I Built Website Features Live by Just Talking to My AI Agent

**[00:00:00]** Hello everyone. My name is Armando Piscareno and welcome to the web talk show live. Today we are going to be building some features into a directory for moms for local events, places, etc. It's a very fun site and I wanted to add some features this morning but I thought why don't we just add them live. So what are we talking about today? We are talking about so many things. There's been a influx of if you are any sort of way related to technology or software there's so much information around the LLMs what they're doing how fast they're progressing whether coding is done for whether developers are done for and a bunch of other things and depending on what narrative you're following you might be a little confused so my plan today is to show you why AI is so good for small businesses or for everyone in general, but in the specific scenario of if you want to build something, there's absolutely nothing stopping you from building something today. So, I am going to go through a list of features and I'm open to any questions. If you have any questions, comments about what am I doing, how am I doing it, and then if you actually have ideas for features for the things that we're doing right now, you can tell them to me live and I'll if they make sense, I'll add them so you can see how fast all this works. Give a little background before we get started and then dive into what we can actually do. But it is tremendously interesting, I think, where we are today with these tools. So, let's start off with the basics. LLMs, large language models, allow you to talk to a computer and it might give you a response. And we've talked about this many times on the show. LLMs evolved and got very good at coding. So, developers started using them for coding. And then platforms came along like Replet and Lovable that allow **[00:02:00]** you to just build full sites or applications directly just by talking to the thing. And all that's really nice, but then some people start saying, well, what happens to the code? like, but nobody's checking the code. Is it good code? Is it a mess of code? Well, at the beginning, it might have been a little bit messy. Nowadays, the models are really good at doing these things, especially if you use top-of-the-line models like Opus 4.6 from anthropic or Chad Codeex 5.2, I think 5.3, and they're they're tremendously good at coding. So, the barrier today is no longer knowing code or understanding code. It's whether you have a clear idea of what you want to build. So if you know exactly what you want to build and you can speak a language like English or Spanish or French or German or whatever language you speak, then you can build something because the LLM is the one that's actually going to build the thing, not you. So, it's very interesting, especially again if you use one of the better models, but then also if you mix it with some of these tools that allow them to go on their own. So, you probably heard about Maltbot, Open Claw, Clawbot, and all its different iterations. Currently, it's called Open Claw. It's an amazing tool. It's basically just a wrapper around whatever LLM you want to use. For example, let's say you use Claude from Enthropic, which is what I use. Well, I can use Cloud in a multitude of ways. I can go on the website cloud.ai and I can talk to it via the chat interface and that works and it's it's very good for many things. And I can also use it in the app on my phone. So, I can also ask things and it'll give me a response and it can help me with spreadsheets. It can help me check an email. It can go and research. It can do a lot of different things. It's very powerful. Then in the computer, there's also the co-work feature. And then they have the cloud code which also allows you to use it within your coding environment. So I've talked about this as well and it's not **[00:04:00]** the topic for today. But if you use it sort of as a pair programmer or if you use it as a helper within your codebase, you're building code, you're doing something, you find a bug, you can ask it, hey this happened, can you help me figure out why, how it happened? And it is very good at finding edge cases, finding things where you didn't realize what was going on. So very powerful in that sense. Now the problem with these and it's only a temporary limitation. They're they're they're doing things about it. But the problem I would say is that these like Chad GBT or cloud models, you can give them a task and they will do it. But it's typically more of an interactive experience. So you cannot just say go forth and build an app for me and do and it'll do the full thing. it it doesn't necessarily work that way today. Right now they're adding things and if you give it a good enough description it will work through a lot of things but it will sort of stop at some point. It won't it won't do the whole thing. Now there are other tools like manus that that will work through it and replet and lovable have their loops in there as well to do these sort of things and and it's really good with opencloud. What OpenCloud did is that they built in an open-source package or program that you can install on a computer. And I don't recommend putting it on your main computer because of security risks and you have to do your research about this, but it's a very nice tool because you can install it on a computer like an old computer like the one I have back there and it's isolated on its own and then you connect it to a model like Anthropics Claude. I mean, yeah, Claude Phobus 4.6 for example, which is what I use. And then it allows