I Wasted 100+ Hours in "AI Hell" So You Don't Have To - AVOID THESE MISTAKES!

Whoa. Let's talk about it. Yes, you should have a healthy amount of skepticism for everything that a model says. Specific topics, it will just tell you something that is completely wrong and send you down a rabbit hole when you could have just checked source material. Let's talk about it. They're just not always right. And I've had this example come up very recently. Recency bias drives me towards this. Also just had a good friend Dan who reached out with another similar case where you run into a wall. So this is a daily upload. Parker Rex is my name and doing AI stuff is my game. Oh god. Anyways, yeah. So first of all, we usually do QA news and strategy. We got a lot of stuff to cover today. So let's jump into it. But on the point of the video, the title, the reason you clicked here is there's a couple phrases that will throw you into AI hell, aka running in loops and being very sad. You don't want to be sad, right? So, industry standard, this one big jargon buzzword. If you see this industry standard, it's like there's a hole and you're walking and industry standard is the palm frron that's over the top of it. And then you go down into AI dungeon, meaning it's going to say, "Oh, industry standard for Python is to use Flask for this." And you're like, "Is it though?" because you didn't read the spec and flask misses out on a lot of the stuff and you told me industry standard would be adding all these other things around it to make up for the fact that it doesn't have what fast API has. Or another example literally just got a DM about this from Dan who I feel like Dan's gonna be like the star of the show today. I don't know why but but he sent me this video and he's like, "Yo, can I get your take on it?" And I'm like, "It's really quiet." And it's funny because most of my videos were really quiet for a long time. So you do it wrong a lot of times and then you finally do it right. But he gets told that this is how his limiter should work. But then you can't hear his video very well. So it confidently tells him that. You're like, "Wow, I picked a really smart model, so it's totally going to be right." No. Boom. Back in the pothole. So just always verify. And you can quickly do this by just, I don't know, using your brain. Find other people that are doing the thing that you're trying to do. verify against it. You could probably AI this if you wanted to, but it'd be overkill where it's like, oh, I'm gonna use Exa or I'm gonna use like Plexity and chain it together. No, just do a little bit of homework. So, jumping on to that, since we're on the DAN topic, this is a big theme that I'm working on right now, which is the AI developer software development life cycle. And so, I gave a talk on PRDS and I gave a talk on the AI life cycle inside of our community. It's called Vibe with AI. It is primarily a Discord, but I do a weekly maybe other every other weekly. Depends. I'm not sure if I should do workshops every two weeks or every one week because I don't want them to just be diluted. I don't want a watery glass of lemonade. I don't want to have a watered down version of workshops. But in terms of AI SDLC or software development life cycle, we're trying to map this something like this. There are more steps to this. You could go as far as what's the release plan, how we're going to do that, how we're going to do communications, what's a PR FAQ look like. But in my talk, I basically said, look, there's all these tools. Most of the people that find out about these tools are either on one side of the skill spectrum, which is like newbie, and they go and they use Bolt and they're like, "Wow, I'm an engineer." Or they're like a pro and they work at Microsoft. We have Microsoft people in our community that like found me and they're like, "Wow, I didn't know about these tools." So, there's this like weird gap. A lot of seniors get annoyed at the juniors because they juniors like the people that are vibe coding think that they're like gangsters and they're not. But point is we're trying to figure out a system that works for all of these and so you don't have this moment. And I really do strongly believe that this is it where you want to be here. The only thing that I'd add to this after I made this talk was markdown. It' be a little bit more specific, but markdown lives in there. And then maybe even just like zed because the more I'm using cursor the more I'm realizing it runs into performance issues and that's going to be something that seems to be a trend because it's a VS code fork whereas zed is finally adding a bunch of the stuff but point is we're trying to map the AIS DLC to tools or patterns really and Dan's talking about it I'm talking about it people are talking about it at OpenAI trying to figure it out. So, I had started with this where it's like, "Hey, cool. If I use Taskmaster, because I make a video about that and everyone's like, oh my gosh, like it's going to be the new Jira." And you're like, "No, it's good for smaller things." So, I think it's good for people to use if they don't know how to make the tasks themselves, but it misses out on a lot of the context. This is something that I like the first version I was working on looked something like this. That's a lot of files, but you need the