And you can be as vague as you'd like. Uh, what I personally found is that cloud code is quite smart, um, way smarter than the majority of coding models out there. So it will understand you from basically

Thomas: half of a word M. And these files, if you have here spec MD and uh, front end engineers, did you store them somewhere on the hard disk or did you load it into cursor? I'm not sure.

Mikhail: Just like a folder on your hard disk. That's it.

Thomas: I see.

Mikhail: Ah, I think uh, you can go like, I mean of course like more advanced. You can create a GitHub repository and commit there if you'd like to. But I Mean, it depends if you're doing this as a solo vibe coder, I think this is, this is a sufficient setup.

Thomas: Yeah, I think, you know, like as the more complicated it gets, the harder is to start. Right. Like for me, I'm not a, I'm not a software engineer. I don't know, your background looks like, uh, you're a software engineer.

Mikhail: No, I'm not. I mean I'm a software tinkerer, I think, rather than engineer. So I'm coming from an economics background. But that's the beauty. So essentially the barriers to coding are so low that you don't really need to understand any of those concepts. Uh, I think you'll see the reason why is that like um, what is currently happening is that two things. So one is uh, Claude code breaks, spins, multiple agents and uh, each agent is working on specific subtasks. You don't even see it because it all happens in the background. And the second thing is that uh, uh, the second thing that is quite important, um, is that it sets up your own environment. Because uh, one of the biggest issues with coding is that you had to create an environment. You had to create a container like Docker or whatever. You had to load up dependencies, libraries into it and they were conflicting all the time. To be honest, I always found it to be a pain in the ass. Uh, and it was like a huge barrier for any project that uh, develop. But here you don't really have to think about it. So I'm basically sitting here drinking coffee and chatting with you while Claude COD is doing all the heavy lifting. So you can see that it says set up project foundation and infrastructure.

Thomas: Uh, okay, crazy.

Mikhail: Yeah, I know, I know. And the thing is, right, so when you have this one, uh, sided, uh, kind of chatgpt type of an interface interaction, you're not getting the luxury of this. Because what CLAUDE code does, it interacts with your entire project, with every single file and folder and it selects certain chunks of those files that are relevant to your query. Um, it's similar to using this augmented retrieval like a rag. Uh, in a way.

Thomas: Interesting. Where did you come up with the spec md? Is that like something you just thought, okay, let's recreate a trello and write it down or head from GPT or

Mikhail: it comes from GPT. Okay, I just uh, asked ah, GPT to write me a spec. And uh, aim for as much feature parity as possible.

Thomas: It's really funny how you use one AI to feed the other one. We did it like with the intro video of the just product. Right. We always like GPT. Uh, okay. I want to create a video or please make this and that. Right.

Mikhail: Yeah, yeah. So it's kind of a matter, you know, so one AI feeds to another AI. And uh.

Thomas: Yeah,

Mikhail: and yeah, it's still. The.

Thomas: Still.

Mikhail: The work is still ongoing because it's a. It's a heavy lifting project. Uh, but, uh, um, it's okay because, like, you can just go and do whatever you want meanwhile.

Thomas: Yeah. And did you ever. I mean, meanwhile, we, uh. Wait, did you test other web coding, you know, tools like global or something like that in comparison to that setup?

Mikhail: So I did. And, um, um, I would kind of group them into like buckets. Right. So, ah, I started Vibe coding even before it was called Vibe coding. When I first like ChatGPT01 model appeared because it was. Wait a second. I think it PM mock interview. Okay. Spectre. Let me see. It might have taken, uh, like an incorrect specification.

Thomas: Okay.

Mikhail: Uh, yeah, it was the other project that I was working on. M. What the hell? I think. I hope you'll cut it out.

Thomas: Well, you know, like, actually, I think this is just what happens in reality to many people starting. It's just interesting what you do now. Right?

Mikhail: Yeah, I'll be more specific project. Uh, it's the Trello. Trello clone project. Be very specific about it.

Thomas: Okay. Interesting.

Mikhail: You're absolutely right. I apologize for the confusion. Let me read the correct spec for travel clone project.

Thomas: Okay, let's see. You know, because I think this is, you know, something which happens to each and everybody. Right. Yeah. Uh, stuff is not working the first time.

Mikhail: Yep. Yeah, sorry.

Thomas: Like with, um, comparison, you know, lovable. And, um, the setup. And you. You just shared that you started with, um, GPT1. I think that's where we lost your story.