you to talk to it via your phone, via Telegram, via WhatsApp, via Slack, via Teams, whatever you want. So then it turns into more of an assistant kind of tool that you have with you at all times. You can be on the streets, you can be at a park, you could be at the movies, wherever you are. You can send it a message and it's like you're talking with the cloud or with the Chad **[00:06:00]** GBT on your computer. However, it keeps a longer term memory. So it knows more things about you and the project that you're working on and it's not stuck in one project. So typically in cloud code if if you develop software for example what you find is that the history and the sessions are linked to a repository. So if you're in a specific folder that's where you have the connection to the the cloud code. If you move to another repository that history is no longer there. You have to sort of find it and then bring it in if you want. That's a possibility, but it's not it's not like always saw and always there. Why what I found with cloudbot or openclaw or moldbot currently openclaw is that you can just tell it to move between them. So you can be talking about let's say mom's guide which is what we're working on right now and then you think of something for another project and you say oh by the way can you go to my website and adjust this thing or change the footer or we added a new blog post. and you put something on the header and it'll go to that project because it'll know it can just move over to it, make the change, publish it if you want and and that's it and it works. And so it's really powerful in that sense because it can just go back and forth between different things. Now the issue here is that with traditional ways of doing it again you have to stick in that repository or you have to stop or it has to stop you get that response back then you have to answer again gets another response etc. So, OpenC Claw and these sorts of tools, and I'm sure there's going to be plenty more, and there's already some coming out, and people are using them, it allows you to tell it to do something, give it a plan, and it'll just go forth and do it. And it's very exciting because you can tell it, I want to build this application that does **[00:08:00]** these things. And then you can have it build a plan. If you're specific, the the more specific you are, the better. it will actually go and do it task by task. So it will start with the first task and build it, start with the second task and build it, start with a third task and build it, etc. And it just just works. It's very powerful because you can have it do things overnight, but also you can have it just work with you as you are um doing your work. So it's really nice. So what I wanted to show is how fast it is in real life. How fast you can actually build things. You can actually do things within the tool because it it's it feels like science fiction still to me to this day as we're using it. You literally just just talk to it. So here's what we're going to do. The idea is I'm going to share my screen here so that you can see how the site is and then we're going to think of a feature or a change that we want to do and this is we're actually working on this particular project and we might want to make some adjustments to the thing. So I connected it to the the code and it has access to do things that I ask it to do. So let's see if I can share my screen here and again if you have any questions throughout the stream please feel free to drop them in. So I'm going to do share screen and then I'm going to share this tab and I will do share. Okay. So you should be able to see this here. Now this site is a directory for local activities and events and places for moms with kids. Right? So, it's here for San Antonio and so you it's it's a way for people to find different things like parks, um museums, caverns, right? Anything you want, right? You can go in and you can see more information about the things and you can see things on the map, etc. And **[00:10:00]** it has a nice little places search where you can find different places and you can zoom in, zoom out, and you can filter by different things like museums, price range, etc. And then you can find them in the map. And it has the same sort of thing under events where you can see all the events that are happening today. And it has this little nice map that shows you u based on what you're showing here where the events are, what's happening, etc. And more details. And it's colorcoded to things that are like closer to you in time like green and just gets gets further away with the different colors. You can see uh that sort of scheme here on the left. That's really nice in that sense. Now, one of the things that we were looking at is, okay, so how do we make it as easy as possible for people to search for things? So, if I go here, for example, I can see that I can search for something like witty museum. And that's that's straightforward. It should find the witty museum, of course. And then if I do dinosaurs, dinosaurs, it should find things that are related to dinosaurs, right? including the witty museum, right? I didn't specifically say witty, but it understands that there's dinosaurs because of tagging, etc. So, it's good at that. However, we found that you might want to search for not only places, but events as well, and not necessarily one or the other that you go to places or go to events. We were thinking, okay, how can we make it such that it's easy for someone to find both things at the same time? Is it possible to have two completely different types of objects sort of portrayed in the same way? So, we're