Why Non-Developers Are Building Better Software Than You Think

**[00:00:00]** Hello everyone, my name is Armando Perez-Carreno and welcome to the Web Talk Show. Today with us is Jose Duarte from Pangea Money Transfer. How are you Jose? I'm doing well. How are you? Thanks for I'm doing great. Thank you for asking. I think there's going to be a lot to talk about related to coding, AI assisted coding, marketing, and even some apps. So before we do, why don't we give the audience a little bit of a quick background on how you got into all this? Yeah. So I am a performance marketer. I have been working in marketing around mobile apps, finance, um loans and and those kind of things for about 15 years now. Um and I've always done a little bit of everything. I've done some product management. I've done editorial, I've done creative, and my biggest uh area of focus has been performance marketing. And so in this last year of with AI being such a buzz word, we had um I was telling you I was saying earlier one of a vague company goal, which is not usually a good thing, but in this case, it was nice where they just told us, you know, AI is this new thing. We don't know, we don't know what it is you can do with it. just go try to figure something out. And I was talking to somebody from our product team and they had, you know, one of the things they did was they launched a feature in record time. It was super And I asked them how they did it and they used this tool called Lovable. And so if you haven't used Lovable, it kind of looks like chat GPT on one side and then on the other side it basically builds you a website as you chat with it. And so started playing around with Lovable and I decided uh for a long time I had this way that I analyzed Facebook ads and it usually uh the way the way that that I had to do it I had to download **[00:02:00]** a CSV. I had to transform the data. I had to make a pivot table. I had to you know fidget around with a pivot table for 20 minutes and then maybe I finally was able to get the insights that I wanted. Um, and so with Lovable, it took me uh about two to three weeks and uh more allnighters than I pulled in college just because I got so into it. And I was able to build that that, you know, basically build that spreadsheet in into a web app. And not only was I able to do that analysis that that, you know, took me so long to do and I did it once a quarter because it was so long, I was able to do it instantly. And I was able to do it from multiple different angles with multiple different campaigns and it just really helped us perform really well toward the end of the year. Like the last three months at Pangea have been record months consecutively. Um to the point where like we're scared for the next month to come cuz like what if it isn't a record month anymore? Um, and yeah, and so after, you know, after Lovable, I, uh, that was around the end of the year, uh, and there was a big update to Claude Code that came out around that year. I think uh, I forget if it was exactly when they introduced it, but, uh, the agents feature that Claude code uh, got a lot better. And so I decided to jump into that. And, uh, and then after that, I have a lot more to tell, but I figured I'd stop there and and let you talk a little bit more. That's a great beginning because keep in mind if you're listening, Jose is not a software developer by trade. Yes. No, I uh I went to school for economics and I got a masters in accounting and then sort of backed into marketing and I've worked with developers throughout my career in in kind of semi-technical positions, right? I say if you you know it's kind of like if you you say I I can go to France and and I can't speak French but I you know I I I'll be able to get around. That's kind of how I feel about software development, **[00:04:00]** right? Like put me in a codebase and I won't know what what it says but I can kind of move around it. That is a very interesting point that we're seeing across the board. people who were not in the software space ever and because of where the tools are now, it's become so easy to get into it. And your example is perfect because it provides a glimpse into the shift that is happening with software. before you would never spend what might have taken you two weeks. It would have taken like an actual engineering team perhaps months to build before a few years ago. Nobody would spend the money or resources to do that just to analyze your ads, right? Unless of course the size of the company. So what has happened is now you could build that and again for the listeners here the the fact that he spent a little bit of time playing around with it building the tool means that now he could just what download a CSV and put it up there directly and then analyze it for you Not even I was I was able to figure out how to plug into the meta API and download the data myself. So not not even download any CSVs anymore. So that's where it's really powerful. Like I find myself nowadays generating apps for calls. Like I might have a call with a stakeholder and I can build a web app that displays what we're going to talk about in the call. Or we can make something that you're going to use because you're going to go to the plant with someone and do a walkthrough and you want to look at the devices. like it's now on demand software for what you You know, one of the things that has been helpful cuz it has been a bit of a whirlwind because I've been learning a lot of this. **[00:06:00]** Um, a lot of this is cutting edge. It's brand new. Like there's not even words for it. But one of the things that's been really helpful for me has been to think in terms of metaphors. Uh, one thing that AI being so knowledgeable about pretty much everything is, um, if you don't understand a complex topic, uh, you can kind of find a metaphor that'll work for it. And so for the development, the way you're talking about it, uh, I don't know if you've seen the latest Tron But that's actually that's actually been super helpful for me to picture it because they have this this kind of concept in the movie where there's this barrier where there's the computer world on one side and then you can cross over to the real world, right? Mhm. And the way I think about software up until, you know, last year that that translation layer happened outside