Transcription
My favorite enterprise service is Cursor. Every one of our engineers, 100%, is now assisted by AI coders, and our productivity has gone up incredibly.
If you haven't been paying attention, Cursor dropped some really interesting features in the last week, which I think are indicating a big change in how we approach software development. We're moving away from an IDE, or intelligent development environment, to something more like an agentic development environment. We've got plan mode, integrated computer use, hooks, slash commands, and more. I'm going to cover all this in the video and show you practically how I would apply it. I'm also going to build out the same app twice. One using Claude Sonnet 4.5 and the other using GPT-5. Both are great models, but I'm going to summarize at the end where I feel each one is best suited.
Also, if you want to ship your app idea in just a couple of weeks without writing the code, I highly recommend you checking out switchdimension.com. I have a course and community there where we have weekly meetups, project support, and so much more.
Okay, let's get stuck in. So, IDEs are fantastic for being able to jump in and edit the code as you need. But here's the thing. In the last 6 months to a year, I haven't been writing any code. I've been guiding agents to create the code for me. The bottleneck for me now is not how fast I can write code in my editor window here. It's how fast I can actually review the code that the agent is producing. Any pro-developers out there will tell you that the number of pull requests and the amount of code reviews they're doing has increased a lot, because I know it has with me.
So, what you might not have noticed is the advancement of Cursor's new agent window. So, basically, I can just click this button here, open agent window, and I get this whole new agentic interface here. So, let's take a look at one of my previous conversations. It's the same, pretty much as what we're used to. We enter in a prompt, we can add context, everything that you would normally through the agent window. On the left-hand side, I can see all my previous conversations and agents. So, basically, I can run through all my changes really quickly to see what's changed and if I'm happy. And they also have a brand new browser built-in, which I'm going to talk about in a sec, which I think is pretty cool. And then I can also open up my right sidebar here. I can see my commit history and I can see any of my files. Now, I know what you're thinking. Doesn't this just look like a reverse version of our normal IDE? So, it's kind of the same, isn't it? Like, to be honest, I can just drag this over here and it's all going to be pretty much identical.
So here's where I see the big shift. Let's take a look at this.
"What is the right interface for managing multiple coding agents?"
If you're just getting started coding with agents, I don't recommend immediately trying to juggle multiple agents. I mean, let's be honest, are we really being productive running nine CLIs in parallel? Probably not. Probably not yet, though. If you've watched any of my videos before, when I am working with multiple agents on different projects, I tend to have them all open in various different instances of Cursor. I have to switch between each one. That's why you see different colors on the top of each one. I want to know which one I'm working on. I use Peacock for that, by the way. Really good extension. So having this agent window, or agent development environment, or conductor, or orchestrator, whatever you want to call it, basically allows me to run and keep an eye on all my agents and separate them out.
So the vision here is not necessarily for right now, but for what's coming. So I'm going to be running multiple different agents. And what we're getting soon is the ability to have work trees so that I can have one agent working on one particular part of the project. I can be having a conversation or doing some research with another project. I can have a background agent running on something else, and maybe multiple work trees on issues I've pulled from GitHub, and they're all going to be separated. So this is basically orchestration of agents.
Now, all of this being said, if you're a beginner, you need to learn how to build one app with one agent before you can manage multiple agents working at the same time.
So another interesting feature that's been released is plan mode. So when you run this, it's going to give you a file, a little bit like this, that will appear in your Cursor plan folder. So basically, it breaks down your environment, your database structure, authing the plan. You move from plan into agent mode.
Now, out of interest, I ran this planning step with Codex, GPT-5 High, GPT-4, Gemini, and Claude. GPT-5 Codex was too concise and not enough detail. GPT-5 High took about a minute and did a pretty good job. The same with Sonnet 4.5. What was kind of disappointing was what I got from GPT-5 Pro. It took 7 minutes and cost me about $2 to generate this plan, and it wasn't in any way better than what I got from GPT-5 High or from Claude. So, I'd refrain from using it for that kind of planning.
