$200/month Cursor Ultra or Claude Code Max?

200 bucks a month. Who you got? Cursor Ultra or Cloud Code? What is it that you're going to be cooking with?

I know it's a lot of money to be spending on these AI systems, but if you're seriously coding with AI and using AI as your use case to generate a lot of code, I have a really special treat for you today. I have someone, his name is Robin, and he has about 20 years of experience in the industry. And not only that, he spends his mission to teach people, um, like people who haven't coded before how to actually use these AI systems to code, make apps. And he's got a whole thing that's going on. And um, I wanted to have this discussion today because of the latest pricing changes that have been going on with Cursor, uh, and also the other agentic platforms as well. So it's no longer 20 bucks a month to code if you're primarily using agents. It's now 200 bucks a month, you know, if you're going to be using it like really full-time, like the way most people like to do, you know, vibe coding with. And I code a lot with using the latest clait models and I'm primarily using it in the terminal using Claude Code, uh, combined with Cursor. So, um, Robin, thank you so much for joining the show and excited to have you on. Tell me a little bit more about, uh, your experience here.

Hey, yeah, thank you for having me. I'm I'm happy to be here. Um, yeah, a little bit about myself. As you've already so nicely introduced myself, I've been a coder for over 20 years. And now with AI, I rediscovered my passion really, and I really like everything that's happening in the AI space. I started picking picking it up very quickly and started making videos about it because I thought I have a lot to share with people. I think obviously, you know, everyone can build something now, but as engineers, we do have an edge, right? And I wanted to use this edge to teach especially non-technical people how to use these very overwhelming tools at times.

Yes. Yes. Yeah. It can be overwhelming if you're starting out, especially to start apps and kind of get started. And I myself, es like, just because I have experience sometimes as an engineer, it can be a little more challenging to kind of guide someone along who's just getting started. And I'm so happy to find someone like you especially online, uh, because there are a lot of people who join my sessions and like, hey, how do I get started? And so I I think you're a really great person to kind of get talking a little bit about this. Um, and so what as far as like coding tools, like when you're using Cursor, um, like I know like you know, we've been doing a lot of like just telling the agent to do these types of things, um, like what's kind of your your go-to bag as far as like how how would you like advise someone to kind of get started and like why did you choose that 200 bucks Ultra plan as opposed to all the other options and all the different things that are out there?

Yeah. So I have a little bit of a secret, you know, that you cannot tell anyone, right? But I actually do not only use Cursor.

What?

So yeah, I know, right? Unbelievable. That's a like a like a bombshell. But no, I I actually use I mean obviously it's our job to use all of them, but I really like Cursor. I know there's a huge witch hunt going on right now, right? Everyone's like hating on Cursor. And I think some of this outrage is, uh, is justified, don't get me wrong, but it's a good tool, right? It's just a great editor. It's not perfect, but I just today I used it all day. But there are also days where I use Claude Code all day, right? So, it's like really I'm always torn between Claude Code for like maximum output, which I think is really good, and Cursor if you want more of like a full-service suite, if that makes sense.

Got it. Yeah. So I went from coding in Cursor, and the thing that I was missing was this whole thing that would allow you to do the to-dos, uh, and then giving like stuff to the agents so that they can go do these other things. So I was super impressed when Claude Code came out for me a while back, but then I was like, okay, this is, you know, I could still do a lot of stuff in Cursor, so this is not really much of a big win. Uh, and then it wasn't until those other features where I just started seeing it do more agent tasks and I didn't have to give it much more things. So I would see it do its planning and then it would run a reasoning step and it's like, oh, let me go ahead and see if I could find this file in my in the codebase, and it would go find it and it would come back and I was like, oh, that's so nice. It's it did it a lot better than when I was using Cursor. Sometimes Cursor would find it, sometimes it wouldn't, or Cursor would be very anxious to go and just write code immediately and just add more to my codebase. So maybe something for me and maybe some other people too who have experienced this. If you're an experienced developer, you kind of know not to immediately just start writing new code to solve a function. There's probably something in that codebase that already has it solved. Uh, especially if it's mature and it's been there for a little while. So how do you think about something like that and leveraging systems, um, in your code bases to either like use an existing framework or kind of how do you, you know, differentiate your instructions and stuff?

Yeah. So I have so many answers for that, right? It's um, so number one, yeah, you're totally right. I think when you are a non-engineer building with AI, you're gonna have all these like Claude especially loves to do that, right? Like Claude Sonnet and even Opus, they're like they they will sometimes they'll come up with like an like a super complex thing to do and then they want to write like a documentation for it, and then you ask them, okay, cool. like in in like two or three chats later, you're asking them something and it already exists there and they know it exists but then they come up with a different solution and then they think like, oh yeah, we can make it simpler, and then they take the same file name and say like _simple or something like that, you know, it's it's ridiculous. So it's yeah, I think the there's like a trigger that especially non-engineers that they need to develop in their minds and that is always ask always ask questions and slowly I see this like in the last few weeks I see it a lot more that the the correct word that is that is starting to be thrown around is context, right? like how to engineer the right context so that the AI actually does its job, and that is something that I I think we as engineers we know that we know, okay, we already have the function to look something up or something like that, but when you're not an engineer, there's always a case to be made that you should be asking the AI, especially with models like 003 if you have access to it and you ask it something and be like, yeah, I want to do this, can you like check if this already exists if there's something that we can reuse or if there's some duplicated code in the in the codebase or something like that, right? But this is just something that it's a bit much in the beginning, but yeah, uh, if you're a nontechnical person, definitely ask. And if you're an engineer, well, you hopefully know.

