Cursor AI Just Dropped the Ultimate Big Codebase Playbook

I'm Parker Rex. Welcome to the channel. This is a daily upload where I give you the best AI news that you should care about. I go through a couple AI resources that are helpful for me and we keep it short and snappy. If you guys have any questions, we cover those, too. I'm going to clap because my audio is off 200 milliseconds. I can't believe it. I just see all these test clips. Bananas. So, let's jump into it. First of all, we're almost at that point where people can just have an AI write all the code for them. Kind of crazy. We're going to get into that after we cover a couple of questions. So, probably a lot of more more so just remark. Dan asked why Stripe? I've been trying to integrate it into Next.js projects recently and had the worst time of my life. That's no fun. Dan, here are two starters that will help you out if you guys ever want to integrate Stripe into your projects. There's a nice SAS starter that Verscell put out that has it ready to go. You literally run a script and you're off to the races. And then if you don't like that and you want to use Polar, then I recommend checking out Midday's project. They're really, really helpful with showing you exactly how to do that. So those are the questions for today. Now I want to talk about this. Now I saw this on Reddit and there was a spoiler alert. I just covered this and total scam, totally someone just making something up trying to promote their own product. Moral of the story is be careful what you read. Get really good at the tools that you already have. I highly recommend getting good at writing PRDs, using Taskmaster, using cursor. Don't jump on every single trend. Get amazing with what you are comfortable with. Last on the news front is with Shopify. So, we saw this letter come out from or sorry, not Shopify, from Duolingo. We saw we saw a letter come out from Shopify a little while ago with something similar that basically is a stance that a lot of CEOs are taking to say, "Hey, we're using AI. We're going to be AI forward and you need to get with the program." So kudos on you for even watching a video like this because it means that you're thinking about it. It means that you're pushing yourself. But in that all hands email, the founder and CEO of Dolingo essentially just covers, hey, Dolingo is going AI first. It's changing how we're working. It's not a question of if or when. It is currently now. And that's another thing that I've talked about quite a bit on this channel, which is AI is not coming. It's here. It has already produced a real estate agent that's done hund00 million in sales in Portugal. Kind of crazy. And we need to be prepared for it. We need to figure it out so that we can help others as well and so that we can build happy lives around it. Next on the docket is another post that I really agree with, which is why dedicated vector databases are a scam. And if you were building in the AI space right when all the chat GPT stuff was coming out, then you probably recognized this trend that was going on, companies were raising a whole lot of money for this new fancy thing called a vector database. And there are still cases where if your dimensions of the vectors that you're creating have a higher number, then you kind of have to do some tricks to get it to work with something like Superbase. But in most cases, it just works. And there are ways around that. But he basically talks about how he was spending $800 to process all of the SEO pages that he wanted to have vector search for. And then when he switched to PG Vector, he was able to basically have all the vector searches run on it for $20 a month. So the monthly cost would go from $200 to $20 a month. And one of the biggest biggest advantages is being able to leverage standard SQL capabilities with my vector data. Can now use foreign keys, joins, and all the SQL features I'm familiar with to work with my vector data alongside my regular data. Having everything in the same data database makes quering and maintaining relationships between data sets incredibly simple. Yeah, I agree. One of the biggest nightmares with Pine Cone was keeping the data in sync between Pine Cone and my Postgress database on Superbase. I have multiple data in ingestion pipelines into my system and need to perform daily updates. Please don't be like me and fall for the dedicated vector database scam. The article I'm sharing echoes my real world experience. Using your your existing database for vector search is almost always the better option. And he puts a nice template in here which I'm going to be using as a workshop for I vibe with AI which is this one. So this has superbase off using SSR of course that's the new way of doing things but then it also has retrieval augmented generation or rag and tavalene which is web search. So I talk about this but it's a good way to learn is to just kind of sift through these projects and figure out how are they working. So, you can hit the T button and then just start typing in, well, how's the main page work? Let's go and check out the let's say the AI chat page. How's that working? Oh, they're using cookies. They need that cuz they're making a call to something that's probably for the authorization. They're using next headers. This