If I was personally going to pick a model for using plan mode, I would use Claude because the questions that it asked at the start before it proceeded to build the plan were really great. It asked me about my MVP, what I wanted to add in, what I wanted to take away. And this is one of the big problems I see with spec planning. They try to build way too much too fast. And Claude was the only one that consistently asked me what do I want to include in the first part of my build. So, I thought that was pretty good.
So, spoiler alert, I built this project using GPT-5 High and also Sonnet 4.5, and Sonnet just won hands down, and I'll talk about that in a few minutes.
And so, one of the new features I said I'd talk about was the browser setup. So you can see up here, there is now a little browser context that I can add in, and Cursor can now use computer use to interact with our web app. So I can have it to test the interface. I could do something like paste in an exact design from Figma and then tell the agent to make sure that the pixel widths were the same or that the design adhered to it as much as possible once it's able to see the interface and see an image at the same time. It can do that kind of stuff.
Now, I did find this experience to be hit and miss. It's not foolproof. It doesn't work all of the time, but but like anything I've watched grow in Cursor over the last 6 months, I think the functionality here should improve. I can just add it as context. So if I type in @browser, um, I can actually prompt it to say, hey, step through the workflow. So this is something I just ran there a second ago. And you've two different options. You can run the browser in the agent window, but I found that still to be a little bit buggy, or you can run it externally via Chrome.
So, let me just open up Chrome here. So, it opened up this window by itself, and then it's got a list of to-do here, like it wants to step through the workflow and test it out. So, you can see it's gone and clicked on the button. Now, this is something we had available already through MCPS, but I love the fact that this is baked into Cursor now. So, awesome stuff.
Basically, when we're running browser mode, we can see the console logs of the browser. So, that's the client, and we can see any errors that show up with JavaScript that's being run. But we can also have our AI watch the terminal. So the terminal is going to show exactly how it's running, if there's any errors or server errors or any kind of logging that you've got in place. If you run this terminal yourself, the agent can't usually see that. So if we actually get the agent to run the server for us in a background process, Cursor usually pops it open here. We can see it running, and then the agent can actually see the running process as well, and everything that gets piped in.
If you don't like doing that and if it doesn't always work for you, I have a workaround. It's in my package.json. Whenever I run npm dev myself, it starts to run the terminal output directly into this Cursor terminal log file. And then I've got a little slash command here called terminal check, which basically goes and checks the terminal logs for me. So if I'm over here, I can just type in /term check and it'll pop up there for me. Sorry, I have two of them running there, but you see what I mean. Essentially, if we can get our agent to see the logs of our server as well as the logs of our client browser, it's going to be able to move a lot faster and solve bugs and issues without our jumping in.
So, the next new addition to Cursor has been slash commands. Now, slash commands have been in Claude Code and other CLIs, even Cursor CLI, for quite some time, but they're actually just a really useful way to invoke certain prompts. So you can just think of them as basically prompt management. So we've got different built-in ones here like reset your context. So if your context hits a certain percentage, you can reset it. You can generate automatic Cursor rules from a conversation. And you can actually go and create your own commands as well. So I showed you this one, terminal check. So all your commands are stored in Cursor commands and terminal check. And you can actually import them directly in from Claude, and it should work because these are just natural language files, and you can set them up for anything like security reviews or prompting any different kind of processes. So it's great to have those baked in.
So we also now have hooks. So what the heck is a hook? Well, if you're a Claude Code user, you'll have heard of hooks. When an agent does anything like call a tool, edit a file, start, stop. That's a point where you can trigger a hook. So basically, you can add in your own bash script files, and you add them in up here under Cursor and hooks, and we can do things like, you know, play a completion sound whenever your agent finishes. I believe Cursor does this already, but I use this in Claude Code myself. So I'm just using this as an example. So whenever the agent starts or finishes, it's going to give me this little sound file. So you can use whatever it is that you want. There's a few examples here in the Cursor docs to kind of get you started.
