The FATAL FLAW in Popular AI Coding Tools Nobody Talks About

This is the future of AI coding. We want to focus on the things that won't change over the next decade. People will always want to make better products. People will always want to get more out of the LLMs and people will always want to get more competent. Those won't change. The models will always get better. Things will get cheaper. We'll figure that out. But those who have competency will get more. So, let's talk about it. I gave a talk last Friday for our community about one piece of the puzzle, specifically PRDS. As a product manager for six years and as a developer for three and as a UIUX guy for a few years, I feel like I have some competency in this space to talk about it. But there's a lot of pieces of the puzzle. If you've seen my channel before in Parker Rex, I led tech for a startup that sold for 23 million bucks. And then after that, I tried to start up in the music space. dumb really good tech though. So I've done things the right way, the wrong way, and I have been all in on a AI for the last few years since OI's SDK came out. So you have all these steps in the traditional software development life cycle and it's our job to try to get to a level of competency where we can get the best outputs from the models. So the model might get better or worse, probably just better every day, but you can also get better every single day. So, one area that I've noticed people have a lot of problems with is pitch PRD, all that stuff. Also, architecture. So, I've been hard at work trying to figure out how to stitch these together. So, I'm going to be flip-flopping between a couple screens here. Stay with me. I promise you'll learn something. So, if we go to our whiteboard, the lovely whiteboard, we can see yes, things will always come out. A cursor is going to get better. Windsur's going to get better. There's going to be this battle going on between the two of them. We've seen battles like this in the past with different industry changes. I think I could read these every video and go through like where's the new alpha, but I'm learning through my own building of software because I don't just make YouTube videos. I don't just host a community. I actually build the majority of my income comes from building software for others and now myself. But what I've learned is like you will actually want less. And you pick up some of the trends, maybe try them. If you go tool hop for a week, you're going to be depressed because you're just getting better at a meta skill. It's going to be irrelevant soon. So, use the tools to learn. But yes, here's one workflow. Another one, they're trying to find standards because it's a total mess with all these rules. I'm trying to develop my own standards. Now, with cursor, you can have different forked chats, which I think is nice, but I'm also noticing that it's crashing a lot more. as someone that's used it since September, like I forget which year, but basically like right after beta came out, been using it and for the first time ever scratching a lot, which I find weird. And so instead of being relying on people's agents, let's just find primitives that work is where my head is at because you can see people in our community and Discord are even talking about this, right? Oh man, like things just keep hitting walls. I was really hyped on Taskmaster. I think that these things make really good demos. I'm not hitting on Taskmaster at all, but they're making up for skill deficiencies in other areas. In particular, with this one, you can't get very specific on your tasks if you aren't providing enough context and you're not doing the thinking ahead of time. So, where I think this goes is just markdown. Markdown, good models, and really no tooling. Maybe a CLI that just wraps Markdown and that's what I'm working on. So in my case, I was able to put something together that combines repo mix, which just takes all of your stuff, and this is one area that I'm experimenting with because I've noticed that the models get watered down when they get put inside of WinSurf or Cursor. And I've noticed this just in my testing today and yesterday where I'll have a code base open. So this one on the left, this is that CLI that I'm working on. And then I'll pop open an instance of AI Studio. And the outputs in AI Studio are far better from the same model than if I were to put it through cursor because cursor has to then figure out how the different tool calls work and there's another layer of fat between. So you just want to cut the fat. And so when I look at this, it's like, okay, cool. I could not have made this tasks prompt if I was just using cursor or if I was just using taskmaster. And what this allows you to do if you do all the right steps in the right order is to get better outcomes. So that at the end of the day, you know what, like I said earlier, like what won't change is the models are going to get better, but won't what won't change is like your ability to give it the right inputs. So we just need to focus on like making sure we don't do any sort of garbage in garbage out stuff. So we can get to a point here. So I propose that in the future, you'll probably have a better way of chaining these together, but in my head it's okay. You have an idea prompt that takes a pitch that you've written. So you name the pitch and then you give it a responsibility. What are the expected outputs? All these things because again models will get better. The token usage it's only 1300 tokens. So it's bringing all of the different prompting that you could do in other areas and just having this mental jousting which you would have done to mirror like we're trying to mirror things that you do IRL to these each one of these prompts. So you're teaching them as like team members. And I know people and companies are trying to productize these things, but at the end of the day, it's just prompt chaining. So you'd have an idea prompt. You come in, you'd actually manually write this out. So that's the step everyone wants to skip over. Everyone wants to find like the magical prompt, but you spend some time thinking about it and really focusing on what the problem is. I highly recommend going reading high growth handbook by Elad Gil. That was part of my presentation where crisp product requirements can make a world of a difference in driving concise agreement on an execution of the