tree in order to get more accurate, in order to get something like this where on the lefth hand side we have the task list. That's actually all the things that we need to do. Inclusive of library, inclusive of snippets. All the blue stuff is a sauce on the right hand side. Taskmaster, which is just super vague. Another bit on the problem space is we're talking about it in here. Like this guy, ML engineer, talks all about this amazing set of prompts that he's going to use, but then you end up with this and you're like, cool, I'm going to give that to an LLM and it's going to literally run in circles for 4 days and it's going to cost me 50 bucks. Like sick. So that's not it either. And so what Dan's proposing is a pattern that basically just uses readmes in a fractal way, meaning it's specific to the different subdirectories in the project. And I think it's interesting. I think it needs work. You can check out his channel here. Just make sure your volume's up. And that's not a hate. I just clear markdown. It's perfect if your volume's all the way up, but I was trying to maybe some. But he made a GitHub. You can check it out in here. And it's similar to what I'm doing where everyone's trying to figure out what to do. But I had tried this in the past, but it wasn't as detailed as his. I find it pretty interesting where I was like, "Oo, I'm going to have readmes everywhere." And I guess I could try to find a project where I showed that. Let's see. We can cd into code and then I think it's map front end. And then let's do v dot. Let's open that up. So this is like an example in my apps folder. I believe I had followed something where that's not it. But I had basically made like readmes in different areas of the project. Maybe I removed them. Yeah, I already removed them because I just found it annoying having all the stuff everywhere. Like you can see already like you're just jumping around too much. So, I'm interested to see where it goes. I could be totally missing the boat on this. I watched part of the video, skipped skipped to the parts that I thought were the most interesting, but there is sauce in here. So, shout out Dan. And then also Dan was asking if you are interested in joining the community, you can get your business to if you work for a company, you can get them to expense it as an educational basically like continuing education thing for you to get them to cover it. A lot of people need a website. We don't have a website. So I'm going to build one of those. But that's just another thing. Optional XML tags if you want. Wow. Now it easier. Where is it? But it's things like there he is. Okay. Okay. And now let's go on to Santiago. If you don't follow this guy, go follow him. But he's talking a lot about knowledge graphs. So how you would ask people MCP, which I'm allergic to MCPS because I think that they're really janky, but I also understand that I could be totally wrong. So the way that I'm thinking about doing this and adding it to my AISDLC is to have the MCP that basically is always running. And the gist of it is he has something set up where it's using graffiti and then in the background it is constantly updating that as a node DB and the gist of it or like the short of it is this would be an example right I would tie this into the workflow meaning that each time that we go from one step to the next then it would call that MCP and I should show you what the AIS LC looks like and I should show you my API key jk you can't see the rest of it that's a thumbnail demo that I'm working on for Echo. But in this case, I would think about infusing some of the elements of what Dan's talking about with readmemes and then some of the elements of that MCP. I'm going to do a separate feature branch for the MCP because I don't want to block it. I want to get this out. And what this does is essentially you have your task management in here. So you have doing and then doing is the start of a new work stream. A workstream can be for debugging, it can be for new feature development, or it can be for ramping up on the codebase. So in each workstream, you have a set of prompts that follow the software development life cycle. And I'm focusing mostly on new feature dev. That's going to be the first total workstream that I'm completing where you have an idea prompt. And the idea prompt starts with a very basic pitch. So you'd run this CLI. So you'd say AISDLC init and then it'll create this file structure and it's basically scripting around just moving you through a bunch of prompts. It's prompt chain with a wrapper, right? Everything's a wrapper. TSMC is just a rapper for rock. Apple's just a a rapper for TSMC. But you'd have this idea prompt and I'd go and I'd fill out this pitch and then it would joust with me to try to make it better. So, we'd go back and forth and then once I felt good about that, I'd go to next and next would then create it from if I had filled this out to the PRD prompt where it's going to inject that pitch into the PRD prompt and then run the actual call so that it generates the prompt. So then it' be 02 I prdello world. So it's dynamic to the name of the feature and then you just go through these. So then PD plus prompt is an additional layer. And I'm trying to figure out how to make it so that people can't just like cruise through because the value of