the computer, right? And so you brought as a business person, as a person who wanted to do something with computers, you brought your intention, you brought, you know, what you wanted to do, what you wanted the computer to do for you, you brought it to a development team who then translated that into zeros and ones. And what's happened with Claude and with all the other tools is that translation layer has now moved inside the computer. And so you don't need to worry about translating from English into zeros and ones. You just have to worry that you know what you're talking about and that you know what to ask for. Exactly. That's a great way to put it. And now with the interview tool, you don't even need to know the whole thing because it's going to help you. And that reverse prompting exercise is really good when you're experimenting with these tools because you might know exactly what you want, which is super important. Obviously, you're going to make a mess, right? But sometimes you don't know necessarily how to explain it. And doing the reverse prompting where you tell it, interview me to get clear direction, and it asks you questions, it makes it so much more powerful. One of the one of the best tools that **[00:08:00]** I've used in making my you know because one of the things you worry about when you're doing Vive coding or software with software development with AI is it it gives you spaghetti code is what people call it is just garbage or it becomes super messy and it's hard to refactor. because I know I know I I know that I don't know what I how to put it. I know what I don't know. I am able to build in certain gates at every step of the process where you know we write the specifications and then I have another agent as a reviewer do a review of that. Then we write the user stories that the coding AI is going to use. I have a reviewer reviewer agent code that. Then the coding happens. I have a QA agent and so on and so forth. So that at every step I'm kind of doing an adversarial review of everything that I've done up until that point. And so that helps keep the keep the quality high because you're not Yeah. You're not just like I have this idea and I have this idea. I think the problem a lot of people have with software is like it's very easy it's very easy to come up with an idea and then see if it works and then yeah it works but if you don't take a step back or you don't take the time to review and have the AI fix it um then that's when you get in trouble. That's a great point. When you mentioned lovable earlier, I like what they did in the sense that you, as you mentioned, have a chat interface, but then you have the page right there. So for anyone who's not involved in software at all, it's very easy to jump into that and start creating. If you're using cloud code or codeex, typically you're in the ID, you're in the development environment, you're coding, then you have to upload a separate window and that's where you view it, etc. So, it's a different paradigm, but a lot of people can do the jump pretty easily once they've they've got a taste over there, right? But what you're saying is super important which is the whole agent side because you might be doing some things in something like lovable and I know all of these tools get better every day but **[00:10:00]** as a general example if someone just goes in and generates something it might look good but then once they start iterating on it a lot of stuff can get lost it can get confused and they don't have good source control or something depending on the person right um they might not have steps to take care of things and so what you're talking about it's sort of the next step in in doing things with quality where you have these agents that give you an extra layer of you're not just handing the keys to one which might I don't know lie to itself uh create its own fake tests and things like that. There's someone else to be accountable to and so they can sort of keep each other on track. Now, this might sound super sci-fi to a lot of people. So, if you can, and it doesn't have to be in the technical side, um, because this is more for for whoever's listening. Can you explain in your words how agents work like how you set them up, how you run them, what makes a difference from a typical just chat flow? Yeah, it is it is hard to explain. Um, I think a good starting a good starting point to think about them is if you think about uh your typical chat GPT window. If you start a chat on one window about cats and then you open up a new window and you start chatting about dogs, they don't know each other. And at no point one chat, you know, you you're not afraid that one chat's going to start, you know, the cat chat's not going to start talking about dogs and the dog chat's not going to start talking about cats. So agents are kind of in a similar way a way to keep that context separate from them. Another important concept to understand first too is the way that um LLM tend to process your requests is there's kind of three things usually there's the system prompt. So that's what tells it you are chat GPT you are a useful AI asset like the the thing that open AI programs or or anthropic programs the the coder to be their identity. **[00:12:00]** Then after that is your user instructions. So if you go on settings on chat GPT if you haven't done that you can tell chat GPT how you want it to treat you. So you can tell it talk to me for example I have mine to say talk to me like I'm a product marketer. I don't understand software. So, if I need to make technical decisions, explain it to me like I'm a product marketer and and tell me what the trade-offs are because I if you start talking to me about classes and selectors and all that, I'm not going to understand. So, you got the system prompt, you got the user prompt, and then after that comes basically as much of your conversation as it can fit starting from the bottom up. And so if your conversation gets too long, once it gets once it starts passing it on, it forgets the top part, which is usually the most important part because that's how you started the conversation where you tell the the AI what you wanted to do and that's when AI starts getting dumb for you. That's when people start having issues. So now if you think about the that context as a vertical list of