So, you said a keyword there and I haven't heard that for a while, and it's using the 003 model. So, I forget about it sometimes because I'm so heavily into Claude Code, and 003 is one of my go-to models that I use on my phone with ChatGPT, especially for web search. The agentic web search and the subsequent calls of of its its flow is really really good, and when I do do some queries with just Cursor by itself, uh, I t 003 is actually my default selected model, and so um, what are you using 003 for like in ChatGPT and are you like switching models all the time kind of like what is your kind of thinking there?

Yeah, I mean there's there's one thing that I want to I mean people think I'm a Cursor fanboy, right? There are good reasons of why I use Cursor, right? And like having access to multiple models is a really good benefit of using not Cursor specifically, but a more full suite of editors. You might use VS Code, if you you can use Tray, whatever it is that you want to use, right? But 003 specifically, as you already said, yeah, web search, doing research with it is amazing. But coding specifically, every time I have a bug that I don't really as an engineer, I could go into the code and find this bug, but it's way faster to just ask 003 and be like, yo, this doesn't work. Analyze it. And I keep you not if you do not have experience with 003 and seeing 003 in action then I have seen 003 go away for five, six, seven, ten, twelve minutes and analyze code bases. There's not even that much code in that codebase. So I'm I'm working with very small things, right? like it's very like small demos and small apps that I'm working with recently thanks to the YouTube channel and whatever. But it just creates almost I don't know it feels like it creates almost its own little like neural network to like think about how this actually works, how everything comes together. It's not just reading lines, but it is an incredibly underrated model and amazing. Yeah, use it for that particularly.

Wow, that is really really good insight that you say that and maybe something that we can share with our audience as far as how we think about coding. So when I start a new chat, it's like you basically, you know, clean off your table and everything's completely fresh. And so in essence, you're setting this intention for I want to do a problem-solving task, which is to look through my codebase and help debug things. And so you're saying 003, you know, go and do that thing. So 003 almost acts like your assistant to start going around, you know, all around your house to grab everything that's relevant for you to solve this problem, right? so in essence, um, you know, Cursor is that like table and that everything is coming to the table and and it itself as an integrated development app, um, may actually already have context and it's possible that it's even using other models underneath the hood to pull that in so it can give it to 003 to build that, you know, basically everything that you need to do to then hand that off to the model and so forth and then you'll see it do these subsequent steps, um, that it's like doing this, it's thinking, it's doing this and then kind of following that chain, uh, to go maybe yeah, solve a problem or do something like that. And so, um, you know, from there then it makes the code and so forth. I think one of the the nice experiences that I like about Cursor is the fact that it does these little checkpoints. So if you're having a conversation, you can easily snap back and just revert all the code. That's something I don't really have I can really do like on Claude Code at all. like I just have to make Git commits like at every question I ask sometimes. Uh, cuz I know if I know I'm going to do something pretty big, I just Git commit and then I ask my question so I can revert back. Um, but it's just it's not as a as as a clean UI. It's not as a clean experience. I'm already using Cursor as my development environment. So, um, it is what it is. So it's just kind of like, okay, I'm going to use the the Claude Code as a shotgun and then I just kind of picking up the pieces afterwards and so forth.

Huh. That's interesting as far as like debugging and so forth. Um, what about like rules files? How do you I know a lot of people have a lot of different theories and stuff. I kind of want to get your your input on this. Like, you know, what's your take?

Yeah, people ask me about rules files all the time and I have a very very underwhelming approach to them. People always they ask, Rob, what are your rule files? Like like it's some like magic, right? But there there's no magic. I have if I go into my rules and they are basically the same across all the agentic tools that I use. I have maybe three or four of them, that's it. There is no magic. You always see them like on X.

Wait, wait, wait. Is one of them saying, uh, call me Poppy? Do you say like what?

No. Every time I chat with you, call me daddy or something like that. I've seen these.

Yeah, I've seen these. No, no, no, no, no, no. I don't have that man, but I've seen this. Drop drop drop your rules rules files in the chat. Yeah, someone made the case that this is actually like working, you know, that it's like somehow it's hard to argue that it might there might be something that's going on that make it like more listen to you. I don't know, man. It could be, you know, but no, I have I have three or four rules. One and it's basically whenever you come across because these models, they always behave in a specific way, right? and they have a surprisingly predictable outcome despite being quite random sometimes and so they make the same mistakes again and again. So for example, it might make when you initialize like a JavaScript project with like Next.js and you want to use Tailwind or whatever it will very often make a mistake of like mistakingly using some like old syntax, you know, old setup method or something like that. The same is true with this like I'm going to butcher its name, but there's this like chat UI library and it it changed its name. So it changed this like package name and then I have seen probably like 200 times it tried to like pull it in and then it always gets like a a deprecated me like message and then it says oh no it was renamed. So I just whenever I run into problems 50 times at some point I get tired of it and I make a rule for it and that's how I come up with my rules.

Wow. Yeah, I think that's a good way to do that because people ask me the same thing. How many MCPS do I have installed? How many like rules files do I have installed? And I think the secret now is like actually removing as many rules as possible and just starting fresh.

Yeah. So whenever a new model comes out

Yeah. Go ahead.

Yeah. Mhm. Yeah. No, sorry. No, even even MCPS. So I noticed that I'm I'm working on like a new video where I will compare MCPS and I noticed that when I add them into Cursor, yeah, Cursor I use Cursor every day guys, then at some point it's going to like tell you even at like 50 tools or so that more tool the more tools you add the more confused the AI gets, and I think people are not aware of that. I wasn't aware of that. So now I actually had to go through because a lot of these tools like Supabase and Stripe MCP they have like 50 tools on its own. So, it's like it's pretty it's pretty wild. You always need you almost need a layer on top of these MCPS to like have fewer tools. It's like yeah, I don't know how to do it.