Mikhail: Yeah, Just, uh, a second. Um, I think what actually happened is that I didn't save the spec M. Uh, and it referred to a different spec in a different folder. So it was looking across my entire basically project repository, which is also like a pretty cool feature, is that it just steps outside of the boundaries that you set for the project. But you also have to be careful when you like, do this at scale because it might interact with the files that you don't actually want it to interact with.

Thomas: Yeah. So Glauco basically has, um, access to your hard drive, right?

Mikhail: Yes, yes, absolutely.

Thomas: Um, how do you limit, you know, like, if you say like, okay, it sometimes goes just, you know, out and takes something else. How do you limit the data access or, you know, like, the things you want CLAUDE code to consider, um, I

Mikhail: think you can do this in the specification as well. Right. So you need to be like very specific about what files or folders uh, cloud code can interact with and uh, you just basically limit it to it and that's it. Okay, going back to your previous question about different uh, type coding tools. So back when like ChatGPT01 was released, I played around with it quite a bit and I think uh, it was decent for prototyping and early stage MVPs, uh, and all that. Uh, it actually returned workable prototypes and you could actually debug them, make interactions. But you didn't have the luxury of working in this uh, engineering environment where you could uh, feed an entire folder into it because the context was so small that you could, couldn't really. You had to like kind of break it down into multiple pieces and work with each piece separately. It was very confusing and uh, it was like after, after a certain amount of uh, queries it just refactored the entire project which made it like very cumbersome. But then I think there was a wave of this uh, vibe coding tools like Replit, Bolt, dot, uh, dev, lovable, uh, uh, base 44, uh, Orkid and a range of others. Um, I think that they're quite promising for very, very early stage uh, users. What I don't like about them is that I do think that they output very generic uh, landing pages or prototypes, very generic ones. And if you want to make them, um, if you want to custom tailor them, uh, you will just run out of credits, uh, just cost quality ratio wise. I think it's just not worth it. Uh, for a project like we're doing right now, uh, Claude would be a much better alternative here. And the thing is this whole, whole setup might look intimidating but it's not complex like you've seen. Um, I'm literally speaking natural language with it.

Thomas: Why do you run it within Cursor and not just use the terminal to do stuff with cloud code?

Mikhail: Um, yeah, it's a good question because it's easy to switch context. Right. So you have cloud code running here and if the cloud is, if Claude code does some changes in the files you can immediately access the file. I just like kind of the convenience, uh, it creates. Uh, you could also use agents here. Like so Cursor has its own agents. You could choose an LLM, uh, whichever you like. But uh, personally I think that Claude code beats them in terms of the quality of coding. Okay. Yep. So vive Trail, abort Breakdown is ongoing. Yep.

Thomas: So do you um, how do you do it in, in your company? Do you suggest the engineers work like that or the project managers? Or is everybody, everybody the freedom to test out their own tools or how do you do that?

Mikhail: Yeah, so uh, it's also a great question. Uh, so what I'm doing right now is that I'm pushing all the product managers that uh, somehow work with those early stage projects, uh, to do the prototypes and build, uh, build like easy stuff themselves. Uh, they all have like GitHub accounts, they do commits. However, what is important is that uh, you just cannot come up with some random stuff. Right. Um, you need to follow certain guardrails. So for instance, even if you're creating a landing page, it has to work in a certain environment that your organization has if you want to load it in a subdomain. So our engineering managers have prepared a set of guardrails which, which product managers just simply use as specifications here. That's it.

Thomas: Okay. Um, um, just make one or two examples how your product managers like what, what they do. Like you know, landing page sounds more like marketing, um, department. Right. Like what do product managers do? Do they, do they really do functional uh, changes on your website or like what's the scope of things uh, the product managers are doing?

Mikhail: Yeah, so for instance, um, it's not. Well currently I'm a bit cautious about kind of uh, putting uh, this cloth code and fraud to action on our big legacy architecture in the existing E commerce because it might end up with unexpected consequences. Uh, but uh, when it comes to completely standalone projects, and you're right, like uh, a large chunk of them are in the marketing and growth department. Like any kind of marketing LinkedIn pages or whatever. Doesn't make any sense to use any resource to build them. But apart from this, um, there are some local projects, uh, in product teams. For instance, uh, we have a supply chain department and they have the own app for couriers. Right. It's a standalone app and you don't really need to have like super complex integration there. Uh, you can like build it yourself or you can make tweaks yourself. So I'm pushing product managers to be very hands on with this.