you know of context that gets passed to the AI agents kind of work sideways because you can as you're going down the conversation you can send an agent to go on a side quest and go do some go investigate something and just come back with a one-page report. And so now the main AI that you have didn't get a bunch of text to fill up the conversation and more easily forget the beginning. Now I just got maybe a onepage report. Another thing that you can do with agents and it can get very sci-fi. For example, once I started getting into uh development, I was like, why not have Steve Jobs consult me? And so I have an agent that essentially impersonates Steve Jobs to give critiques of product decisions. And that's different from another agent I have that is trained only on Swift and Swift UI which are the Apple programming languages. And so now I can have when when **[00:14:00]** I make a new feature, I can have fake Steve Jobs review the feature as he would do it and then I can have the Swift UI agent review the code quality and they don't they don't talk to each other. They don't know that the other one's doing anything. And so that kind of keeps things tidy and also keeps things um focused and and and and and in their Yeah, that's that's a great way to explain it. So for those who don't know, as Jose just explained that context window depending on the language model can be larger or smaller. So, some higherend models have larger context windows or some newer models, some have smaller. And so, sometimes when you're chatting with them, this might have happened to you. It's it seems to be very good, but then it starts to sort of not be so good. And that is typically around the time it it starts getting close to that context window. And what some of these models do is they compact once they're they're getting there. they start summarizing the past conversations so they can keep track of it but still there's something lost there there's because it can't keep the full fidelity of it. So what I think what you're explaining, the way you explain it is really clear because this means I can have this conversation and maybe I'm using 10%. And then I send this agent to do some research that takes a lot of context because it's browsing websites. It's doing a lot of stuff. It has its own context. So maybe that one uses 20% of its context, but then like you said, it'll create a very simple summary for me. It doesn't use a lot of tokens. gives it back to your main agent who then can can use just that and it didn't fill up the rest of it. A good example is um any kind of layout issues that you have with with Swift or Swift UI. It's you know Swift UI especially because Apple changes it so much. It's not like that these models are trained on it right because the **[00:16:00]** the cuto off dates for when they train are are so old. So like for example the stuff that Apple released for Mac OS 26 is not on the right training models yet. So I can have an agent and go look at the entire documentation of Mac OS 26 which would kill the context window of the main agent just to get the answer of like so how do I make a table right? And so it can go look at the entire documentation, figure it out because everything changed and then just come back and be like, you do the table this way. And now the main coding agent just knows that and it doesn't need to get everything else polluted or forget. And that's why it makes sense to have like a coding agent for this particular stack or this particular language and then maybe you have a reviewer agent who's good at certain practices. So do you give it instructions and say like you are this kind of thing and this is what you do and these are the So yeah. So so Claude has this feature built into their their command line app and so uh when you initiate you usually initiate an agent you can choose it to be for the project or if you want it to be global and then you just kind of give it a description. So you just tell it in plain as if you're prompting a regular chat conversation. You prompt it what you want. And obviously you can be as as detailed or or as little as as you can. So for example, for for the example I gave you about Steve Jobs, it can be as simple as you are Steve Jobs and you critique products with an eye as he did, right? And the reason I picked Steve Jobs was not just because he was the head of Apple, but because that has so much baggage about who the person is and the AI will know how that person thought, how the person acted, how the person spoke um and so it can train it that way. But if it's for example um for Swift and Swift UI, you tell it okay and then you you because you don't for example for that one I specifically told that your model is not trained in Mac OS 26. So any question you have, you have to go to a website to look for the answer before you **[00:18:00]** give anything. So you can give some some information like that. You do it at up front and then you don't worry about it. In the uh in the last few updates, Claude has even given agents memories. So they write their own they write their own files and so they can remember uh you know I can ask an agent how we did something today and then in five days maybe we have the same bug pop again and I can ask hey how how did we fix that thing and it can remember oh we fixed it this way and then it it does it. That's really nice. I I don't necessarily want to segue that far yet, but I I really like how OpenClaw handles that where it has the little memories for each one and it's been very useful. We use it for for some hobby projects and things and reminders and precisely that that memory part is really nice because it's just like, oh, you just go to that one and be like, oh, let's push this code to Cloudflare. And I in a new conversation I would have had to explain you have to connect using the CLI use wrangler use this thing and connect and this is the environment variable and etc. But since it has this memory files it's like oh I saw my memory I know how to do it and it just does it What is your flow for like something that's agentic where you have agents? Do you start in a main conversation and then tell it use this agent, use this agent, use this agent or do you have it set up in a way where you just tell it what you want and it understands which agent or agents to use? So almost every