LLM Experience is the New Developer Experience (or at least part of the AI pic ya you understand)

Continue working on the express to fast API migration. Okay, cool. I can only do this because LLMs don't care about your feelings. I learned this the hard way of choosing some languages, some frameworks that I thought would be fun to learn just to get to the fact that I want background agents running a lot of my code. And it doesn't make sense if you're not going to pick something that they know. So, let's talk about it. This is a daily upload channel where I talk about a couple things. I talk about comments that come in I cover some news that's relevant and some strategy. So this will be the main piece that I get into in this video. So, first of all, on the new side, if you're interested in learning more about why I'm so bullish on GCP and the actual implementation details beyond the business details, it's obvious they have a money printer. They're not beholden to anybody. They have the science, they have the TPUs, and they have the data. Yeah. And their platform is just built to deploy. Now, there are certain things that you should use it for and should not use it for because it's an enterprise product. you would not want to use it for, let's say, a database if you're not at scale because you could pick up something like Superbase and it costs like $20 whereas the entry monthly for Cloud SQL or Alloy or those is a few hundred. That's just to have it stood up. But if you want to learn more, you can check out this playlist that I made. You just go to my YouTube and then go to playlists. It's called GCP. It's got 10 videos in it. Some of them are a little bit more boring than others, but they're really educational. The second thing I was starting to watch today, and this is a funny thing cuz I was really trying to get rid of Next. I was really trying to go and go all in on Tanstack. I watched every Tanner Lindsley video about Tanstack start. I was sick of depending on next and some of the foot guns that I've found by using it, but the LLMs are so good at it. And they have the capital and they have a model now that's really great at design. So I think for not I wouldn't have this for an API. I'm going to have a video of my full dev setup on my main channel that goes out. Today's Wednesday. I'm going to film that after this. But it's going to be part of the stack. It's going to be a thin back end and just front end stuff. But in this talk, he explains how whenever you're learning something and you want to get people that are like you that'll want to either be friends or build community or market, you should start teaching it. And so that's how he got into open source was just teaching basic stuff online. And it just opened a bunch of doors. And I'm finding that myself with this YouTube channel. It's this marketing flywheel. Yeah, I make a couple bucks off the community, but that's not the point. The point is long-term that it opens doors to relationships because I live in a tech desert and there's a lot more effective ways to make money than doing community stuff, but it's not the point. The point yourself with people that are interested in the same stuff is just awesome. So, let's get into some questions and answers. We see Danie Danieli Maro. He says, "Thanks for another insightful video. Could you make one explaining how you would add a complex new feature to one of the existing projects? Sometimes we can simply start over and have to deal with the lack of rigorous planning at the beginning. Or should we? What to do in these circumstances? Good question. So for this video that he's referencing is a road map that I was explaining and I'm building a lot of tools and that's actually what is part of the community is we're we have a repository full of different things that I'm prepping to release later this month. One of them is actually just the YouTube pipeline, but that's not what we want to talk about. One of them is this AI SDLC. SDLC is software development life cycle. And this included just a couple of the steps in the software development life cycle. So it had an idea prompt. It had a PRD prompt, a PRD plus prompt which pokes holes in it, an architecture prompt which is full of examples both for Python and TypeScript. So that's the two languages that this will be adhering to. But it has a bunch of examples of different architectures that you'd use based on what you're building. Then it has system patterns based on the architecture based on what you're building and then tasks prompt which is really important because then it gives you actually really high quality tasks that you know look something like this and then it can go and actually do the work on the right hand side. So when he asked the question, he's like, "Hey, how would you do this for a new feature to an existing project?" Yeah, it's the same. It's just more context. And that's one of the biggest bottlenecks that I found with tooling so far. I'll go into this more in depth on my video on my main channel, but I started using Augment uh maybe two months ago off of a recommendation from a great member of our community named Brian and he was like, "Yo," and he had this crazy workflow and I don't even know where. It's somewhere in here, which is why we're building this product so that we can find the best stuff and then have it not be in the thread. We also have a co-pilot that'll help with that. But he was like, "Yo, you