So, the app I decided to build was a YouTube script generator. You can just pause the screen here if you want to see my prompt. Now, this is just a dictated brain dump. There's no prompt science here. There's no prompt engineering. It's actually really quite rough, but I gave this same prompt to each one of the models to see how it performed.
So, after about 40 minutes with GPT-5 High, this is what I got. Essentially, I can create a new draft. I can see my dashboard. I can create a new draft here. Dump in my thoughts on script. Create the draft. It generates a load of different ideas here. It didn't escape the JSON. It kind of got that wrong. Again, the design is really rudimentary. When I click on keyword generation, it didn't necessarily work. When I tried to generate an outline, didn't work either, or a script didn't come into fruition. So, so this was after about 40 minutes of working with GPT-5 High. This is as far as I could get. The majority of the time, about 17 minutes, was just working away on plan mode, putting it together, but it really just didn't get anywhere.
But on the other hand, Claude 4.5 did actually really well. So essentially, this was a one-shot development from Claude. I got all of this within about 30 minutes of it working away autonomously. So I can create a new script and I can enter my thoughts here or brain dump them. I can generate a title. I can click on this and then go ahead and generate a script outline. Personally, I don't use scripts for my videos. I basically just do a brain dump over an hour or two on video and then spend way too much time editing it to give you the concise version. But I thought this might be a fun project.
Anyway, another thing I really like about Cursor is your ability to choose any model that you want, and that's really important to me. I normally switch between using GPT-5 or Claude. Mostly, it's just those two models. I do test other models all the time. There are cheaper versions of these, but they're just not quite state-of-the-art, so I stick to those models. Pricing is not a huge concern for me. I just want to use the best that's there. I feel like any amount of time saved on writing good code and bug-free code is going to save me more in the long term.
Anyway, so having worked with it now for a week or so and running a couple of tests, I feel like Anthropic is the more creative of the two, particularly when it comes to design and front-end design. I just get better interfaces with Anthropic. So, the benefit of the Claude Sonnet models is of course their creativity, but that can be a downfall as well. So, it makes them really good at reading between the lines, but it does mean that they go and they do a bit too much. And when you're trying to be precise, that's not exactly what you need. I remember when Claude 3.7 came out, it was overenthusiastic and it did way more than it needed to do. It needed to be rained in. So Claude still has a hint of that enthusiasm, which is great for greenfield projects, but gets you in trouble when you're doing focused refactors. So that's where I thought GPT-5 did better when you wanted to give it a precise spec to get something done. It was much more precise and just focused on exactly what you wanted it to do.
Another key takeaway is Anthropic is just that much faster in terms of keeping you in flow. It's just a more responsive model, but that's not always a great thing. So with OpenAI's GPT-5, it does have that ability to think and work for longer on a problem. So for that reason, I'm kind of putting Claude Sonnet in my S tier and OpenAI falling into my A tier. They both have their own uses, but if you were to give me just one, I'd probably be picking Anthropic now at the moment. But to be honest, this is really down to what your job is or what you're doing on a daily basis. I do a ton of prototyping and front-end work. So Claude really suits. If you're doing a ton of refactoring or doing much more precise work, GPT-5 might be your daily driver. So, you really just have to play with these models yourself, think for yourself, and see exactly where they fit.
So, if there's any other models or tools you'd like me to test next week, let me know. But here's my big advice. There is new CLI, there's new tools, there's new models coming out on a constant basis. We've kind of reached this plateau where they're all just getting fractionally better at the moment. So, I want you to stop focusing on all the new tools for a minute. Pick a tool like Cursor, Open Code, Claude Code, whatever you want. Ship something in the next week. I dare you. Make it small and achievable with the aim of showcasing it to somebody by the end of the week. You can follow me on X or Twitter or even in the comments section here. I want to drop in what you're going to build, and I want you to come back one week later with a demo, a URL, or something that's not a localhost as a form of accountability to just get building.
Okay, see you next week.