is a server component. You can see that they have a document component. What's that do? And then stuff opens up on the right side with all the different symbols. So you can kind of click around without even leaving. And I found out this track from John who is in our community. When you hit the period button, it actually opens up a web editor version of this, which is really cool. And then you're actually in the IDE and you don't have to pull it down. So shout out to John if you're watching this. Thank you for that awesome tip. And now you basically have VS Code running in here and you have the fancy modal thing. That's a rare pattern from Next. Next up is an article on large code bases. Oh, sorry. I did want to wrap with why the dedicated vector databases are at scam. At the bottom here, we have a lot of people agreeing. But then one thing I wanted to just cover is if you don't understand how vector works, then it's not that hard to understand. Imagine you need to store simple XY coordinates in a project, maybe longitude and latitude. To search that in raw SQL, you would have to sort the values and do a B tree search. Instead, you can search if you have a vector database or vector data types by finding the closest nearest coordinate in your data. Now, imagine instead of X and Y with two dimensions, you have 248 dimensions. So, it is a simpler way of thinking about it. Now you can know that text embedding 3 large uses 372 but you can still use this in PG vector without the hnssw index. I actually don't know what that means but I just know that if you wanted to have the very best then you can check out this one. It's better than open AAI only has 1024. You can store them in int8. But in a case like this, if I didn't know what this means, what I do is I just pop open this and I'd say, can you explain this to me in a context I understand? And because it has memory of me, it's going to write something that's helpful. What's happening? Text embedding three large returns 372 dimensional vectors. PG vector is a Postgress SQL extension that lets you store and search vectors like embeddings. You can store these in 372 length vectors no problem in Postgress but the HNSW this is what I didn't know index which is the really fast approximate search method and PG vector supports only vectors up to 2,00 so you can still store and search text embedding three large vectors you just can't use super fast HNSW index you have to use IVV flat or brute force search which are slower as a database grows And then you get more and more and more examples. So it provides options. But that's typically how I do my research. So hope that was helpful. And I wanted to just quickly touch on this article that I came across on the cursor website. So it talks about using cursor for large code bases. So working with large code bases introduces a new set of challenges than working on smaller projects. Drawing from both our experiences scaling cursor's own codebase and insights from customers managing massive code bases, we've discovered some useful patterns. In this guide, we're going to walk through them. So, you have this nice chart where it talks about building the codebase, understanding, defining the outcome or the diff that you need to get done, planning the changes and then implementing them. So, you'll use chat to quickly get up to speed on unfamiliar code. So, you want to cruise around basically and let's see what it says. With chat, you can start asking questions to find what you're looking for. And in this case, it looks like they have a nice read me. Let's just watch the video. How is the chunker for indexing implemented? Show me the relevant code with some examples. And what's interesting is they're actually in agent mode. So, when I think of chat, I think of inline stuff, but I guess that is chat with agent because then it's allowed to do all the different searching and the digging through and graphing of things. to give cursor a heightened understanding of your codebase structure, be sure to include project structure from settings for improved performance. And then it mentions writing rules for domain specific knowledge. So if you're onboarding a new collaborator in your codebase, what context would you give them to make sure they can start doing meaningful contributions? Your answer to this question is likely valuable information for cursor to understand as well. For every project, for every organization or project, there's latent knowledge that might not be fully captured in your documentation. Using rules effectively is the single best way to ensure cursors gain the full picture. For example, if you're writing instructions for how to implement a new feature or service, consider writing a short rule to document it for posterity. Cool. So they say basically write rules, add a new VS code frontend service, define a new service. Okay, so basically just saying if anytime you want a new front-end service, this is how it should work. That makes sense. If there are common formatting patterns that you want to make sure cursor aderes to, consider autoattaching rules based on glob patterns. Yep, this one's nice. The glob pattern says these are the things these are the files that you should attach this set of rules to. So in this case, anything that ends with