building out this little page called explore. Let's see if it loads. Yes, here it is. And so, this explore page is now showing us events that are here right now. So if I do free events this weekend for example, then I see events that are happening **[00:12:00]** this weekend and then I see places. That's a start. It's good. So we can see both things at the same time. But if you find I don't know if you notice if I did free events this weekend. It's actually understanding that I'm talking about the weekend and it's applying the weekend filter over here such that it will only show me things that are re related or relevant to the term. Also, it's filtering for whatever I said this weekend or today with one of our actual filters, which is pretty neat. Now, still, if I if I do this, I I find that the events here look okay, but then the places look odd because the places, I don't know if you recall, have a split column view. So, it would be nice if if we have this sort of same split column view if we're using the explore page because otherwise they look I don't know I think they they look too big. And then also I I'm not too happy with these. We're experimenting on having these little pills that you can click and it'll give you that as a search, right? Uh so it's sort of a a way to to to show people what's possible. So, okay, that that's sort of nice, but I'm not sure I like the little pills. At least I don't like them now. So, I'm just going to tell my agent here. I'm going to say explore page looks good. Let's remove the pills for suggestions like splash pads, etc. Okay, so what's happening now is I told it I told the agent exactly what I just said. Explore page looks good. Let's remove the pills for suggestions like the splash pads, etc. So, I that's what I said. I I just said in plain English that I want to remove those little things right here. Okay. So, typically what would happen in a regular scenario is we would get a request or we're building something and people say, "Oh, can we change this on **[00:14:00]** the website?" And so, we would go in and and look at the website and and if we built the code, well, we will have access to the codebase and we would be able to then say, "Oh, okay. Yeah, we we have to remove this section. So, I have to go in here and I have to find what's the name of this section. Where is it? Oh, it's uh this one here. So, okay. So, let me look at the codebase. Let's find where we have it. I would find the code and then I would comment it out. I would delete it. I would remove it. I would do something with it. Right? So, that's that's the whole thing that we would typically do. So, again, a person would have to go in make those changes. Once those changes are done, then they could be tested in a staging environment, local environment, and then we look at them, they look good, great. And then from there, I would push them to the server, in this case, like a preview um server, right? And so what I told it to do is remove the pills for suggestions like splash pads. I don't know if you know if it came through the microphone but I I heard a sound bling which was the uh agent sending me a message through telegram and it says done suggestions removed and the preview has been updated in explore v2 whatever which I think is the same page so let's refresh the page and that's it they're gone right so it's I didn't do anything I just told it that it should do that and it went ahead and do it so now I'm going to go ahead and tell it something else like and and again this is already pushed into the cloud. It's already running and it's running on a preview instance. We're using Cloudflare pages. Um and so it's running in a preview instance that works exactly like the main one. So if I want to make a domain one, it'll just make the main one and now that would be live on the live website. very powerful because it'll do everything that I've talked about, but especially if you're doing a production website, then you have to be careful not to touch production without um being without **[00:16:00]** actually having tested it before. So, this whole thing does it locally, then pushes a preview, then tests it, make sure it looks good, and then it can publish it to production. You can have whatever pipeline you want. You can have it where it actually opens a browser, goes and checks, verifies, and then makes a PR to your or pull request to your main branch and then that way you can receive a message or someone on your team can receive a message to review it. Oh, it looks good. Then publish and then it can go live. So now that I know that it did it, I can tell it, yeah, push it live. It looks good, right? Which I'm not going to do right now because we're still working on it. So another thing that we might want to do there is say okay so we were talking about these these sections right so I'm going to do something like museum okay right and so just to filter it out a little and what I want to do is show it that this looks good but I'm not I'm not very happy with that bottom part because it it looks too and we're experimenting here right we're live but let's see so what I'm going to do is I'm just I took a screenshot and I'm going say, I'm going to paste a screenshot in my chat and I'm just going to say events look amazing, but places with full width look a bit odd. Can we make the places show in two columns just like we have in the places page without changing anything else? anything else. Okay, so I sent it that and now it's going to go through and it's gonna say hm. Okay, so to do that I'm going to need to change column widths, change the container, something, right? Depending