How Cursor’s Developers Actually Use Cursor (And You Should Too)

See that? That is a sign of some frustration trying to work with LLMs that don't know how to use a new framework. That's we're starting with that hot tip. Do not use LLMs with brand new stuff because they're trained on stuff from six months ago and they'll bias against it even if you feed them the right info. Anyways, I'm Parker Rex. Welcome to the channel and today we're talking about some AI news things that have come out. It is Friday, May 2nd and let's just jump into it. We cover Q&A but I answered most of the questions yesterday. So, a couple short and snappy AI news stories and things going on. So, first of all, yeah, if you want to use Gemini, you can paste in the repo. That's pretty cool for the folks that want to stay away from IDE. It's not me, but it's there. We're going to talk about tips from Super Maven and one of the guys that was on the founding team there. What is Super Maven? Well, they had a big part in the way that code bases are indexed in cursor because cursor bought Super Maven. And the guy breaks it down. It gives us the sauce. We're going to talk about YouTube automation stuff, which maybe I just think is like a wild trend. And then let's just cover this. Who do you guys think will win this bet? Who's going to have the best model by the end of May? There's actually a wild amount of bets that are going on in the poly market. I think it's Google, but I don't think that's like a controversial thing. That's like saying, "Hey, is the best person in the world at golf going to win the next golf thing?" Yeah, probably if they're the best in the world and they've showed that they're awesome. Yeah, it's kind of crazy. So, let's jump into the Super Maven stuff. So, I want to just say that there is no one-sizefits-all. A lot of the times when you're using cursor, when you're using whatever tool, it's dependent on where you are at in your journey. If you're a zero out of 10, you've never coded before and you get a PRD and you're like, "Oh my gosh, like why isn't it working?" Well, it's because you didn't know that there's like 10 other things that you have to do. So, that said, this is Adam. He knows what he's doing. So, I've been working with AI assisted coding tools for a long time. Super Maven. One of the biggest things uh one of the area I've put a lot of thought into is how to use these tools most effectively. Here's what I have found makes the biggest impact. Make sure to understand the problem you're trying to solve before making changes. Tada. Product management 101. Start in the problem space. Narrow it down. Figure out what it is that you're doing. Why are you doing it? It's the why that matters. Otherwise, you're just wasting time. You need to be solving a problem. Plan with ask mode until you're satisfied and then change to agent to do the implementation. Okay, agree. Right now I'm relying on 03 for ask a Gemini 25 pro for agent. Nice. So yeah, plan ask plan ask plan ask. That's why you don't need a lot of tools. I actually made a meme of this for today's workshop in I vibe with AI where you just don't really want to be this guy. I for the workshop went and tried every single tool and that's why my eyeball exploded was because it's just it's such a waste of time because you'll go and you jump and you jump and you jump and then you realize, wait a second, I just need this. I should have just stuck with what I knew. Second biggest problem before we jump back is people don't realize that there's a lot more to it than the PRD. Everyone's searching for this prompt. Oh my gosh, give me the prompt. Yeah, it's helpful, but like did you actually think through the idea? Did you narrow down the problem? Probably not. And so only then can you go and expect that a PRD would be good once you've validated that this is a problem. This should exist, whether that's with a customer or yourself if you are the first customer. Always recommend doing that if it's possible, but if you're working at a company, validate the problem that the other person's having. Then you can go into architecture. That's when you'd get into stuff like diagramming. Let me pull up like mermaid for instance. Like this is what I'm building right now and it's for automating this channel, right? So this would be the architecture part, right? I wouldn't start with this because why would I architect before I narrow down the problem before I come up with the PRD? Then I go and architect stuff and it's like, okay, cool. You're using GCP, you're using Tanstack start, you're using Superbase, you're using Flask. Great. How do all these things work together? So yeah, um let's get back to what he's talking about. Oops. So he just hops back and forth. That would be that second step, right? Like you don't even don't go there if you haven't done this the ideation. So to get most out of the agent understand what the limits are. If it's consistently implementing the changes in one shot, try increasing the complexity of the changes to get more done in a single request. Instead of a single task, ask the agent to execute a series of changes. I think this come back to context windows. So, I talked about this today in the workshop, but you have a context window with a maximum input token and output token. And um you basically your quality of work that's going to come back goes down 60% once it's at 50% of its context window. A lot of percentages. Let's use easy numbers. You're using a good one like Gemini, say a million. A million's the the size of the context window. when you get to 500,000 expect it to be worse. So you want to continuously refresh those windows and it's easy and then you hold context of what you're doing in some documentation. You can use the memory bank feature with client, but I just recommend frankly just have docs like whether it's in subdirectories for getting context quick within a thing that you're doing or if it's having uh progress with like all the atomic steps for the purity that's associated like you need to break the work apart. You can't expect to get 50 screens done in a one shot. No, sorry. It's not going to be good. And depending on your skill level, if you're good, then you don't even need Taskmaster. Taskmaster basically exists to take a PRD that's good and then come up with a tasks. But if you're very specific on the tasks, then you're going to end up going into the task.json that it makes and then tweaking it because it won't get it right. So you could just raw dog it and literally have the best PRD that you ever written assisted by AI and then an atomic checklist associated with that PRD. Rinse and repeat to get most out of the agent. Yep. We just read that here's an example of a task where an agent would could do the work in a single prompt instead of three different ones. Models are changing so consistent experimentation is key. Agree? But it still comes back to the fact that you need to be