Yeah, I I feel like I have to be very surgical and say, "Okay, I'm going to do database work today and just turn that one on for Convex and then just ask the questions and then turn it off when I'm done." Um, and so I was having a chat yesterday with Manuel, um, his he's his name is Program with AI a couple days ago on the stream and we were doing prompting and we were doing like minimal prompting which was really fascinating like in three words he would just put part of a class and it would just generate a Python class just from those three words like HTTP and then like either with put the colon or without the colon or just the way he does an underscore for the name of the class and then the result was very different and I I found that fascinating because what he was saying in his observation of prompting is that whenever he enabled tool calling, the model's behavior went from being very creative to just very restrictive in terms of like very dull. It's like, oh, I'm just going to do this. And so he noticed that his output, like if he was doing creative output or creative writing prompts and it did a tool call, it just got less creative. And it's like, oh, that's interesting. And so the reason why we're talking about this out loud and the reason why I want to have something like someone like Robin across this thing is like we're seeing different things where you we could even be using the same underlying models, but if you're using them through different tools, you actually could get different results or um, you know, different outcomes in different ways. So it's it's really awesome to just kind of, you know, hop on the stream like this to share our perspectives. I know other people in the chat in the discords as well have been sharing their perspectives and we kind of have to talk about this because this type of AI systems are totally new to everyone. Even the people who make the AI models themselves don't actually know what outcomes could possibly be out there, you know, and so, um, that's kind of, you know, our exploration here and and I'm just, you know, really happy to have you basically on on the stream here and, uh, kind of chatting a little bit about your workflows and chatting a little bit more about kind of what you're doing. So, um, in terms of tell me a little bit more about, um, what is like you have like a community and then you have people you're helping and kind of what are some of the things that they're seeing when they're using these models.

Sure. I mean I I have a very like streamlined approach, right? Like a lot of people they ask me like what's the best AI stack and whatever and then sometimes they they send me like a like a thing of like 20 20 things like frameworks and I don't use most of them in my projects. I have a very very minimal approach to everything, right? And I think it's, uh, I I try to help people in like two ways, right? It's like the first thing aside from like the the the program that I sell, right? But like in a general thing, I I try to help them in two ways. So number one is like somehow validate your idea. And that is something that that no one no one wants to hear, especially when they like hear that now they can like use AI tools to like build their ideas. Maybe you're not a technical person. Maybe you're a junior engineer and you have a way bigger idea than than than you should have for the skill sets that you have at that time, right? And now AI comes around and you're like, "Okay, maybe I can do it." And instead of and I have that conversation probably three times a week that people are like, "Yeah, I want to build this and this and this and this and this." And the list goes on and then like, okay, I hear you but have you collected a payment from someone? Have you collected some emails, you know, of like people that actually want to use this? Right. Have you considered building a minimum viable version of this project? Right. Yeah. So that's that's number one. And then the second thing so simplification like in the idea and then simplifying the idea also helps you simplifying your stack, right. And there's there's so like I literally I can count the the number of tools that I use or the number of frameworks I literally use like Next.js. I'm a JavaScript guy. I always was, right? I used to be a PHP guy, but wow. But uh, no, I'm a JavaScript guy now. So I use I use Next.js as a stack. I use Supabase for my database needs most of the time because I'm already paying for it and it's very easy to use. And most importantly, language models know how to deal with it. That's that's really useful. and Tailwind CSS and, uh, that's pretty much it most of the time, you know, and where where do you deploy your apps? Vercel. Okay. Vercel. Yeah.

Just really easy to do that. Yeah. So that was that's one thing. So easy. Yeah.

When when I first got started on it, I was like trying to enter the space through Python, because I just kind of, you know, I did a lot of scripting in Python and I did a lot of client-side apps. So it's like it's not not apps but like actually frameworks and system-level things. So, you know, shipping to the web was kind of very foreign to me. And so it's like, oh, I can just run this Streamlit app and then like when you get users to use it, you know, like you can't customize a lot of things. You can't easily throw an off-screen with payments and do these types of things, and it's like, okay, I need a really good framework for this, and then that's kind of how I bumped into Next.js, JS and Tailwind CSS, and um like the maturity of those systems are very nice, but they also like it feels like driving a like a really complex like race car, cuz you can actually go crazy with it. Um but there it kind of looks simple on the outside too. So um yeah, it's really interesting.

And then I think the biggest unlock was when these models started to actually code with them very very well and like generate UIs and that stuff. And that was kind of the biggest breakthrough for me. And so um as far as deploying, I used to use Railway. Um and I occasionally I'll do that if I have like some Python projects I want to have and with some FastAPI type of thing. Uh but then when I discovered Next.js when I was deploying to Railway, I noticed it was very slow. And when I just used Vercel, everything was really really fast.

But um one workflow I did as as an engineer was like I always try to have my GitHub hooks. So whenever I made a feature branch, I would try to, you know, set up a different um actual deployment just for that branch so I can track it separately from the main. So you know, it was like it took a while to devops this stuff for myself. And then when I moved to Vercel, it did it automatically. It's like I didn't have to think about it. I just branched. I was like, okay, how do I set up the web? Oh, it's already set up for me. That's great. You know, it's just so easy. And like even the comments back into So that takes a lot of work to set up. And I do like the developer experience for that, which is really nice. And um they have a lot a lot of these nice little touches that they do um to make that deployment step super fast. To me the platform is the it's a no-brainer. It is I would call it I would call it unreasonably good. Unreasonably good because I get approached, you know, we have we're both like you know largerish YouTube channels now and you know people come to us, brands come to us, and I got like big hosting companies come to us and they're like, hey, we think you're like a perfect fit, and I'm like, yeah, but I tell my like students to deploy on this service, which is like free for most of them, also it's super easy to use and have all these integrated services, and they don't understand it, they just don't get, you know. Yes.