on on what its approach is. And if you could see the chat itself, it it would be saying those things specifically. Oh, I'm going **[00:18:00]** to look at the codebase. I'm going to compare it to the events page. Then it goes to the places page. It's going to check what what it is doing and then um it's going to try it out and then if it has issues, this is the fun part. If it has issues like it checks the code and it sees that there are some TypeScript errors or there's some compilation error or anything, it'll show you it start process and say it's it ran but then it had these issues or these warnings. Let me check what these are. Oh, okay. Hey, the comma was doing whatever or the regular expression wasn't working properly. Let me fix that. And it goes through and it fixes. So it iterates on itself to make the thing do what it should do. So now it says done. Places now show in a two column grid preview updated. So if I refresh, I don't know if what if that was a sort of a big ask, but let's see if it did it properly. So let's let's filter out again something like museum. And there we go. Right. So now we have these and then we have this in two columns which then we would expand and say well these all should have pictures ideally in this directory. We have still not pulled all the data but uh it helps us see hey this one doesn't have a picture and it looks a little odd. So we should do something where we have a placeholder image that we show there when um when it does not have an image. So we can actually see how we're doing it here in places. And what we did here was if we don't have a picture, we actually remove the picture section so that these look like this. And so it's more of a masonry layout than a um side by side. Right? So here we were looking at them and it was side by side, right? So I think we could do the same thing over here. So I would say can you So what I could do is I can copy this paste in the screenshot and then I would go to **[00:20:00]** places and I would do the same and it has access to the whole code right I don't actually send have to send it all the screenshots but the clearer you are the better right so I can say nice always nice to give him some kudos and then I can say something like I like how on the places page We remove the image section if there is no image and the layout is more of a masonry layout. Can this be achieved here as well? So you can be as nice or as not nice to your agent as you want. But I found and actually research has shown that it's better to be nice to them and you get better results and better outputs if you're actually kind to the agent. I expect because the way it works is very similar to how humans interact. And so if you guide it in a negative way then it might sort of interact in the negative way. Now that's not to say that you can't you can tell anyone anything you want. It's just my experience what I found and what some other people found. Some people are mean to the agents and I mean to each its own but I don't think you should because you shouldn't be mean to anyone regardless like if it's a person or an agent don't get frustrated just think of the agent as if it's another co-orker employee friend or whatever and and it's better I think it's better for mental health and everything as well just like chill and so they they can make mistakes the same way now they won't get offended or hurt or anything But regardless, I I don't know. I think it's a negative energy if you're you're sending a negative energy wherever. So, as I was talking, it did the thing. So, C CSS Y. But we can get the same effect with CSS columns done. Places now uses CSS without images will be shorter. Okay. So, let's see. Let's **[00:22:00]** see if that's true. Let's see if it understood what we were actually saying. So, I'm just doing the same query here. And yeah, there it is. So, it did the same thing as the other one. Now, as you saw, I was talking. So, either I'm very good at talking and coding or or the thing did it right. So, it's it's very powerful because you can have a conversation with someone and we've done this for for for some like this this little thing up here. Did you know the element was established as a Spanish mission? So, if you click it, it shows you the full tip because some of these might be a little longer, especially on mobile. So, you can tap them and it'll show you the full the full thing. But then you can you can swipe through them. But this the first iteration of this didn't have the little pop-up box on and then on mobile. So the thing what happened is that it would it would cut off. It'll truncate the last part. You couldn't see it because it was too long of a sentence. So in having a conversation, we said, "Oh, can we do something where maybe you tap it and it shows you more details?" And I just asked it on my phone, mind you. can you do that? Can can you add a little tool tip or something? And it went ahead and it added the little tool tip section. And you can actually swipe the tool tip if you're on mobile, which is pretty neat. So very powerful if you think about it, what you can do with this technology. It's LLMs have gotten very advanced and if you connect them to the right tooling or if you have something like open claw here and you do it in a safe and secure way and you give it certain instructions and restrictions um due diligence right like it can't do recommendations on that part. It is extremely powerful and it opens up it unlocks a world of opportunities for you because regardless of if you're a software engineer, a coder, a business person, a chef, **[00:24:00]** whatever you're doing, if you need