Thomas: Great.

Mikhail: Um, so yeah, you can see that the project was broken down into very like granular pieces. So there is even like uh, database architecture there, uh, APIs and all. So uh, let's just build it right? Take the task breakdown and uh, start building it. Strictly adhere to the roles client in the files back, engineer MD front, um, end designer MD do not um, lose focus of this back. Yep, something like that.

Thomas: Interesting. So if I have here now a front end engineer, could I just put in the specification, hey, use this design library or something like that that it's really like on point with the, you know, branding and everything.

Mikhail: Yeah, yeah, exactly. So you can like limit it to certain brand books, m design systems, color codes or any other specifications uh, that you actually want. And that's why this context is engineering thing makes it so useful, uh, far beyond just the basic prototypes.

Thomas: M. And we can tell the architect, I don't know, make a microservice architecture with whatever security and whatsoever. Right?

Mikhail: Yeah, exactly, exactly.

Thomas: Love it.

Mikhail: Yeah, exactly. So I mean if you think about this, you can launch a new MVP every day. You'll just have to extend your paid tier. Uh, and that's it.

Thomas: Yeah. Do you have something in, in production, a project or something like that? Because you know like people always worry about um, you know like having something like that in production and maintenance. M. So, uh, any experience there, um, how that works for you?

Mikhail: Um, so yeah, I had um, uh, actually had one project almost a year ago which was built completely on uh, this ChatGPT01. Not in the classic wipe coding sense that we're using, not in the cloud code. Uh, I mean it was okay. Um, but the scale of the project was relatively small. It was like generative AI imagery. So nothing like over the top. I didn't experience like any massive downtime, uptime or like fraud attacks or whatever. Um, I'm also curious to know like what would be like how it would look like on bigger scale. And the one project I'm um, uh currently planning to push to production maybe like next week it will be closely related to the substack to my substack account that I'm currently using. So it would be about um, mock up interviews. Like how can you do mock up interviews for product managers, uh, based on real data, real uh, evaluation criteria. So you could practice building stuff, stuff like that here, like using quote code. It's exactly what it's meant for.

Thomas: So you would take existing uh, interviews, videos, uh, transcribe them, put them in here and then um, basically give it the task to respond or how would you. Something like that.

Mikhail: So I'll do it slightly differently. So um, since I have like a large knowledge base of how interviews are done, what the expectation criteria are, um, so I'll be creating a library of different uh, types and specific interviews for different companies because assessment criteria, they vary wildly and then you could basically uh, you practice with uh, real time tutor, like using your own voice and get feedback immediately after it, after the interview. I'll send you a link and uh, with the access and all after, once it's ready. So. Yeah, yeah. Ah, so you can see something is definitely cooking.

Thomas: Yes, Working.

Mikhail: Let me update to the list and you can see like all of the kind of file structures here. Uh, it interacts with it dynamically. There are multiple agents working under the hood. They're making their own decisions so you don't really have to guide them or provide any excessive inputs. Just a second. Setting up project structure.

Thomas: And is that like the quality is better with Claude code um, than the other tools? Um, it also feels like it's working a little bit longer on the, on the project.

Mikhail: Right, yeah, because it's, it's a large, we're basically, basically building an entire Trello from scratch. Something that the team might have spent, I don't know, like half a year, uh, when they, when they released their first Trello. So yeah, it's a large project I think like with smaller ones, uh, or with just basic prototypes, I think we can, we can do it much faster. Of course. Yeah. And the quality is um. Well I'm relying on two kind of uh, two criterias here. So one is objective criteria. Uh, and uh, from everything I read about, uh, evals on uh, quad quality, quad code I think beats other models uh quite substantially in terms of performance. But uh, subjectively when you play around with coding and uh, if you take any pet project and you try to bring it to fruition and you experiment with different models, uh, the level of frustration that you experience with code is probably the lowest out of all the models out there.

Thomas: M and is that based on experience or is that based on reading stuff or.

Mikhail: It's based on my personal experience, uh, playing around with different projects. M Seating tokens. I actually like the interface here. Uh, tells you what exactly it's doing, explains uh, you why it is doing this. And you can clearly see that it takes just chunks of the code. It doesn't refactor the entire project. Mhm.