I feel like almost everybody that gets into Claude with Cly ends up building their own uh plugin that handles their work. So I've built a plugin that has my workflow that I like and it's uh a very standard **[00:20:00]** software development flow. So uh I built a skill that it's essentially a virtual So that's how that's how any feature idea that I have I start as a virtual meeting. So my that that plugin depending on what the project is if it's a web project or if it's an iOS project or sometimes I even you know it's even a blank folder it'll ask it'll figure out okay it's it's a it's an iOS project and based on the code that I see I see that it is an iPhone app for travel based on that I think we should have these three agents in attendance for this meeting and yeah and so the reason I do I do that I do it that way is because uh if if we're in that meeting construct, it doesn't code because one thing you've probably noticed is those coding AIs really like to code and you can tell it don't tell it don't code but it read the wrong thing and it just like changes your code base. So keeping that in that meeting it knows to stay in the meeting and and not code and it lets me rant as much as I as I can and then it orders my thoughts. It orders my thoughts before doing anything. So I yeah sometimes I even tell it I'm I need to rant right now. like don't and so it knows how to do it. And so after after that meeting is done, I have a BRD, a business requirements that I then take to another plugin that I didn't make but I found called Lisa named after Lisa Simpson, which will make sense in a minute. But Lisa is a planner plugin. So she takes the BRD and then turns it into user stories like a developer team would turn things into user stories. And so Lisa is another plugin that um it uses that interview method. So it it it takes the the you know it tries to infer my intent from the BRD document that came out of the meeting as much as possible and then anything that it can't figure out it asks me. And that's usually when I get into like the really technical questions where I'm I just sit there and like I read it three times. I'm like I **[00:22:00]** I don't know what you're asking me. So explain it to me some some other way, right? Mhm. Uh and then after that I I have the specs of the feature that I want to build. And then I would say the centerpiece like the thing that really helped me accelerate is the next plugin which is an official plugin that that Anthropic develops called Ralph. And so I don't know if you've heard of it, but that that you know before Open Claw Yeah. before OpenClaw was the the flavor of the day. Ralph had had a moment. And so Ralph is essentially a plugin that I'm not even sure exactly how it works, but my understanding is it um sets its own tests. It creates tests and they fail the first time by necessity and then it just keeps coding and coding and coding and coding until the tests don't fail. And so sometimes it can work for 30 minutes, it can do 30 loops, 40 loops, and it hasn't finished coding because it still has this one little bug and you don't have to touch it. Usually you have to give it permission or sometimes it'll ask you but for the most part it runs And um and then it it finishes. After that I do a manual QA uh of it. So one thing one thing I've actually recently um incorporated into my workflow is I ask the AI to make me a checklist of things to test. And so it'll tell me we built these five features and so open this, click this button, click that but it tells me everything that I need to do and I just check off that that things And then after that we just I just kind of do another I have another plugin uh called Axiom that is specialized for iOS. So I use that usually to audit the code and and I just keep refining it until I'm happy with it. Wow, that's a great process. I really like the the session the meeting that you're doing. So, how does that work in the sense of a you you said you built a tool. So, did you build **[00:24:00]** is it like a little web app? You go into it and it has a chat interface and then it decides who comes in and then you you chat there or do you chat in the actual context of cloud code directly? It's all it's all in it's all within cloud code all there. So it so in the same way that the Ralph Wiggum plugin exists and the Lisa plugin exists, uh, Anthropic lets you make make plugins for Claude and they're, you know, they're essentially just little bundles of instructions that that the AI can can And then there's um I think this is a Unix. This is how this is how you know I'm not I'm not I'm not super technical. There's this thing called a bash loop, which I believe is a Unix thing. And so a bash loop is essentially it just goes around and around and around. And so the Ralph plugin uses a bash loop to go around and around until it until it finishes. My plug-in uh stays in the meeting loop until I say meeting adjourned. And so that's how I control it. That's how I control it. And the coding agent can decide based on the question you ask me. This seems like it's a design question. So we're going to ask the designer or it's a product question. So, we're going to ask the product marketer or whatever. Nice. Yes. Well, there's a lot to unpack there. The what a lot of people don't realize is that these these chat conversations with the LLMs are typically that session, right? And so, before they added all the nice quirks and features, it it used to be that you started a chat conversation with an LLM and it's always and actually behind the scenes it's still this way. It doesn't know anything, right? You send it a question and receives it and gets you back. You reply, it does not remembering what you said it. It's actually receiving the whole conversation again and then it knows how to answer and that's why the context grows so much, right? Then they added a bunch of stuff onto it because of how people have been using them and making them better and all this good stuff. So what happens is you tell it to do something, it might do something, but there's nothing **[00:26:00]** to tell it to keep going. like they don't have a they're not alive like us like and so that's something that I think uh for example openclaw or the Ralph loop uh did very well they turned it into a little bit more an autonomous process because they