And like exactly what you just mentioned, right? The automatic deployment of the of the of the branches, for example, when you combine something like this with the autonomous background agents like from Cursor or Codex or even Augment Code, all of them and now you can literally I know I'm I'm shilling Cursor again but like now Cursor has these like web agents right and you can like be or in Codex, you know, if you're a Codex guy, that can you can do that too. And you can just ask it some I said I saw like a like a like a thread from someone on X the other day where someone was like, I found the perfect use case for these background agents. The guy was like, sometimes I'm just like in bed and I take my phone and I I tell like Cursor or Codex or whatever or something and then when it w when he wakes up he has a branch deployed that he can test. Yes. Yes. It's pretty cool.

It's pretty cool. I did that recently because uh my front end I I changed it from Cloudflare to so that's the other problem too. So with Cloudflare like the branching is not as easy because I have multiple branches sometimes and you have to specifically select the branch that is going to be available for you to be deployed on the Cloudflare. So you can't have multiple feature branches that are being deployed, which kind of sucks and it's like ah so I switched it over to Vercel and so I had this same idea. Was like, "Okay, the footer for my landing page needs to be updated. So, I want to give Claude Sonnet, you know, from like using Cursor background agents, right, on the web." I was like, "Can you just go fix my footer? I don't have a lot of stuff available. I want you to dig through the code. Um, and I'm trying to launch this MVP. I literally have zero and I'm trying to just go to one and I just need the bare minimum." So, it was cool. It was a perfect background agent task. I was on vacation. I was on a boat, you know, and so it did its thing. I didn't have to worry about being connected to the internet, you know. I just had to like have enough internet just to send that over barely and it spun it up and then I was able to preview the page on a Vercel link because Vercel gives you previews for all of your branches and features and stuff like that and then like I was like this is sweet, merge done, you know, it's like it was that fast I was like that's actually really cool so you know I I could dig it I'm a big fan of that.

Hell yeah, I had such a similar experience I was probably on a plane and plain Wi-Fi. And I was like, "Yeah, I have some ideas, you know, and I had so that's going to it's going to trigger some people because I'm I'm talking positively about Cursor." And now I'm going to say that Cursor actually gave me early access to to the web to the background agents and they influencer. But they gave me early access. So I was like, you know, on the on the uh plane and I was like, yeah, I was exactly as you said, I I didn't like it like playing internet sucks. I mean, now it's getting better with like stalling or whatever, but not that was not a stalling plane. So, it was pretty bad internet. And yeah, I was just like testing things and there was no way for me to actually test them because the internet wasn't strong enough for that. But I I had just enough internet to shoot off like any idea that I might have. And I did, right? And then when I got home, I could actually test them. Well, half of them couldn't be deployed because they were like Typescript errors, but hey, they were they were there, you know, it was good enough. Yeah, it's pretty cool that that that whole process could even kick off. And background agents are pretty cool. And so background agents actually are included in even like the $20 a month plan from Cursor, which is really great. I think the thing that's are they? Yes. Yeah. Are they? Because someone someone just asked me about this and then I tried it and I didn't have usage basing on. Oh, I do. Now, now, now, now there's my chance to to to trash Cursor a little bit to make, you know, make it a bit more balanced. Please clarify. We're trying to figure this out together, y'all. I actually think it's uh So, what I tried because it's super confusing and again like now is my chance to like bring a bit negativity into Cursor. But the they say Max is included. Okay. And then on their background page, background agent docs, they say like, "Oh, you need to have max mode enabled to do that." Okay. But then you try to like put one on like launch a background agent without usage on and doesn't work. Doesn't work. You have to turn on usage. Oh, I see. Usage. Let's see. Extended usage limits on all models. So for many users with casual agent usage, access to background agents. That's what they say. Oh, access to background agents. Oh my gosh. Access to Bugbot. Access to background agents. Yeah. No, Bugbot is not included at all. Right now, there's a trial and it's cool, but it costs in my small projects, it costs about 25 to 30 cents per Bugbot uh initiation. I guess it's not cheap, but it's a cool it's a cool thing. And you don't have to do it automatically. I I looked into the documentation. You can actually turn off the automated Bugbot and you can mention Bugbot in the commit message if you wanted to take a look which is cool. Oh yeah, there's a toggle if you go to I think it's like /agents or something like that on Cursor's website and then at the bottom they'll say this toggle say like just I just want you to call it when I hit the app symbol otherwise yeah it'll be running and kind of churning through all of your ideas and PRs and stuff. That's super duper helpful.