your outputs are directly tied to the value of the inputs. Garbage in, garbage out. And so I want to figure out a way where it's, hey, you should probably do just the PRD prompt like three or four times before you move forward. And you can bake that into the CLI. And then you get into architecture. This is supporting both Typescript, obviously JavaScript because it's just a subset and Python. And the way that it's supported by it is because I put in a bunch of the patterns. So these would work for other languages, right? Like patterns go elsewhere. But in our case, I have both the architecture system patterns and the tests and the test libraries tied to the like language specific. It also includes a bunch of the out of the box tools that you should have on your Mac. Shout out Mac. And that includes like tree. So file tree. And if you had no tree to begin with, net new project, then it's going to need to know that you can come up with a plan. And then a proposed file structure. I should add lines of code in here because that's nice to have too. You can see if things are dry or not just off the rip. And then you have system patterns. This is the first living and self-healing document. This is from Nick at Klein and he's a G. But taking that kind of pattern and then putting it into rules. So maybe not rules is the best name, but it's like living docs or something. Whatever. If you have feedback on a better name, then good for you. Just drop a comment. And yeah, this would basically be the one that it would go and check and see if there are rules to this to follow. If there are not rules, then it would put them in here for the actual like system. So right now it's empty. And then once you get past that, then you're actually doing the task specific to that. So it pass in the pitch, the PRD, the architecture, the system patterns, and has a big example of what it looks like. So this is actually 16,000 tokens, and we need to focus on what won't change. So what won't change is you're going to have to use your brain. You're going to have to have competency in order to get the edge, in order to be ahead. So you can go and stop watching this video and ram your brain full of mush, but just don't do that, please. We don't want to ram our brain full of mush, but you'd end up with a great task list. And then finally, you have a prompt, which I should add that. I'm going to add prompt for task prior, not prioritization because it's doing that already. But it's basically once you're actually going on the list, there's going to be like 50 of them. And so it's for task for actually doing the tasks. Cool. So we need that. And then this should check against the token limit somehow. That would be really cool. So if I end up making the CLI just be like raw calls instead of the cursor agent. So that way it's like tool agnostic, then I can figure out how long the context window is and then like meter against that. Similar if I go into here and I have this up here, like that's the sauce right there. So yeah, but that pretty much summarizes this. Then you have the test engineer. I want to add as you see step for actual like deployment slashinfra. And it could be argued that this should be earlier. So we'll see. But that's another black hole for a lot of people. That was literally like the biggest problem that I had in multiple projects was just like figuring that out. Also think about repo mix as part of the flow. So repo mix if you're not aware really good for just right clicking stuff getting the whole content of the selected files. But you could just not even do that. you could just run instead of doing a file tree, you could do their thing which is it takes the whole codebase and then you can define which areas of the codebase that you want to have optimized for LLM readability and then you can select if you want XML or markdown and so that's what I think it's in here I deleted already but yeah so I'm exploring this and this video is pretty long already 15 minutes wow a lot of stuff mermaid diagrams thinking about putting those in. So, let me put that in the read me. Mermaid diagrams question mark. Okay. And I don't know, I wanted to check out this zero thing. This guy's built like a really good task manager, but honestly, not worth the time because I just want to build it into CLI. So, other things that I'm doing now in the AI strategy is what I learned this week. I just want to start covering that because I'm going to add it in as accountability inside of our Discord channel because a lot of this when you're learning new stuff is not so much the content like I could make a bunch of content but it doesn't mean you're actually do it. The it's similar like if you had like access to every workout plan on the planet you can do like you can just go rip anything you could ask chat GPT but the way that you stay accountable and actually do the work is by having a group around you or a trainer or something. So, I want to emulate that in here with both what I learned this week because a grouped knowledge pool is far better than you trying to figure out yourself. I always made this reference to Mr. Beast when he was getting started on YouTube. He had five people that were super obsessed. They'd talk all day about what they're learning. And that shared knowledge pool moved everybody forward faster. So, I