got to use Hogman." and I went and I used it and I was like, "Well, this is cool." And then they reached out and they're like, "Hey, will you make videos about augment?" And I was like, "Yeah, as long as I can not have to read something and I can give like raw feedback." So, this isn't one of those videos, but I'll be doing it in depth on them. Augment solves the problem for context. Why does context matter when you're building something? If you're in the middle of a project and you have something that looks like this where in my case this is a monor repo and it uses turbo and this is augment on the right but it has packages right and it has a discord shared package and then it has apps so you have a fast API you have a express API that I'm moving away from you have discord bots which I'm going to be scaffolding that he's has each co-pilot within there and then you have web app which is next. And when I want to add something, it's really important that I find all the relevant pieces. So, I would follow the same exact steps as the AISDLC where I'm like, what's the idea? What's the problem I'm solving? And then actually write out three solutions for those. That's something I'm adding now off of the way that Paul Coppelstone, the founder and CEO, co-founder and CEO of Superbase, because they're super open source. If you just write one solution, you get really tied to it. So they have a request for comment when they put something out that's new. So when I want to write the first step in a new or an existing codebase, I want to come up with three solutions and then I just run through those steps. Now there will be an improvement to this and that I made this in v 0ero yesterday. This started as a dock that looked like this and it's all the steps that I think are necessary when you want to go from zero to production. So all the way through performance optimization and knowledge management, user feedback, all these things. It's 15 steps. And so I just made this little interactive thing. You can check it out at this link. But this will be how I'm scaffolding what a an agentic workflow would look like for background agents. Yeah, this is where I think it's going. And that plays into why you need to use LLMs that are good at the different jobs. So yeah, I hope that answers your question. It is basically the same but with more context. So check out the video going out in an hour. I cover this. Thanks for the comment. AISDLC. I'll just reply after this video. Hi Parker. I'm curious about the security of my source code. Does cursor potentially have access to my source code while running these background tests? Yeah, I would go and read their privacy policy, but I'm assuming that they do. So, just be aware of that. And then let's do one last question. Why does Val look suspiciously like an NFT collectible, though? Maybe because I'm an NFT, bro. No, it's because I took a GitHub co-pilot announcement article that I liked and I was like, "Wow, that looks cool." And I screenshotted it and then I brought it in and I made a mascot out of it. So, that's why. And you make this little brand.json. You can check out a video that I do on that particular task on this channel. So, let's get into strategy. Oh man, I forgot. I got Gemini diffusion. I don't even know what I'm going to do with it. Maybe I'll just grab some code. Let's say turn this code. Yeah, let's say translate this into Rust. I just I have no idea how it's going to work. Let's see how fast it is. That's a bunch of Okay. It would be a complete rewrite. All right, never mind. I figured that'd be a dumb way of doing it, but that was pretty fast, man. Yeah. So, back to the LLMs and how they don't care about you. This is me at 3:00 in the morning last night watching background agents in both augment and cursor. Now, cursor, their background agent thing is so buggy. I know it's in beta, but it just crashes. I can't even touch cursor anymore just because it's so buggy. But this is me last night. I spent an extended spike. That's what we're calling it. Where I went and learned Tanstack start over the last couple of weeks for a product that I'm building called Echo and I was also learning fast API at the same time. So yeah, in total it's been like six weeks of build since I built the MVP. But I get to the finish line, I'm like, "Hey, it works." And I'm like, "Okay, agents go." And they just run in loops and it costs money. and that's stupid and it's going to be forever until they're smart. So, just use something that's usable and that they know. Then I found this channel where this guy's actually here's how you would do it if you wanted to use Django and Celery with different web clients. And I watched this guy's video and I was like that's perfect. Like I'm going to set something up like this. So I made a template that or I made a project that basically allows you to have fast API and next and it's really simple. Nothing crazy. And so that's what the pattern is that I'm going to be using for moving forward. And you just have to think of these things will drive these LLMs will drive a lot of the choices that you make in technology. Now I know that OpenAI moved to Remix, but I also have read that they're not happy with that and they're not the fastest tool. T3 chat is basically a bajillion times faster and it uses convex and it uses next but they like wrapped it at the root so that it would be client side but it's very quick and using