can run like a cron job on the server which basically is just here's the message and here's some memory and then do your thing and when it finishes it just goes in again and says hey this is where you were this is what you did and it can it could just go on and on and on and on because there's something triggering it to do it again. And a lot of websites work that way. When you have any website that is constantly fetching new information from a service, getting new times, getting price rates, stuff like that, they have to do it in a background chron job. So, it's all like you were saying, they're they're like Unix internals that are very common to the operating system, but that weren't being used in the LLM space. And now that people started using them with the Ralph loop a lot and it did explode. I remember the the the like I mean press in the space of course but the media and X and everything talking about Ralph and on YouTube and then it sort of went away with open claw but it's still it was a huge thing when it came out. These these tools opened our eyes I think into what it would be like in a more agentic world. And now you see codeex and uh anthropic add these features. Some of them already had it but not going to use them. But now with scheduled tasks with co-work like they're starting to add these into sort of the default stack. Yeah. I I would say for you know obviously going into the command line is is pretty intimidating and you know even I've been a Mac nerd for 30 years and it was still a little bit intimidating to go into the command line but the stuff that they're doing with claude **[00:28:00]** co-work is bananas and so I you know I I I usually have three terminal windows with cloud code open at the same time but I go to claude coowork for a lot of things still because the the interface there is so good. So if you want to get a taste of what you can do with with claude, I think co-work is a really good place to What do you like it specific? Like what do you like about it specifically? I think it's because they are tailoring the use case so much for business. Mhm. Um to give you an example, every week I have to prepare a weekly report and for the last month I've been able to just it's the same report just numbers updated and you update the insights. I've been able to just upload a red line, you know, here's the here's the report with the new numbers but no new text. Here's last week's for context. Can you do your best job at at updating it? And it's not like I can send it over, you know, I can't send it over blind, but a report that usually was finished by Tuesday evening is now done by Monday afternoon. So, it saved us like a day and I think one fewer person has to touch it now. Wow. And is that scheduled? Does it run on its own or I haven't, you know, I haven't. I've been so heads down uh with Ralph and developing that I haven't played with Open Claw too much or with scheduling. So, I haven't played around with the scheduling, but it's definitely on my list of things to try. Right now, I literally just drop Monday morning when I get the template. I just drop the report into co-work and give it last week's. And now I actually just have a folder where I just keep dropping the reports every week and I just tell, hey, the new one's in there, Uh so, I think I'm I'm I'm very ready for it to be scheduled. But more so than the scheduling, I think what's impressive is how good the writing and the insights are. Yes. Be it's not **[00:30:00]** just, you know, cuz like I think in the beginning, um it was just regurgitating table, right? Like and anybody can just table says cactus 20 c 20, right? Anybody can can regurgitate stats, but it was actually uh especially once I started giving it older reports, it started telling me like, hey, you know, this trend or or there's this angle that you haven't considered or it just started popping out a lot more interesting things. It kind of gets rid of that the the level one analysis because it's so easy and fast to do that. You can start getting into deeper That's an amazing point. the power that these tools have. Sometimes we we forget because we use them for basic things. But they can do very understanding of knowledge and calculations with code of course. They they can abstract data or information from data which is which is very interesting because like this morning or last night I was we were working on a voice agent and we gave it a price sheet and we were testing the price sheet so it could respond with prices and then I asked it something very basic. I just looked at the price sheet and just asked it something really basic. one of the terms there just to get the price back. But it said like, "Okay, so here are your options." And I'm like, "Options?" But and then it started telling options. It was like that one. But then it also read ahead and there were like three or four more instances of that word within the context of something else. And it was like if you want this but with this variation, it's this much. And if you want this thing that includes it as a bonus, like it went deeper. And I was like, "Oh, wow. Okay, thank you. So, No, I I think **[00:32:00]** so. I think the most impressive thing that I've seen Claude Co-work do, so in the in the app that you can download, uh a few weeks ago, they also launched a plugin for Chrome where it can control Chrome. And it was a little scary, right? But I just wanted I just I was just curious to see what it could do. And so, I'm going on a trip to Japan for my 40th birthday in a couple of weeks. And I don't know what got into me, but I just decided to let it wild on my Gmail. And I said, "Hey, there's my Gmail. I'm going to Japan in a few weeks. Here's the dates that I'm going. Uh, you should be able to find I didn't even tell it what to look for. I said, "You should be able to find plane tickets. You should be able to find hotel reservations. And you should be able to find like event tickets for places that that we're going to go to." Um, also told that be careful because I've been wishy-washy with some hotels. So, you might some hotels you might find that I actually already cancelled. It took about 20 minutes. It went through my entire Gmail. It found the hotel reservation emails. It made a list of them. It even told me, "Oh, but these ones you canled." Um, it told