think it's going to be more in this direction for the community, which is like shared knowledge pool plus the guy who's shephering everyone, aka me. So I do have monetary gains from that because people recognize, wow, like I'm doing work. I'm driving traffic and like organizing it. So it'll be a combination of like workshops and then the community actually like being part of your accountability group. So I think by just having something simple like what I learned because I also see this in different channels like we have an iOS channel and Nick is learning how to do iOS dev. He went from like zero to hero. He's already built stuff which is dope since he joined. One of the first members shout out to him. I want to have a dedicated spot and then I'm not going to paying people to do it but I'm just going to start doing it. That's what I've noticed is if you want to get people to start doing a behavior that you want a community then you have to do it yourself a lot and then people see that and if they gain value then they start doing it and then I'll do a what I did like post week maybe Sunday or Friday for the normal people and then it would look something like this but yeah like whip.co co is like a work in progress thing. Keeps people moving and grooving. And on the case of what did I actually learn? I learned about the repository pattern. I hadn't done that before. And I think it's something that's like pretty niche because I have a bunch of different storage places. There's database for Superbase. There's a storage bucket for Google Cloud for videos. There's signed URLs. There's a bunch of stuff. So in my re research, I found that this was the best one to do. I wish I had that AISDLC before I did this because it would have saved a lot of time, but I was kind of like co-developing them. And then I learned more about fast API. I spent a ton of time on UV because I had a very headline understanding of package management and Python in general and understanding environments, understanding, oh, should I use a dev container that just does all of this and run it on its own little island where it has its own extensions. So I don't have all these extensions. Like if I'm working in JavaScript or TypeScript, then I don't want to have all of the like rough linting and stuff like that. Or vice versa, if I'm in Python, I don't want to have all these other extensions that are like in the way. And then I learned a bunch more about Tanstack router doing demos with it because I made the first version of the app and then I kept hitting the same bugs and I realized, wow, I'm using Tanstack start, but Tanstack starts actually just like 95% Tanstack router with an escape hatch for SSR. I was like, I need to learn more about this. So I watched a bunch of talks from Tanner and then on the what I did I completed the V0.1 of Echo that is the automation content creation software and so it does a bunch of let's call it like 15 different pieces of metadata around a YouTube video and then obviously functionality around uploading everything except thumbnails is working. So I'm going to work on that this week and then I completed V 0.1 of AIS SDLC. So I'm using that now. I gave a talk on PRDs inside of the community. I tested zed. I'm really happy with it. It was the first time I've used a new IDE or I've jumped into a tool where I just had no complaints. Like it's just good now. And it took a long time, but I know they rewrote the whole thing in Rust. So, it's fast and has native notifications. So, if you leave and you come back, you can follow the agent. Very cool. Stuff that you can't do in a fork of VS Code. Tested codecs. Mostly good. Mostly good. And then what I'm doing this week, I'm getting back in the gym, fellas. and maybe the one girl that's walking this watching this because this last two months I've been getting a lot of stuff on the ground or off the ground between client work, getting the community going, getting my own sasses running, but you get like achy from sitting all day. And I need to get back to 315 on deadlift, which isn't a lot of weight, but I need to just get in there. And then I'm going to be building the multi-stage thumbnail generator using pill, which is the Python imaging library. There's a fork of it called pillow. I'm going to be building the membership site with s tanstack start. This should be really quick if AISDLC does what I'm hoping it can do. I'm adding discord accountability channel. So what I learned this week will wilt adding that. I'm adding discord validation because we have people that churn who mostly are people that just don't understand that you actually have to use your brain and work in order to learn this stuff. They jumped into bolt. They made a marketing site. They made tic-tac-toe. they think they're an engineer and then they realize they're not. And then I'm going to be releasing AISDLC to the community so we can get feedback and then eventually release it outwards. So that's all for this video. If you learned anything, I hope you did. Then make sure you subscribe to the channel. Also like the video so you see more of these models aren't always right. You don't want to end up in here. Subscribe, like. I already said that. Check out the community if you want to. Drop a comment below on any sort of feedback. And I'll see you in the next one. Peace.