Yeah, I'm kind of curious because if people have some questions and stuff, somebody was asking um that they're still debating which is better. Um would you throw $200 plan on OpenAI as well and kind of what what what are you thinking like here? Actually, Rob Robin, like what do you think as far as like your usage so far? Things that you've heard, things, you know, are you still curious? So, if you have like another 200 bucks, what would you throw it at? Like if you had to rethink everything again knowing everything you know now, you know, what would you do? No, no, not at all. I I'm actually going to lean myself a little bit off out of the window here and make another like enemy in my influence at times, you know. But OpenAI I would say is like the most the most overrated subscription and it's the only AI subscription that everyone has that I unsubscribed, you know. I don't even have Plus right now. I have nothing. Nothing because why, you know, like why would you have ChatGPT? I pay already for Claude because I use Claude Code. And uh hold up, hold up, hold up. The reason why you have it is because they're part of the mafia. OpenAI is part of the AI mafia. I don't mess with them. Yeah, you don't mess with them. You have a subscription. I Exactly. I I have a subscription. Blink three times, right? Blink three times. It's It's just not worth it. Now they They're not paying us, by the way. No one is paying us any money for talking about them. We're talking freely just like you and me using this apps and actually how we are. Um just because we have x amount of followers doesn't really mean anything to be honest. It's just you know everyone's showing up together and that's kind of actually why I wanted to have these conversations because it does kind of feel that way sometimes. There is that like odd pressure that people like may feel if they're getting paid by a company where they have to feel like they have to talk positively about them or maybe it's even in their contracts that they can't actually put them down. Um or also legal liabilities and stuff. So with that said, everything that we say here is our own opinion and our own expression and please do not use this as any advice or anything like that as far as you know, please do your own diligence and please kind of make sure that you do um you know your own research and this is kind of on your own things. uh you're listening to our raw opinion. It's just our own thing. Uh I'm talk I'm kind of giving you insight as us talking as two human beings. Um and so anything we say here um should not be taken as advice or anything like that. Uh with that disclaimer out of the way, let's go.

No, but like I honestly I think like ChatGPT's problem is that their free tier is so good, right? that they free let me talk to it and like ask tons of questions and yeah I I don't even know I I might still be I don't know is I unsubscribe from Pro but I'm I might still actually be on it let me check that's wild I might be yeah actually I actually am I'm I'm I'm canceled but I'm still subscribed but I don't miss it it's like it's I barely use it you know that's interesting because I the 200 bucks a month Pro plan. And so the primary thing I've been doing is just flooding GPT-3 with tons of requests. Um I'm not sure if you get the memory feature on the free plan. I think that's kind of a big deal for me. Um a lot of my workflow as a um as a person who uses this a lot is like the memory feature is huge because I have a lot of these like project folders and each project folder for me inside of ChatGPT deals with a specific use case. So um my use case like for example launching this app I'm having this conversation with marketing and like I'm treating it kind of like a marketing person each little folder in ChatGPT and then GPT-3 is really good at digging through all the memories. So I'm not really sure does paying for different tiers get you an enhanced context window. Uh and if it does you know how helpful is that? Actually that's a really good test. I should probably cancel my pro subscription and then see if I can use it for free and feel like that, you know, like do I have that drug dealer effect where like as soon as I don't have access to the drugs, I'm just kind of like I need memories.

Yo, so do you have do you use GPT-3 or GPT-3 Pro? I use GPT-3 a lot as like my default search for everything and then GPT-3 Pro would be occasionally like if I'm feeding in like some code stuff. So, um, that's kind of sometimes where I'll pop in repo prompt and it's like I want to refactor this big large repo. Can you just think through it? And to be honest, I actually prefer the output from Grok a lot better most of the times for my architecture. I love Grok. Yeah. Yeah. Grok is so under I mean that's maybe a topic for another day, but Grok But wait, let's let's let's rock up the OpenAI because take back what I just said. take back. Okay, because I didn't know. So, I didn't know that I was still subscribed because I canceled a while ago and now I'm going to continue being subscribed because I just checked and you need Pro or Plus for memories. I'm not going to give up my ChatGPT memories. That's too powerful. So, there there is the differentiator, right? So, I'm going to continue with Plus, but I think the Pro subscription overrated big time. Okay, 200 bucks a month. That's that's that's super insightful. Super duper insightful. I mean, what y'all What's y'all rolling with? Um, there's Codex. Codex did work for me, but I was a passenger driving when I was lying on the couch. Yeah, it it's to each its own in some ways, but I don't know if you know this Cursor background agents allow you to use GPT-3 and also allow you to use Claude Sonnet and Claude Opus. So, you can select all three, give it one task, and it'll do all three of them. You have to have probably pay per use obviously enabled and it'll cost you but you can get one task multi-threaded in a back all three background agents will kick off in that chat. So to me I actually find that a better use case than Codex um because Codex it's just so slow to spin up and Cursor's background agents are actually really fast to spin up and then the if you have Vercel hooked up the review and like the actually looking through the the code cuz I was able to spin up uh GPT-3 and then the Sonnet 4 and actually prefer the output from GPT-3 and I was like damn that's really good. Um so yeah it's it's good to know though but yeah a lot of people are on these different plans and stuff. Yeah, ChatGPT canceled. Thanks for the reminder. Oh yeah, exactly. That's so awesome. I think uh I think Cursor even has I mean Vercel probably too. They even have GPT-3 Pro. I haven't tried it yet because that must be so expensive. Oh, interesting. Okay, but what about Ultra Background agents? Is it still pay for use? I I think so. I think Yeah, I think so, too. I mean, yeah, I can I can I can test it right now because I have I have the Ultra plan. Um, I have the Ultra plan. What is the what is the the key again for for the Do you know the the actual like key shortcut to launch a background agent? I forgot. Oh, I I just clicked the cloud icon at the very top. Um like in that uh chat window or sometimes if you hover the mouse to the left side, a little pane slides up. Let me see if I can find my Cursor things. Um yeah, they just have this new feature now where you can slide it in uh from the left side, which I found really interesting. Um yeah, I found that extremely annoying to be honest. Yeah, I think you could disable it through the toggle. I just enabled it. Yeah, because I'm always going to the left and reaching for um my code and stuff and kind of reviewing those things. But um if you are Yeah, it's been a little while since I've used um but I think I think you're a full Claude now. Yeah, but I'm actually with with our conversation today to be honest. I'm actually thinking of of trying, you know, here command E. So, show background agents. So, then this will bring up the background agents window here. Um, I'm actually kind of wanting to go back and the reason is is mostly for the to-dos and that agent thing that they've been trying out. Um, and also trying out because they stole some of those Claude Code people, you know, this thing's cooking. Yeah. Yeah. No, for sure. For sure. But see, the thing is the everyone keeps talking about the to-dos. I haven't I haven't seen Cursor do a whole lot of to-dos on their own. No, it's not. It doesn't work in the same way that Claude Code does. I see. Claude Code you give a big prompt and it gives you like 10 to-do lists or something like that. But in Cursor I have seen it once one time it came up with this on to-do but the way more interesting I don't know maybe that's just my workflow but the way more interesting feature for me are the queued tasks, right? so if you it's just I don't I don't know about you but like sometimes you're like you're coding away and then again I'm a JavaScript guy so I don't know maybe your workflow is different but you run like a you run the test server or the development server you test the thing in the in the browser, but you already gave it another prompt to work on something else. But then you see, oh no, this button sucks. So then you go back to Cursor and you're like, do this, this, this, and then, oh no, you also need to change this. And it's like you just keep queuing up these tasks and then you just hear like bing bing task done. Bing bing task done, you know, and it's just I love it. It's just it's so much better than the than the to-do. Oh, I like this. Yeah. I I don't know why, but you're kind of selling me on this workflow again because I kind of missed some of this UI stuff. Um the the thing is, do you feel like you're going to hit a limit soon? And if you hit a limit, it's per month. Is that what I understand as far as the number of requests that you can send over to these models? Cuz that's kind of what I'm confused about with Cursor Ultra. Like I have talked to a lot of different people now and I ask people for like their uh their like usage and because I yeah I want to I I really want to understand and here's something that is super super weird. So some people they on the Ultra plan and even the Pro plan and they used maybe like $30 or $40 and mind you they're still complaining right that they they they pay $20 and get like $40 in in in you know API credits but okay but others have like $80 usage or $120 usage or so on the Pro plan which