me, it figured out all the ticketed events that I had. And then I told that, you know, oh, I don't like to shop. We're like, you know, we want to have some some our my travel style is to go fast fast and have rest days versus having a ste. So, I told it a lot about how I like to travel. I told uh my brother's coming with me and he doesn't like to do this, he doesn't like to do that. This is the food we like, this is the food we don't like. It built the most incredible uh travel itinerary that it then uploaded to a website that I was able to send it to my brother and it had like even little touches about like hey if you want to splurge on food like you should go to Kovi because you're going to get really good beef there and like go to this restaurant and that's where you're going to get the like it was so thorough and so impressive. **[00:34:00]** Um and like I said it was literally just said look at my email and see what you find. Wow, that's impressive. And the the amount of time, effort, and money that that would have cost you to have someone do for you. And also, a lot of it couldn't be done by anyone because you would give a travel agency your email access, right? So, there's a lot of hidden data. I'm not sure you should give it to AI either, but yes, of course. But at least, well, you have to decide who you trust, right? But the way it can get access to especially like yesterday two days ago they released Google released a CLI I don't know if you saw that that it now provides easy access to all the Google tools to an LLM. So instead of as as you know if you built an automation in an A10 or anything like that it's it's hard to get everything set up and the a and all the scopes and the instructions and and permissions. They just released this thing that just lets LLMs, if you give the right permissions, just do everything on Google, email, spreadsheets, documents, calendars. It just understands everything and it tells them like an MCP kind of deal, but but it's still like and so it's amazing because now you can do this programmatically as well. So I what you did is a great example of how this sort of personalization of software is evolving because yeah I mean you're giving it to the eye but you're not giving it for training etc. It depends on what models you're using and who you trust and all that and what you believe but at the end of the day it's just processing it and getting rid of it. It's not like it's storing all your data. Now could there be gray areas? Sure. But again, if you trust Google to keep your email, I mean, **[00:36:00]** keep your email, right? That was kind of my thinking. They already have it, so why not get some use out of it? Yes. Yes. So, so at least now it's not just for the marketing companies, it's also for your benefit. So, the the way these tools are now useful and a real dayto-day thing. Yes, of course, we're business owners. Definitely everyone knows that. But also in the day-to-day like you just showed in your example, you can have it make you a day itinerary. You are traveling just to get to the city and you know which points you should visit or even you you might not even know which points to visit. So you can have it go on Yelp and figure out using the browser extension or or MCP and it could go figure out which are the most popular places and then build a little itinerary for you like it like it did for you, right? It's it's really impressive how we can use them, but the speed at which they do it and the cost is what's really ridiculous. I mean, at least in the business side of things. Well, yeah. I think I think the the cost is to be determined because right now we know we're in a heavily subsidized era of AI. So, right now I'm trying to get the most out of it before it it ends. But I think to pick up on the other thing you said, the speed, um I think what what people haven't realized about the speed yet, it's not just that you can do things really fast. It's that the cost of starting over is almost zero. Yes. And so one thing that happened with me when I was doing for example the meta uh the meta tool to evaluate ads that I told you I started that tool over from zero just deleted the project three or four times because by the time I hit certain walls as I was as I was learning by the time I hit certain **[00:38:00]** walls I had learned so much not just about how to code but about what I wanted that it was a lot easier to just start the process over and tell the AI exactly what to build me again because in the end it wasn't it wasn't that um what I was asking for was that complicated. It was getting that intention out of me and being able to communicate that you know kind of getting the ideas out of my head into a way that they actually understands me. That was the hard part. Wow, that's a good unlock there. regardless of what you think of it, it's actually a very helpful tool to help you understand what you really wanted, which is very doesn't make sense, but but at the end of the day, having to explain things to it sometimes like a kid or like an intern and be like, "No, but I mean do this this like having to do that sort of makes you do that whole teaching cycle that is widely known that helps the best way to learn things and to understand things is to teach them right so it gets you into that sort of cycle and I find at least for me it's been it's been a way to sort of oh now I I understand what I had wrong in my in my like I said a lot of times when I hit a wall I'm just I I just ask it to explain it to me in a certain Or for example, I think I mentioned earlier with that meta tool, I kept hitting a wall with uh pulling the data from from because that is a very that is a very complex thing of hooking up to the meta API. So I kept hitting this wall where I I was either pulling in too much data and timing out or I was calling the API too many times and and meta was blocking me because I was making too many requests. And to make a long story short, I **[00:40:00]** started working with the AI in a And we came up with this metaphor together that the ads were like people at an airport that in the data they were picking up their luggage. And so just through that, explaining that metaphor of like, oh, and then people walk here and they pick up their luggage here and they go there and then do they go home and then come back for