a tool, it can be built. So you can have it build it or you can have someone like us build things. And the speed at which we can provide new tooling is like nothing that has ever existed in the past. Now it finally makes sense to have ultra personalized experiences, ultra personalized tooling for your business. So, if your front desk staff is always filling out a spreadsheet whenever someone calls or a block of notes or they're writing things down and they always have to then move that over to another system or they have to send an email. All that work that is constantly repeated within your business can either be automated or just simplified. The data capture step could be greatly simplified in most cases, whether it's for the front desk, the back office, the people who are actually working in whatever widget you're building or service you're providing. Even in the auto industry, we've talked about like car shops. You could have now tools that are specific to how you do the work. And that's really powerful because before you would need to fit yourself into a box for what Soho does or what Monday does or whatever even air table like Air Table is such a great tool because you can sort of build out your data structure the way you want it but it's still within their system and within their constraints and you might not like that sort of input structure but still I mean it's a it's a huge leap forward versus the old traditional systems. But as you may have heard, SAS software has taken a dip recently, a big big big billions of dollars. And that's because people and investors are starting to **[00:26:00]** realize that now you can build anything. So if you hire a company like ours for example to build tooling for your business that's specific to you that follows the exact process that you already follow that you've followed for 20 years and it just puts it in a nice interface for you to do the thing without distractions and have that then move the data over to your different systems and just keep everything the same way you have it but but a lot easier to do data capture uh a lot less human error etc. and it automates all that process, then you'll save a ton of time and you can get a turnaround in like two weeks. That's ridiculous. We've been seeing that we can build things that would have taken months, 6 months, a year, and they're ready in a week or two. Of course, you have iteration cycles and testing and all that good stuff, but if you have a clearly defined scope, the speed at which these tools can be created is is amazing. It's it's there's just so much that that can be done. Anyway, let me stop sharing here and come back to the screen. So, the I think the power that we have today should get us all thinking and not whether AI is good or bad or or if I should use AI or not. is how should I use AI? Because it's here to say how can I use AI to help me achieve the things that I otherwise would have taken a lot of time to do. So think about that. Think about the things that you're currently doing in a repeated way. Things that you feel you're losing a lot of time in. Things that are just frustrating to do because you have to do them again and again and it's just draining. All that can very likely be automated and then also with what exists now you can probably set it up in a way where the interaction is so much simpler. Now I'll leave you with one last thing which is just the **[00:28:00]** communication part of it. So we're building tooling for from businesses where their staff might be in the field, they might be doing installations, they might be doing service and they cannot go into an app and enter data into an app. They they already communicate with the people in the office by messaging Telegram, WhatsApp, Teams, whatever. So, we've been building these agents that allow them to keep that same channel of communication that they're used to and just chat with the agent and just say, "Hey, I'm here." And then the agent automatically understands, "Okay, that person is there." Checks your location, sends it back, puts it in the database. This person checked in, they're in the right place, and they're actually there. And then throughout the day, the person can say, "Okay, here is my evidence." And it could go in through, shoot some pictures, send them over, and then the agent will know where those pictures should go, what project that is based on the GPS location. and then do a bunch of other things. Let let the supervisor know to do this, do the visit, etc. And it's all handled by just regular communication. And of course, that's just text, but we can put in voice. We can have where someone can call and we can have the agent understand it and then do a few other things as well. It's it's just incredible where we're headed with natural language communication. I think that's the the real kicker here. You now can interact with the computer directly with voice. Not just let me have it type something or not just Alexa turn on the lights. No, no, it's you have an actual conversation. It will understand the context and what you're saying and what you're doing and what it needs to do to achieve the task that you're asking it for. And we're there. We're there already. Really impressive. And it's just going to get better. So, I just wanted to leave you with that. If you like this stream, please like and subscribe wherever you're listening to it. We are on YouTube, Spotify, LinkedIn X, Instagram, Tik Tok, all the places. But if you want to hear the show, the web talk show, you can find it on Spotify on YouTube and Apple Podcasts as **[00:30:00]** well. So I will see you in the next one. Thanks.