Costs $120 right, and then on the Ultra plan, Cursor says that they give you at least $400. I have blasted through like 180 or so now. So, I'm nowhere near, but I'm going to get there by the end of the month.

But then I saw a guy on X who said he blasted through like $800 something dollar worth of API credits, right? So, even if they inflate the API costs, which some people are claiming, I haven't verified it, then you're still getting a significant amount of value out of your subscription.

But what Cursor has to be more transparent about is how, how, somehow, sometimes your limit is hit for a specific model. And then what a friend told me today is, so he hit the 03, or the set limits, then he went to 03 and was able to continue coding, which is weird because he blasted through enough of his, like, you know, credits that he shouldn't be—it doesn't work like that, despite Cursor's bottom line. Then later he went back to the previous model and was able to continue coding, right? Oh, so clearly there's something going on that is not just a plain any dollars credits and you can use them on whatever. No, that's not how it works. Clearly, that is not how it works, right? And but they are just not transparent.

And my hot take on that is that they have, because they always say that they signed a deal with the frontier model providers like Grock, OpenAI, etc. And I think that they somehow made specific—because there's always a word that Cursor uses. They always use “capacity.” Which models have capacity that you have to—if you pay attention to that, they always say that, right? So I'm thinking that they, that these like model providers, that they have some, some weird deal with them that allows them to use models maybe at specific times or something like that, like different, like different amounts of times at different times. I don't know. It's a very complicated deal. It's not like—I don't think they have a deal that tells them, yeah, we get you, we get you like, you know, $200,000 worth of compute per day. No, I don't think that's how it works. It's very complicated. So, they somehow they allocate it way more dynamically than they make us, uh, believe, you know. I think this nails it right on the spot. They're, they're basically pricing it like electricity usage. Like low-key, it's like, okay, are you tier one today? Is this like a brown day or a blue day? If the day is a brown day and you're working between, you know, 9 and 6 during peak hours, you're going to basically get this. But by the way, your peak hours depend on your time zone and who else is using it at the time. Yeah. Okay.

And so for the most part in Ultra, like the vibes is that even if you do hit a limit for whatever reason for that time, you can come back later. There's sort of like this cool-down period. But they also encourage you to use a different model, which you can still be okay with it sounds like as well, on top of that. And I guess the advice in general is like don't stick to the Opus model for everything. You know what I mean? Just because you have Ultra, don't feel like you have to get everything from it because it's actually probably not the smartest choice anyways. Um, I use Opus a lot, but like I actually switched to Sonic, cuz Sonic is so good. And then 03 is really good too. If you use Claude Code on the max plan, especially the $200 max plan, then go all out with Opus 4. I get it. But if you're buying like a $200 subscription from Cursor, I, I referenced this because there is a guy like that, a student, maybe he didn't know better, but it was a guy on X who had like a thread that kind of blew up and he said like, "Cursor, is this real life? It's been like, uh, it's been like 3 days and my subscription is already over." And then he posts, posts this history and it's all Opus 4 thinking, and I'm like, are you kidding me, man? Like, are you kid? You blasted probably through $1,000 worth of credit, you know? Like, it's just, it's, it's wild. It's completely wild, bruh. Bruh. Yeah. I, exactly. Come on, man. Come on. Come on. Come on.