more. Just through walking those scenarios logically, I was able to abstract the more complex topics that I was hitting. For example, in that case, it was something called concurrence, which is how processors manage threads in a program. And so, it's like a very complex like how how do you handle the actual silicon on the computer that the AI was able to walk me through just by an analogy that I understood. That's amazing. Using analogies and metaphors is something I haven't heard a lot in the space, but it makes it makes perfect sense. At the end of the day, you're talking to a language model, So, that's the Yes. So, if you think about it from that perspective, right, uh LLMs are just big autocomplete and so they know literally how to talk like anybody to anybody about anything. So finding the way that you best talk to an LLM is the best way to get an unlock uh on how to on how to better collaborate with the tool because then you're speaking the same language. And another benefit that I've discovered is having that common language sit outside of what you're working on actually helps the LLM pay more attention and make fewer mistakes. So, for example, one of the allnighters that I mentioned, uh, it was with Lovable, and I kept getting into this, and I don't know if you've had this, where, in this case, I think it was a a box needed to move five pixels to the left, **[00:42:00]** and I just kept saying, "Move it. Move it five pixels." And then it did it, and then it didn't do it, right? No, no, no, you didn't do it. And it just back and forth, back and forth, back and forth. And out of frustration, I don't know what possessed me, but I typed, "Sometimes I feel working with you is like walking that dog from up, but without a leash, and you're just seeing squirrels everywhere, and you just go off running, and you chase squirrels everywhere." And when I said that, it instantly was, "Oh, I understand. You said that because of this, and you told me that because of this, and I behave like that because of this, and I should fix the behavior." And it just understood, and it fixed the bug right away. Oh, that's funny. And so the the conclusion that I or the working theory that I have is that the the and I even built it into my plugin where I have I have a command called squirrel and so I can say hey squirrel here squirrel there because it kind of stops it in its tracks like squirrel that's a weird word for you to say right now right because like they're all pattern matching and you break the pattern like squirrel what's that? Um, and so I even built that into into the code. So for example, instead of writing comments on my code, I have a pilgrim that protects sacred code. And so it's like it gets like very weird. Like I said, it gets very weird and like sometimes sci-fi, sometimes fantasy, but I've had in other projects, random my eyes, like you said, a brand new session, never talked to us. AI goes into one other one of those coding runs and it stops and was like, "Hey, I was about to change this, but it says this weird thing about a sacred place like what do I do with that?" Even so that prevented a bug. Yes. Yes. Oh, wow. Interesting. Yeah, that's a great way to go about it. And and again for for our listeners there's all these tools also have especially like the CLI based tools like cloud code and codeex and they they have their like claw.md file and their agents file and **[00:44:00]** things like that where you can set certain rules certain instructions certain things about that particular project or you can have like global things and so you don't have to repeat everything every time. There are things like what you're saying now that you can sort of embed into the embodiment of the thing so that it knows to use those or or even have those restrictions like you were saying uh marked in a way that's different and you can decide how they should behave like I'm not saying it's best practice. No, no, but it's but it works for you. And that's the thing like I think you you you said something very important earlier that everyone who starts playing with these builds their own like little plugin or they build an orchestrator or they grab someone else's and they sort of adjust it. It's because we sort of want to put our own flavor to it. And this reminds me of the old IRC I mean you're about to be 40. The IRC uh era where where you would we would chat through IRC. It was this thing. You had a program on your Windows computer cuz nobody had Macs back then. And you would like type something. You could you join a channel and you could talk to random people in random places, but there would be channels for like cities and stuff. But people would mod these. They would theme them. And so everyone had their own not everyone, but people who got into computers and stuff, they would go in and modify. They would skin their their program and it looked different. then it had little emojis or things or or asky characters, not emojis, but um and and the same happened with Winamp when it came along, right? So, I think that's sort of coming back where people are starting to use these tools, but then they want to sort of make their own flavor of them so that they're not fit into a box necessarily. Yeah. And I think because they are so text based, it's a lot easier to just mold them because at the end of the day, it is just communicating. And I think that's one of the things uh having worked because I was I wasn't **[00:46:00]** like AI skeptical but I didn't understand where it was going to go, what it could do. And you know there's a lot of talk about effects on jobs and what that's going to do and and I do think there's a lot of concern there. But the one thing that I've I've come away with is the one thing that AI will do is it will make creativity that much more And I think even the models themselves will need it because one of the things that that people start talking about is they're running out of things to train the AIS on, right? Because they've already trained it on everything, right? So the only way to the only way to give it more training data is for people to be more creative and produce more stuff and write more and make more things. Wow, that's a that's a great concept. So for it to actually evolve and become better, it still needs the humans to be creative to start generating