product. PRD should clearly articulate primary features and product needs. Ben Horowitz says that it's literally the Bible for product. And I do want to say I think Taskmaster is fantastic. I think they're doing a great job, but it's just making up for skill deficiencies in other areas. So if you have the ability to leverage that, if you've been a PM before, if you've been engineer before, if you have these things on the left, if you gave those specifics to an LLM, it will just degrade it basically. And so you just don't want to skip over the thinking. You actually want to take ownership of that. And I'm doing that with Markdown. And also, when you get into multiple turns, like things are just going to degrade. That's just the reality of it. and it will get better, but you will gain the skills if you do it yourself. And it's not like you're raw dogging anything. You're not not using a AI at all. It's just each step in that product development life cycle, you're being challenged. And then because of that, you get system thinking, all these school skills. And on the idea of skills and like how to use them across the steps of product development, it's basically like I started mapping this out. I'm like, okay, this is what I'm doing now manually. How do I chain these together? So, I think the future is basically some sort of pattern like this where I have my actual like project development, like my task management in here. So, I'm not even leaving to go to linear or anything because it's just here. And so, I have the prompts that are chained together. So, you'd have the PRD prompt. This is getting a previous step. So, it's just passing it from one to the next. The CLI, again, it sounds like I'm showing something. it was going to be free and open source so haters going to hate. But it's basically just, okay, look at the previous step and now let's turn it into a product requirement document and then you're just jousting again. What I'd love to see is if next week or this coming week if Gemini comes out with their cloud code equivalent, I'm going to just cry. It's going to be so sick because then you don't have any of the deficiencies that are built into these idees because you just own everything. And again, like if you need the training wheels, then go get them. But this is where it's at in my opinion. And then you'd have a PRD plus. So this makes it even better. Then you have the architecture prompt that is printing out the different file trees. So this will be supported for Python and TypeScript. And then it will if you have a proposed file structure, you can give it to it, but it'll take the existing one. So if you have no project set up, then it's going to need to go bake one. And then it has a bunch of the different popular architectures in here for both JavaScript and Python. And so I think it's gonna be great. I'm testing it now. I feel good about it. Then you have the one piece and a lot of this is inspired by memory bank. But this one piece basically system patterns. So it will go and print a rule and update it so that you have a self-healing code base as you go. But this cluster of different markdown files would then just be going from if I had moved forward from step one of the pitch, then I want to go to PRD. Then I just hit next. And I'm going to tie it in as a rule, which when I get out of cursor, I'll need to figure out. It'll just probably just be CLI. But you just go through this chain and then you have PRD plus architecture system patterns tasks. This is the one where you actually have examples of it and it's contextual to your codebase. So this is 1600 16,000 tokens. You should never go beyond 40% to be safe of the context window. And then tests. So all the popular libraries in here. I'm going to be adding other things beyond this, but I just wanted to dial this in up until here because a lot of areas of this people just don't understand where it's, oh, what about like infra? How am I going to deploy this? How I'm going to think about rolling this out and marketing it? And I think that's really cool, too, because then you can bake in product loops and all the things around the psychology of the product of how to get it out. I don't know where that'd be nested because if you do it postfacto then you're not really thinking about it at the start. But if you guys have any comments on this, let me know. it would hold all the changes in memory which I think is cool and then yeah figuring out repo mix is gonna be another challenge but it should help with this whole like stepby-step thing because when after I gave this presentation I was like man like I was trying to wrap it up for our community and give it to them. I was like it's still confusing but I think it's going to be better now and this is what it was looking like before. I think having self- enforcing rules makes a lot of sense, but we're again just trying to map it to that. And yeah, so that's where my head is at with all of this. And I think there's maybe a couple more things on this Excali draw that I could go through that would be helpful. Yeah, basically less is more. And I'm hoping that Gemini launches a cloud code thing because then my dream is basically just to have a instance of a terminal that's open like this. And then I'd have my codebase, I'd have my Gemini, and then I can just have a task. not even task master because it's really just like your coding assistant that jumps you through it. So I I think that's where it's going to go. I'm excited to see how it pan out. Even people at OpenAI are collaborating to try to figure out like an actual system for this so it's not so messy. If you found this video helpful, make sure you subscribe to the channel. I'm getting momentum back. I had uh I was away for a week, so I'm working on getting momentum back to the daily uploads. I appreciate everyone's patience on this. If you liked the video, awesome. If you didn't like the video, you should like it. even if you hate me cuz you know it's good good for your spirit. And if you want to check out the community to see all of us yapping, it's 50 bucks a month, but we've gotten some great testimonials lately. Like my goal is basically to just get out of my own tech desert and have a lot of people in here that are smarter than me. Got people from Microsoft in here. We've got some Googlers joining. Those are some of the brand names. And I will see you in the next one.