Thomas: Have you used this to leverage your own data? Like for example, I don't know, um, use your own database from the service department to analyze the structure, the calls and stuff like that. Like have you ever used, you know, your, your existing data database and connected with um, Glauco to analyze or do anything with it?

Mikhail: I actually did and I found out it was extremely efficient at this. Um, especially when you're dealing with large data sets uh, so we took our entire AB test, um, data set, we loaded it into cloud code, um, and also with additional user, uh, journey data, um, from dashboards. And uh, we back tested it against the actual AB test data to see whether in it gives correct predictions.

Thomas: One second, I didn't get it.

Mikhail: So you take historical data, uh, from your past AB tests, you basically put them all in the folder here and then you uh, layer all the user data that you have at your disposal. Right, like funnel flows, I don't know, um, just like clv, behavioral profiles or whatever. Right. And then you can um, try and try to do historical predictions. Right. So, uh, and compare, compare it against your actual data. I think that that shows kind of the uh, the accuracy of, of estimations that Claude Quad is doing. And what we've seen was that the swing was between, between like 20%. So for product purposes I think it's quite decent. That's one, one uh, thing. And um, if you put this into practice and think about actual applications, um, so, um, for instance, I have a large unstructured data set. If you load this data set into ChatGPT, uh, it will simply just throw up because there is just too much data. Uh, if you load it in other models, maybe like Gemini could work because it has a super large 1 million token context. Um, but it didn't work quite well for me. But if you load this data into cloud code, it will digest and process the data in maybe with 15 minutes and then it can build uh, conclusions. Like one example that I did basically just yesterday, I took uh, different customer journey maps, uh, and uh, there were 60,000 of different combinations, uh, and it grouped all of those combinations into kind of engagement. So how many clicks a certain user did. Uh, and when I looked at the buckets, I was quite surprised that uh, our web was uh, the lowest, um, in terms of engagement. So 40% of the users have actually dropped off, uh, based on the amount of clicks and the versatility of different customer journeys. All right, uh, with all that said, uh, I think we have something brewed here. Let's ask it spin out local host instance so we can uh, click and play around with it.

Thomas: Now. Time to see what happens. Yeah.

Mikhail: Ah, I love how, how responsive it is. Like, let me spin out the local H, of course. Okay. I mean if you think about this, just like the whole development server setup, maybe like 10 years ago it was such a pain. Now like literally I'm not doing anything crazy. Okay.

Thomas: And you know, any experience because, you know, like I was um, chatting with Sebastian a lot. And he said like, you know, any, anything which is, you know, quite. Oh, that's first.

Mikhail: It's still work in progress so no worries. M. Still ongoing. Yes.

Thomas: Um, and basically he said like whenever you have something, you know, newish, uh, lovable and all these tools and maybe clock code are also not so good. So if you have, you know like common patterns because it was trained on a historical data. Right. So you know, architecture wise or if you want to come up with something new or newish or unique. Right. Do you have any experience in, in that sense?

Mikhail: Uh, yeah, I actually have. So um, newness comes in different shape and forms. Right. And most likely when we're talking about new, we're talking about uh, either libraries, uh, that don't exist. So there is no ready made solution for this which you can just plug and play. Um, uh, um, uh. Or we're talking about a very complex engineering challenge. M. And uh, in both cases what I've noticed was that uh, the models, they just got stuck in this loop of Deb death. What this essentially means is that they're over promising to you. What they're saying is that hey, yeah, of course we can deliver and they just like feed you the prototypes that are not working.

Thomas: Yeah. Uh, and driving hard. No problem, we can do it.

Mikhail: I think there was a research about this, uh, which basically said that uh, models try to be excessively helpful to the extent where they don't see their flaws. So they can't really tell you, well, hey, I actually have no knowledge about this and I have no idea how to build this. Uh, instead they try to kind of mimic this. So yeah, it might not work out however, uh, if you break this down into very granular specific tasks. But you need to have an engineering mindset for this. Right. It might actually do the trick.

Thomas: Yeah. Interesting. You know like where I saw this behavior before.

Mikhail: Yeah.

Thomas: I was living in China and you know like they, they cannot say no or you know like they cannot say I don't, don't know or I can tell you or something like that. Right?

Mikhail: Yeah.

Thomas: That leads to really awkward things because you know, like you just approach somebody, say hey, um, what's your name? And if they don't understand they just

Mikhail: say yes, yes, exactly, exactly.