react and next.js gro react with Nex.js and shad CN literally is in the system prompt for every open AI system prompt. So that stack is going to be the one that I'm going with for front end. And then what else? Yeah, Google, you know, they made Angular, so they're pretty dieh hard on it. I think that it does influence the way that we build things. I'm even in my architecture with VI, which if you didn't know, it's the new product we're launching in June that looks something like this. This is one hour in V 0. The new V0 is actually really good. And I hate to say it now, ignore the UI of it and just focus on UX. So I was like, hey, what would it look like if we just took like the Versel thing and cloned it for VAI? And it would be something like this where you have the overview and you have the things that you need to do. You can see the new members that are coming in and click into them and see if you want to get to talk to them based on their skill set. And ultimately like I want to just have people talking and learning from one another. upcoming events, repositories that you have access to, announcements. Again, don't focus on how ugly these outlines are. Focus on just the features. So, the UX, this would be AI integrations. So, we're building out a playground where you bring your own key and then the playground allows you to use the different prompts in the prompt library. Ignore all these cards at the top. V 0 does like to throw cards everywhere, but this table will be a playground where you can see what the summary of the prompt is, how popular it is, the inputs, so what are the things that you're going to be putting in, and what are the things that are going to be coming out of it. Most of these will be written in XML. What model we find to be the most successful based on people using it, the intent of it, so what is it that you're doing? And then you can filter them. and then who is it submitted by and then you can copy it and then you'd have the playground where you can go and test all these things a members directory. Again, this is one hour of putting all this together and now it's actually in a production codebase and getting tied up to all the different external APIs. So then I'm going to go and actually design it and make it look nice. You'd have the integration so connecting your GitHub, your Discord, yada, and then settings. So, general billing, invoices, and notifications. And then news feeds. This looks horrendous, but this will be baseline news feeds based on stuff that I check where I used to be like a news parent on this channel. And we'll have a data pipeline. That's what I like doing. I don't like doing frontend stuff as much because I've just done it for so long. You can't tell by this because this like trash, but I was a designer for a very long time, but I really like doing the orchestration behind the scenes. And then we'll have the learning tracks. So, it's similar to what this looks like or it's some sort of video game like feeling. And then I have playlists on YouTube of sources that I found to be really helpful. So, we can put that in there and then we can have usergenerated learning paths as well because I just find that it's like when you talk to people that are competent, when you collect information from Reddit or YouTube or the docs, most importantly docs, that's kind of the best way to do it. And so I want to collect all that in one spot and then we'll have different co-pilots that are trained on the different frameworks or languages that you're trying to learn. And then that's powered by this scraper called context brains. And this is off builder IO. So what you do is you basically I actually built GPTs for devs off this which is like a popular iOS project. But you put in what these are. It scrapes them into a JSON and then you put that into the context of either a rag engine or just literally a GPT. But when I'm building this and how this is shaped by the LLMs is when you look at something like, oh, we want to have the ability to have a mobile app later. iOS is still really shaky on stuff and if I didn't build this the right way, then I couldn't open the door for Expo. And I know that it's really good at expo. So I just think that we're underestimating just how much our life cycles, our development life cycles are going to be shaped by the LLMs because they're trained on so much of it. And if you're a large scale company, then for sure you can go and you can actually train your LLMs on your codebase and that's fine. I know Meta is doing that and even a company like Augment 90% of their agent code is written by the agent but that's because they have a team of AI researchers they have the resources they have the capital they have the time to go and tune it to their model and tune the tools to the things they have around us but for smaller teams I don't think it makes sense I think you actually need to pick things up that the agents can do so you can get that leverage so if you learned just one thing in this video make sure you like the video make sure you subscribe to the channel if you want to be a part of EI lock in the half off, then go get it now. Hope that we'll have this out in beta in a week. I'm hoping that I'm using it by the end of the day today. And you can check that out. Make sure you subscribe to the channel and I'll see you in the next one. I got to get thumbnails. So, let me That's how you do thumbnails. All right. It's