good at the thing to be able to know what to do. So skill up and don't be afraid. Easiest time to learn ever. So on the left he has three that are separate. On the right he did it in one and he said the one on the right's better. But then sometimes that doesn't work. So then you break them apart. This just goes to like experimentation. And I think that again like the more I work with these after spending the last week popping every vessel in my eyeball fighting with these tools that just replace small subtasks of what someone already did five years ago. I realize like just agent in the right side of every IDE plus your brain equals a win, right? Your brain needs to be on fire. It needs to be knowing how to write PRDs. It needs to know how to break the context windows out and then have the tasks that are atomic. So if you're bad at parsing tasks, then go ahead and use Taskmaster for sure. But if you already have the tasks and they're specific, then you give them taskmaster, it's just going to throw them in the washing machine and it's going to come out different. Your codebase probably has a lot of similar code to what you want to produce. Yes. So this comes back to rules. You should have rules for how you do data fetching. You should have rules for all the different patterns. So if I go into our private Discord and I go to webdev, I dropped one of these in here yesterday. Yeah, here. Documentation refactoring. Cool. Let's open this up. This would be um I need that Discord thing because I'm building stuff. But this basically helps you document. This is how we do server client architecture. This is how we do state management. This is how we do data fetching. This is if you know just another thing that people already did. That's what people need to really hammer home is these agents, these little Cody tools. They're just doing the thing that you did before and they're very very good at it if they're guided by someone that knew how to do it before. routing, navigation, error handling, logging, tracing, features, so you can break out the features, the components. Having this in a nice readme, really good. Having your types, if you have database generated types in a place, really good. Um, having one master seed.SQL, which like explains like this is how our database is set up. Really helpful, right? And then when it comes to testing, same thing. like you pick if you want to test based on the size of the project. So in our call today, we have people that work at like big fancy companies, right? And their stakes are much higher where they're not just building an MVP for themselves. They're building something for a service that's already existing. There's millions of dollars flowing through it or billions of dollars. So like doing tests obviously matters a lot and having those tests be across the board because when you deploy to AWS or GCP or Azure there's going to be differences and so you can actually put that on them um to do because you don't want to sit there and then you can go do something else. That's leverage. So yep you include all those things. That's what I mentioned with the rules and the documentation. Some patterns I commonly reference. Cool. I think this is okay. I think it's just better if you'd have a rule where it's like each time that you do this, you must add, you know, this utility of the logger or whatever. So, once you finish working on your change, ask for a summary, get the PR. Yeah, makes sense. Yeah. So, I think this is a good little thread. I'm sure he does a lot more than this because it just doesn't it it doesn't get into the depth of like all the different steps, right? There's a bunch of them. And so I just kind of map agents and prompts. When I say agent, I just mean the thing on the right side of cursor or wind surf. I map that to prompts that I'm using for each step. And so design UIUX, you can have a brand guideline. You could also, if you wanted to use a library and pick it up so you can do previews and all that. And then when it comes to, oh, I need like a browser MCP or I need a puppeteer MCP or I need this MCP, it's like, wait a second, use Puppeteer. And that's existed for a long time. So yeah. What else was I going to talk about? Uh, I just think it's bananas that there's people that are making like a boatload of money by stacking together 11 Labs and Image Gen. And I made a list of these cuz I'm like this would be fun. Like is this a weekend project? I don't know. It's not going to take precedence over stuff I'm doing right now. But I am curious how it works. Let me find my AI YouTube automation list. And I just want to take a look at it so people are aware. So this guy, yeah, he's pulling 200,000 views in 48 hours. Kind of crazy. Um let's see. There's just different niches where it's like here's a guy that made 1,500. That's nothing actually. That's no cash. I don't know. I came across this like random yesterday and I just found it interesting how it's like 412K on the month for automating faceless YouTube channels, but you have to still have taste, right? So there's just really interesting niches where it'll be like, "Oh, African history or World War II stories or nightmare stories." I don't know why you'd want to listen to that, but your cogs are basically 11 Labs and then videogen. That's it for today. If you guys found this helpful, make sure you like the video. That's how uh I don't know, you pay it forward, as they say. And then if you want to get a discount half off for our community, this is a long-standing thing. Like this is not actually I'll just show you. This is what I talked about today. It's like it depends who you learn from. So there's a lot of people that are good marketers and they actually haven't done the thing before. I've done a lot of stuff and I'm doing the community because I'm following a three-legged stool strategy that a lot of companies do. And so as part of that strategy, you know, in the case of Disney, there's on the right and AWS is on the left. But for me, it's like code. So SAS, AI first SAS, as you can see here, AI first marketing, which is what I'm building right here. I'm doing it manually, but soon orchestration software that I showed you the diagram of earlier, will be like basically doing 90% of the work. And then community, I think it's just great. So that's why I do the community now. But yeah, it's half off right now on school and then I'm rolling our own product that'll have a bunch of AI first features. So the price will be 100 bucks a month, but it's going to be dope and um really excited about it. So check that out. Subscribe to the channel. If you didn't know, I have another channel that I'll be doing builds on. So that's three videos a week, 20 minutes each of like basically I go and I build something that I need or that someone needs that I'm working with and then I record it and I compress it down to the sauce. Now that's finally it. All right, I'll see you tomorrow.