Thomas: Or I was, I was you know, like walking uh, the street and I didn't know where to go anywhere. Right. So I had a map in my hand. I showed it to the person said like okay, I want to go xyz. Right. She looked at a map and then she Said like, okay, down the road here, left. And then you know, like she left and I was like, okay. But you know, like the, the map was up upside down. There was no way she could understand, you know, like, where would go. Whoa. Trello vibes.

Mikhail: Yeah. This is what we have built. Uh, coming soon. Serenity Mona Sundown. You can actually play around and just look at the design itself, right? Uh, of course, like we set up the specifications for it. Uh, however, like the alignment, the colors, like the typography just looks so nice. Um, I think we need to kind of make a step forward and add some um, at least some functionality here because like we have all the foundations there.

Thomas: So was the instruction in the spec MD that there should be already a board or was it just for the landing page? I think you wanted to have a full Trello functionality, isn't it? Yeah.

Mikhail: Ah, yeah. So the thing is, right, so, um, considering the time constraints, I kind of pushed a little bit further. Right. To deliver at least something visual. Right. Because of how I prompted Claude, it was biased into delivering the visuals, uh, ah, and sacrificing kind of the uh, functionality substance. So, uh, we actually have the entire task breakdown. But if you want to build it fully, uh, uh, probably you'll need to spend a few hours or maybe like a few days if you want to play, polish it, uh, doing like going step by step.

Thomas: But how do you know, um, which chunks the AI already created and which one you need still to create?

Mikhail: Uh, yep, great question. But you just play around with the prototype. If you open the prototype, um, it will become pretty obvious, like what works, what doesn't. Right now we can't access the boards, um, can just change the colors. Um, and once the boards are accessible, you can clearly see what's working, what's not, um, and just ask it to make changes as you go. Yeah. And you can be a very demanding stakeholder.

Thomas: Yeah. Um, to watch. To which degree are you, uh, demanding from a, uh, AI?

Mikhail: Oh yeah, like to, to the fullest. I'm a horrible stakeholder.

Thomas: But you need to take care, you know, like here's a, uh, slight, uh, warning because at some point AI will be smarter than us and they rule the world. And if you behaved, if you behave badly, it will remember.

Mikhail: Actually, uh, no, I'll just contain it in my folder using the specification. Yeah, um, let's maybe see what it comes, uh, up with. But, um, going back to one of the topics that you raised before, um, you can put whatever you want here, not necessarily the specifications or the coding uh files you can put data, you can like interact with the data in whichever way or shape you want. You can put rfds. Right. So you can ask quote, quote to kind of challenge your, the logic, the thinking, the depth of your existing R&D's also using different roles if you want to. So like you can play with it like however you like. So it's not necessarily for coding, but I think coding is just the best use case for it. Mhm.

Thomas: Because coding is quite structured. Right. I think this is one of the easier things for, for the ah, AI to do.

Mikhail: It's easier in terms of like probabilities. Right. So uh, it's easier to expect like a certain function be followed by another function within a predefined syntax rather than like natural language which can be very uh, like random. Okay, so what's going on here? How far are we building API routes for card operations? I think we're pretty deep. Mhm.

Thomas: Reading board page component with the theme support.

Mikhail: Let's see uh, what's happening in the.

Thomas: Did you, did you ever do like uh, you know like for prototyping? You know, often you want something you know on your website. Test, test it. Right. What if I have a button here? What if?

Mikhail: Yeah, yeah.

Thomas: Around with you know, taking screenshots of your website and then um, you know just adding functionality to the website or you know like integrating into um, you know like I don't know, into your existing website somehow.

Mikhail: Um so yeah you can, you can take screenshots or you can like download just basic uh, styles htmls of course if it's possible, if your website isn't overly dynamic and then you can ask cloud code to recreate it, add the buttons or change certain UI components if you'd like to spin it out uh, on a local host. Super easy task.

Thomas: Creating board page component.

Mikhail: I'm hopeful we'll see something by the time um, our session ends. So phase one complete. Uh yeah, you can also see like the project file here basically what has been done.

Thomas: Where is that? That's like uh phase one complete and okay I see that's phase two.

Mikhail: So probably you'll need some database integrations as well like super base, maybe like Gmail integrations if you want to add like a uh, login layer on top of this.

Thomas: Does it, does it ask for these things or um. How do you know if um, that you need database or something like that?