things that they can then learn from. You know, I'm listening to the origin book from Dan Brown again, uh, after many years, and it's funny because it fits perfectly with with the current state of things, like there's this AI and and, uh, it can talk and change voices and things like that. And it's very smart and and it was really sci-fi even a few years ago when it was written. It's not so old. It's not like when you read Asimov stuff from the ' 50s, which is like way ahead. It's still like more of a guard even more so than even more so now I for sure. Um but like this this particular book it's like you hear it now and it's like it's perfectly possible like many of the tech stuff that uh seemed just a few years ago completely impossible and it's just LLMs and that's another thing I think we don't realize. We're just being exposed to LLMs, which is a very basic version of AI in the general scheme **[00:48:00]** of things. There's way more advanced AI, like the ones that's driving self-driving cars and things like that. Um, and in medicine and other there's there's a lot of uses of AI and robotics that we're not necessarily exposed to that are way more advanced, but we're seeing this and we're interacting with it and it's amazing. So if you can sort of step out of that context and think what's going to be possible with AI in general in just the next few months, it's it's mind-blowing. Absolutely. Um yeah, I think I I always dreamed of making apps and I you know I I tried and flunked out of a few Swift courses like I was never able to I just don't have the patience to code. Like I'm you know I'm sorry about that but this has Yeah. Like I said, it's it's it's about how well I can express myself to the AI and and knowing what to ask for. And and that's really been such a game changer and and yeah, it's it's very weird. You know, people people of people say, and I didn't believe this, people say, "Oh, you're just going to be able to speak and and an app's going to appear." And I thought that wasn't going to be possible, but the more I think about it, you know, I think it's come down from developers down to people maybe like me that are that are technical but not developers. And I think it's just going to keep coming down uh into more and more uh you know less technical use cases. Yeah, that makes sense. And I think that was the point like computer programming languages weren't for developers. They were just that's how we could be able to talk with those things that we called computers and they started evolving becoming more advanced and things like that. And of course that a lot of people didn't weren't interested in that. So, so then the developers kept that in their realm and everyone else just moved on and they just used developers and software engineers to do it. But the the real purpose I think um that people like AI lovely uh had where like **[00:50:00]** a thing that I could communicate with as a person and now we have this thing way of actually talking to it in natural language. I think it's amazing. I I I have a lot of fun with it. And I think like you, it opens up the creativity aspect for a lot of people to actually start building new and exciting things that no one ever dreamed possible because now they just Jose, this has been super interesting. I really like how you've been managing everything on your end and your your background coming into this just just tells us how it's possible for anyone to to just actually just build stuff if they want to. So, I really thank you for coming on the show and if there's people want to know more about what you're doing, what you're doing money transfer um and your apps, etc., Where where can people find more information about you? Where can they reach you? Yeah, you can uh pangeamoneytransfer.com is is the main website. And then um my own apps that I'm building I'm building out of It's a pplac.com. Applicat.com. Okay. I'll I'll put it in the show notes. And what what are your apps about by the way? I have one is uh when you work with cloud code, it outputs a lot of markdown files. And so I kept getting so many I had so many uh finder windows open and so many text. So I had I built an app to just kind of look at all my projects markdown files together. And now I'm developing it so that it's interactive on the way back. So now, for example, I can tell it build me a QA checklist. And so I open my markdown app and I can just click click click click click as I test things and and I can leave feedback for the AI right there without **[00:52:00]** having to again pollute the the conversation context. I just leave a comment in the file uh for it to um and then the other one that I'm hoping launches this week is um you know how you have app folders on your iPhone. I basically built that for the Mac OS doc so you can have a little button on your dock that opens up u and you can have up to 16 uh bunch of groups but up to 16 icons at the same time. Uh so you can open apps. That's neat. It wasn't possible before like they can they don't you can you can put a folder for example I've it's a it's a Mac hack I've used for years that I drag the entire applications folder into the and it but it you can only have every single app show up once and Mac doesn't really Mac OS doesn't really like if you move apps out of the application folder so it's not like you can group them easily. So I made this so that you can group them easily. So you can have like I can have Discord in three groups or I can have only three of them. And so yeah, it's it's a little bit more Oh, that's a neat one. And that's the sort of thing like I would never expect someone, especially someone not a software engineer, to be like, "Oh, I'm going to build an app that extends the hon." Yeah, that was just that's just an annoyance that I've had for the years. I use the Mac. That's it. That's just it that now people can just go and do and fix those annoyances that you have with Yep. So yeah, that one's called Doc Pops. So that one's going to be out soon. It's docpops.com. Nice. Nice. I I'll put those in the show notes as well. Thanks again, Jose. This was super interesting and super fun. And if you're listening and you haven't subscribed yet, please uh like and subscribe so that you learn whenever we have new episodes show up. Thanks again and we'll see you in the next one. Thank