So, yeah. I think, um, there's so much, so much design, kind of what, what team are you as far as like if you were to spend $200 on folks, drop it in the comments. Let us know where, where are you at with this. Are you, have you been coding with Claude Code? Do you like that workflow? Um, Cursor Ultra is extremely enticing to me cuz I'm already using the Pro plan. Uh, I get a lot out of it and I'm kind of combining the two. And so, um, I'm kind of going to start to revisit what the Ultra looks like in my daily workflow because ideally I would just like to have everything in one thing. Obviously, the reason why I went to Claude Code for me was that I didn't have to attach files and select functions and kind of do a lot of the hair-pulling I typically would do because I'm very precise with my changes and I want them to not have a lot of bloat. Um, and so I, I, I'm trying to weigh the two options and I'm always going to continue to do this. The thing that's really enticing about Cursor is like this less anxiety about specific models, uh, and then kind of letting the agent do its thing. Um, and then I know you've played around with Augment Code a little bit, just to kind of go off topic. Um, what, what has been your experience there? Because someone else who's been on the stream, tech friend, likes it a lot, but it's not like one model. It's like you tell it to do a task and then it starts to think through and it kind of does a bunch of things, like, have you had good experience there or like what's your thinking?

Yeah, so there I have to give a disclaimer because they sponsor me, right? And, uh, you know, take whatever I say with a, with a grain of salt. But I have tried it. I, I have it installed at all times and it is really good. I think they're overpromising in terms of what you just said because previously, like Augment, like three or four months ago, they, they did 01 at the time for like planning tasks and then set 3.5 or 3.7 for the actual execution. I was a bit sneaky and I just injected—they, they allow you to have like a customization prompt, right? Like an instruction prompt. So I just told it that in every request or in every answer it needs to, like, start in parentheses with the model name at use, right? And so every time since I did that, I found out they are exclusively using Sonnet 4 right now. So that's all they use, but they reserve the right to use any model on the background to do what you just described, right? But right now, as far as I can tell, and I've tested them quite a bit, they're using Sonnet 4. It's really good. In my opinion, they have the best web search. They, someone the other day brought to my attention that Claude Code has a web search. Yes. I didn't know that because it never does it on its own. Right. Right. I have never seen it do it. Cursor also didn't do that in the past. Now it searches a little bit more often. But Augment, it has any doubt, any even remote doubt about not knowing what, what it might, what it might need to do, it goes on like a rampage and like reads 200 websites, you know, to get all the information that it needs. I like that about it. I really do. It feels a little bit slow sometimes. And I did like a test in one of the videos where I compared it to Claude Code, I think at the time or maybe Cursor, I don't remember, and it is a little bit slower than the other tools, but it's really good. Okay. It's really good as well. Okay.

So maybe critique my workflow right now of kind of how I'm doing some features. So, um, I'm still new to web development. So like you said, like I'm using Next.js, JS, and I'm using a lot of the same stack that you're using and a lot of times, sometimes I'll just use 03, like just on ChatGPT itself, um, and say, hey, I want to accomplish this task for my app, like help me understand what frameworks I should be using. So I'm just kind of going back and forth in the UI there just to learn more. And this is just kind of what software developers do anyways before AI. So like you would have to talk to people, look up frameworks, search through GitHub and see, you know, who, who wrote something because you never want to repeat your code. Uh, see if you can find a nice lightweight framework to accomplish the task you can do. Uh, see if it supports your needs and you just like evaluation, right? So, um, and so a lot of times I'm kind of sort of repeating that old-school mentality that I had, but just using 03 as a web search to see kind of what it pulls out to kind of take me in a direction. It's like, ah, okay, I've never done markdown editors and so I'm trying to learn more about them and like who's out there that does stuff, who's talking about them, where are code examples that I can review. Um, so once I kind of get that idea down, I then kind of use that to try to formulate some instructions. And sometimes I'll just raw paste in the output saying, um, I'd like to use this framework. Here's kind of what I was thinking. And then just some of the output from 03 in our ChatGPT session so that the model has context. Uh, so then I'll use that maybe as my starting ground, prototyping ground, uh, to build out like a feature for me. Uh, and I find that to be really nice and very direct, um, you know, for that type of thing. And then I may open a new chat window after that and kind of do some more planning of like, okay, um, you know, can you review my code and see are there any hooks or other things that I'll need to do here? Um, you know, let's talk about the relationship between users and how I'm going to store the data, like what should I be doing with this output, you know, and like kind of like, you know, it takes longer to develop the feature, but I feel like in my head I can at least keep it in there. Um, cuz I'm trying to keep the code in my head/you know what's happening. Uh, and it's sort of my way of reigning in, uh, these things because I feel like I have enough context, probably more context in my head than most of these language models, even though they claim to have a lot. Um, I'm still the one that's directing them anyways. So, if I can understand it, then I can better guide these models to just, you know, do these tasks. Um, because I need to have that like pillar understanding of, you know, how a framework works or what are the ins and outs that I should be aware of or something like that. But that's kind of my flow. It takes longer to develop, but I do like the outputs and that's kind of why I'm impressed with some of these models. It's like, hey, they do pretty good output when you direct them really, uh, there. But, um, what's, what's your experience? I know cuz like there's other people that are completely on the other side, you know, new to engineering, so they're just say YOLO. Um, I feel like the results are kind of mixed, right?