a ts should use bun as a package manager. Should use kebab case for file names. Use camel for function and variable names. Use uppercase snake for hard-coded constants. Then it has some stuff around the syntax where it's like, "Oh, I don't want error functions." Cool. And then you want to stay close to the plan creation process for larger changes. Spending an above average amount of thought to create a precise well scoped plan can significantly increase cursor's output or improve rather. So this is where I think the PRD comes in and this is where Taskmaster comes in with this kind of combination. I'm really bullish on Taskmaster still. I don't want to touch root code. If you find that you're not getting the result you want after a few variations of the same prompt, consider zoom consider zooming out and creating a more detailed plan from scratch. As if you were creating a purity for a coworker. Often times the hard part is figuring out what change should be made a task suited for well for humans. With the right instructions, we can delegate some of the parts to cursor. And it goes on to say that one way to use AI to augment the plan creation process is to use ask mode to create the plan and then turn on ask mode and cursor and dump whatever context you have from project management systems, internal docs or loose thoughts. Think about what files and dependencies you have in the codebase that you already know what you want to include. Yes, exactly. So they have a planning prompt which I think is a nice one. So you can include the past chats. We're asking the model to create a plan and gather context by asking the human questions referencing any earlier exploration prompts and also the ticket descriptions. Use a thinking model. Yes, love this. From this you can iteratively formulate the plan with the help of cursor before starting implementation. So context ask plan agent implementation. I talked about this yesterday in our school, but I am working on ivibewithai.com, which will be our replacement for the school, and you should join now because it's going to be half off what the price is when we launch. But this is the same process that I do where you're gathering context, you're coming up with all the files, relevant relevant paths, example code bases that if you're doing a feature and you want to learn from, you can bring in. And I use ask over and over and over again. I'll probably have eight to 15. Sounds crazy. Different iterations of the same kind of p uh I'm basically starting with a block of granite and you're chipping it away and you're shaping it. You're going from this very unshaped piece of work to something that's much more shaped. Then you can flip into the agentic mode. And that's when I'd be using Taskmaster to parse it out. Take the final PRD. Have any open questions. I'm making sure I'm answering those. And then finally the PRD gets turned into the format for Taskmaster. We're off to the races. Picking the right tool for the job. So you want to use using tab for those quick infile manual changes. Command K for scoping to the single file. Chat for the larger multifile changes. So chat is everything on the right. They used to call it composer. Now it's just chat. So each tool has its sweet spot. Tab is your go-to for quick edits where you want to be in the driver seat. Command K shines when you need to make focus changes on a specific section and chat is perfect for those bigger changes where you need cursor to understand the broader context. When you're using chat which can feel a bit slow help it help you by providing good context. This is where I talk about the relevant paths folders for understanding the project structure and then it has some sort of just like takeaway here. What I also find helpful if you're doing stuff where you go like this is all just in cursor but if you want to use grock let's say the closing thing that I'll cover here is a freebie for just generating a tree but like if I want to go from my ID to gro for whatever reason there's going to be a couple things that I'm going to need to do. I'm going to need to do a repo mix and this is all the files that I want to have exported and made loom friendly. Then I'm going to need optional a tree but I believe repo mix includes that. But what tree does is if I do at tree I'm going to get this. And this tree says we're going to go four levels deep. We're going to ignore the node modules and the git. What that'll do is it makes a nice looking tree. So I can give you an example of this in cursor. If I just pop that open real quick. So let's just this is a YouTube to blogs thing. So let me grab that cuz you won't get that in the command line. And let's make it bigger. But if I just run tree then you can see I got a nice tree of everything that's going on in the project. So if you weren't going to use repo mix and you just wanted the tree, you could do that. But that's a nice one that I have bound to at@ tree. Now, I hope the quality of this video is much better. Had to do a lot of work with just literally the audio to get it synced up for whatever reason. A lot of clapping. But if you guys have any questions, make sure you drop them below. I'll be happy to answer them tomorrow in tomorrow's video. And make sure you like the video and I'll see you in the next