Mikhail: Yeah, it can ask you like if you don't specify it will give you a suggestion uh, what to use. But I uh, typically Specify. Because there are simple tools that you can use. Like for database you can use Supabase, and I think that's the basic integration that Lovable has as well for like logging. Uh, you can also use Supabase because it allows you to use Gmail and store it in their database. Like, literally. You can set the up without any technical knowledge.

Thomas: Okay, I need to start a, uh, project on the weekend.

Mikhail: Yeah, I mean, I think you can already push it to production on Monday and uh, start earning money by Wednesday

Thomas: and being rich on Friday.

Mikhail: Um, it depends on what you sell there, but yeah, why not? And depends on your definition of being rich as well.

Thomas: Yeah, that's true. You know the video on Instagram. Hey, chatgpt, uh, I want to have an idea to be rich. Uh, deep thinking, deep thinking.

Mikhail: Uh,

Thomas: take your time. No problem. Yeah, uh, uh, little work. No work. All right, let's see. Ah, no, it's uh, still, it's still building something.

Mikhail: But we, we already have the demo board, um, loading board. So. Yeah, pretty pretty. I mean, not perfect, but it's pretty nice. Uh, I don't know, like, uh, review my backlog priorities for the week.

Thomas: Okay. Crazy.

Mikhail: And I think you can shuffle it as well.

Thomas: Drag and drop.

Mikhail: No, it doesn't drop, doesn't work yet. But you can add it quite simply. Yeah.

Thomas: Um, can you go into the card already?

Mikhail: Let's see. No, not yet. You can just like, um, edit the, uh, edit. But I mean you can layer in just a single prompt. You can add whatever functionality on top of it that you like.

Thomas: Yeah.

Mikhail: So, yeah, let's see what, what's happening in cloud code. All the pages are working. Responsive design demo board is there. Yeah.

Thomas: Crazy.

Mikhail: And I think that, uh, when we're talking about the new, uh, age of, uh, subscription as a service tools, uh, you can build your own, like, web apps from scratch, like literally. Um, if you think about like how it was 10 years ago, this would probably take me, a, uh, group of engineers, like three, four engineers to actually build it. And this will take me a few months for sure.

Thomas: Yeah, this is actually, it's quite insane. Quite crazy. Let's go just one more time back. Let's assume we want now the tasks, um, to be open or something like that. Or how would you go on if you really want to build that project? Right. Would you now go into the faces and see. Okay, let's take the second, the third, the fourth phase, or would you just go function by function and prompt, uh, code to, you know, like build now a, ah, second level, like that we can click in or, you know, like, how, how would you proceed now?

Mikhail: Um, so, because I love tinkering with, with stuff, I would probably go like, very functional. Um, so I'll, I'll ask, like, add like an ability to drag and drop, like ability to open cards. Uh, because like, you can play around with it however way you like. Right. So the project management doesn't matter anymore. And uh, uh, that's what I would probably do. Um, and then, like, once I've exhausted the functionality here, I would probably move on and I would add the layers of what I was missing here. Let's say I would like to have multiple projects. Um, I would like to have kind, uh, of an authorization functionality. I would maybe like to sync it with, uh, I don't know, something else like Google Docs or whatever. Um, yeah. And, um, probably that would be my approach. Until you feel that, uh, your masterpiece is ready to be, to be put to production. And if you're, if you're, if you're, you're questioning how to put it to production, you can ask, quote, quote, to build you a specification. Build me a new manual on how to push this project to production and deploy it to my domain. That's pretty much it. Yeah. Thomas, unfortunately, I'll have to jump off. I have a meeting and. One minute. Uh, but, um, thank, um, you very much.

Thomas: Thank you very much for sharing this. It was really exciting. Uh, I think, you know, like, if, especially if you, if you've never done it before in code and, uh, um, having this set up, um, I think it's super, super interesting. Yeah. So I'm gonna play around with it on the weekend and try my first project and, uh, ship it by Monday.

Mikhail: So you can do something. That's the spirit. That's the spirit. Um, yeah, I think you can definitely do it. Um, and, um, I mean, especially at such a small scale, you don't really have to think a lot about security and all that stuff.

Thomas: Yeah. Thank you for sharing.

Mikhail: All right. Uh, thank you, Thomas. Uh, as always, uh, was a pleasure. Was a pleasure as well. Yeah. All right, so have a great.

Thomas: Bye. Bye.

Mikhail: Bye.

Thomas: See you.