Yeah. Right. Yeah. Very quickly. So when it comes to like research, I do this when I do something that I've never done before, right? So the other day I wanted to learn something about like a special type of database. I've already forgotten the name of it, but it was like a database type that I've never heard about. And so I went to Grock. So what you do with 03, I do with Grock because Grock is very tech-heavy. And I, I find that Grock delivers extremely good results because it is very hungry for, for search results, right? So it searches X, obviously, which no other LLM can do as well, I think. And also it has, I mean it has live X access, right? Many people forget that, like, like Grock has live access to all, all X threads that are currently being tweeted, right? Which is, which is insane. Insane, insane. So that might help, even if it just helps it, like, steer it into some direction. But, um, then when it comes to actually planning out the features too much. So I use, I use it mostly to, like, make up my mind and thinking about, okay, what type of framework. But then two things happen. So the first one is that I noticed that saying the more I guide the framework the better the result isn't always true in my own experience because sometimes I feel like when, when you give a model too many constraints then it, it almost feels like it's almost human in that, in that way that it's almost stopped thinking on its own, right? And that happened sometimes where then when you would have, if you would have just said like, yeah, do this and just remember to also do this, this, this, and this, the results would have been much better, right? And that may have been because my constraints were [__] or because, yeah, I don't know, it requires more testing, yes. More importantly, a guy, I don't know his name or something, I'm sure you know, like, from, from X, he's also some AI YouTuber, like mostly on X, and the guy said something that, what he, he said in a different context, but he said that what he likes about the $200 Claude plan, he's, he's as much of a Claude shell as I am a Cursor shell, by the way. He's like, "The guy loves Claude so much, you know." Oh, I think it's Ian New, I think. Yeah. And you, yeah. Yeah. Yeah. Yeah. Him. Him, him. Yeah. The guy's cool. But yeah, very, very Claude heavy. And he said the thing that he loves about, about the $200 plan, and I think he's right, is that he has the capacity to one time just tell the model roughly what to do, learn from the result, the problems, what it did good. Probably it did something that he didn't even consider, and then scrap it and do it again with a bit more constraints that fit the model. Ooh. And I like that approach. Ooh. Ooh. Ooh. That's a really good nugget to drop on everyone. So, basically, sketch out the rough draft. You're pulling the intelligence directly from the model without steering it too much. So, that way it can activate more of its parameters and stuff. You see what it outputs and just say, "Okay, I like that. Do more of that, less of this." And then try to see what it comes up with. Oh, that's such a great, great tip. Um, I love it. But see when you do something like this, sorry to interrupt you, but when you do something like this, then the Claude max subscription suddenly becomes a little bit more powerful because you have just so much more, so much more like raw throughput than with, than with Ultra, you know. With Ultra, you, you probably have to think a little bit and be like, yeah, should I really run this now, you know? It's like you don't have to do that with Claude Code, you know. I see, I see. That's, that's a really, really good stack. Um, you have been dropping so many gems on this channel and it's been such an amazing—I mean, we're just kind of scratching the iceberg here. We didn't even do any code, anything like that, and you have a really awesome channel. For those who don't know, um, I think your channel name is, it's by Robin there. So definitely go ahead and, and give him a follow cuz he's got some amazing YouTube content and some stuff coming up in the books. I'm actually trying to encourage, uh, Rob to do some more live streams. So, if you all want him to do some more live streams, definitely drop it in the chat. Uh, and then drop it in the comments. So, like, bro, you got to live stream because his, his just his delivery style is like very similar to me, but it's got a whole different take on things. And, uh, obviously very Cursor pill, getting me on the Cursor pill topic as well. Even though I've met them in person and stuff, I still give them a hard time on like things that I really want, you know, feature-wise. Um, and just because I know them and, uh, I've become friends with some of those folks doesn't mean that like I, I want to show their products or anything like that. Um, this is all of our own takes, by the way. This is our own opinions. Uh, it's going to change all the time. I think if you look at this video a couple months from now, we're going to be like, "Yeah, GPT5 all of a sudden made Cursor this new AGI machine that we didn't know was there all the time, you know?" So, it's like, that's what happened. That's actually what happened to me a year ago. Claude, Sonic 3.5 dropped and then it being in Cursor made everything so much easier because I didn't have to copy and paste to a web interface and that's kind of what kicked off this whole vibe coding for me and so, um, yeah, with that, definitely go ahead and follow, uh, it's by Robin on the YouTube handle there. Um, you can also, I think on X with the same handle as well, and then, um, yeah, uh, let us know where they can also find you. So, if you're someone who's like, I'll let you kind of do your pitch there as far as, um, you know, Robin has 20 years of experience and I kind of let you take it off from here and how they can find you.

Sure. I mean, thank you very much. You're way too kind, you know, but yeah, if you, I, I deal mostly with non-technical people, uh, but also engineers. There's obviously a lot of overlap. And if you want to find me, as Ry already said, you can find me on YouTube and on Twitter, the same handles. I have a newsletter that I write almost weekly newsletters to with the latest nuggets in AI and, uh, I also have an AI coding blueprint which is a program that targets non-technical people how to build with AI that you can check out and that's, uh, pretty much it, I think. Enough about me, you know. Yeah, I know you don't like to sell your stuff, but whenever there's someone in the community, um, who's dropping some good gems and, and doing on a similar plane. Uh, it's really, really cool and he's got a lot of folks that are currently following him, you know, 20k plus subscribers on YouTube and, um, like, you know, like almost, I say, yeah. So like that's kind of what we got going on there. Ladies and gentlemen, this is part of my summer of streaming series. So if you're interested, definitely subscribe and then on YouTube there's actually a notification you have to enable to see when I go live. Uh, generally I'm trying to go live right around 9:00 a.m. Pacific Standard Time. Uh, this one was a little bit earlier because Robin is literally on the other side of the world and so we wanted to make sure that, as you can see, he's dark and I'm kind of light here in California. We're have like the, the contrast, right? So California to, you know, I think all the way over in Dubai. Um, what a pleasure to have you on. I'm so excited. Uh, and we're definitely going to have some future sessions. Uh, and then also drop the comments. I'm trying to get Rob live so we can do some live coding. Maybe we'll have some coding challenges or something like that that we can kind of work on or maybe work on an app or something, uh, live that's really kind of lightweight and fun that we can do or contribute something to open source which would be really fun so that we can kind of vibe together and jam. So yeah, thank you so much for, for joining the show and I'll see y'all guys, uh, tomorrow 9:00 a.m. Pacific Standard Time. Thank you very much. Awesome. See you guys.