Transcription
Let's cover 10 pro tips for working with the cursor agent. So, I've got cursor open and my agent on the right. And the first tip is to use plan mode.
So, in the agent input, I'm going to hit command N to make a new chat. And then I can hit shift tab to switch and toggle into plan mode. Now, what plan mode can do is it can actually go look at my codebase, research and understand how my app is set up, and then help me create a high-quality plan to implement a feature.
So, this is on my personal website. I'm going to say let's make a new page to show my top artists on Spotify and I'm going to hit enter. Now, I've already set up Spotify in my app and I'm using it to display the track that I'm last listened to on my homepage. So, we're going to see that it's planning the next moves here and it's going to come back to us with a few questions to help clarify what the implementation's going to look like. So, it reads my current Spotify implementation. It reads some of the code that I've already set up and it's going to help me figure out how I would actually add this feature.
So, let's do uh slash. We'll do slash music and then we're going to do let's do top 10. Yeah, top 10 works. Time range. Um, let's do all time and last three months. Two buckets. Sure. And then for displaying what we want to show, we can do um use your best judgment. I'm just going to see what the model comes up with and we can change it and tweak it from there.
So, it's starting to make the plan and we see as it's making the plan, it creates a new markdown file and it's giving some details on what needs to be created. It's calling specific components and files and referencing them in our application. And it even makes some to-dos that we can use to implement the feature. So, this seems fine. We can just read through here. It looks okay. If I wanted to make any changes, I could go in here and say, um, we don't need a brief intro about music taste. So, we can get rid of that. I'll save the plan. The to-dos seemed fine. Okay.
So, now if this looks good and we're happy with it, we can hit build. And build is going to accept this plan. And then for our given model, which I'm using Claude 4.5 Sonnet here, it's going to go and actually implement the plan. So let's let this cook.
Okay, so the model finished, but it actually didn't work. I tried to load the page and I got a permissions error. So I needed to reauthorize my Spotify token to have the additional permissions. So I just pasted in the error I saw and now it works. So if I go back to my app, I now have this list of my top artists. So thank you plan mode.
The second pro tip I want to show is accessing the context menu and specifically a feature at branch. So I'm going to go ahead and command shift P, create branch and we'll call this top artist. And then inside of our chat window, I'm going to make a new chat. And in this new chat, if you type the at symbol. So let's just do that. What you get is this menu with all these different options. And one of them is get and specifically at branch. So on this branch, what I want to do is review all of the changes and see if there are any issues. So I'll kick that off.
Now, inside of that menu, there's actually a lot of stuff. I can just pull it up here. We have files and folders. You can reference public documentation, past chats, linter errors, uh, all sorts of different things, especially tagging in specific files or folders that will go into the context. So if you know, for example, you're going to use this Spotify file, you can tag that into the context and put it in the context window for working with an agent. So I find these get commands pretty helpful. And it looks like the code seems pretty good. We're having the AI review the AI generated code, but maybe there's one issue with the cache times that are different. Nice. Okay, so overall seems pretty good. That was pretty helpful.
The third pro tip I want to show is creating custom commands. So in the cursor folder, inside of here, I'm going to make a new folder and we're going to call it commands. And inside of here, I can add in markdown files as commands. So we're going to call this one PRD. And inside of here, I can I can call this file whatever I want. I can put whatever prompt I want in here. So, we're going to use this to make PRs. So, uh, your job is to create a PR with a descriptive title. Always use the GitHub CLI. Now, we could add in specific rules or specific guidelines, commit guidelines in here as well, but um, let's also say if you haven't already made a commit, do that first. Great.
So, we have this command. Now, over here in the agent panel, I can do slash and it's going to show up as an option as a custom command I can run. So, slash PR and we even see this little hover of what the custom prompt here is. So, slash PR. Let's run that and see what happens. Now, my hope here is it will use a shell command to use the GitHub CLI to go ahead and first make a commit and then make a PR. So, it's going to check and get status here. Great. We have some, if I go over to my source control view, got some stuff on this branch, uh, and some changes that we need to commit. So, it's going to go ahead and look at the changes, try to figure out what the PR is about, and then make this change for us.
Okay, so this just finished and I see I have a link to my new pull request. Nice.
The fourth agent pro tip is that you can pass images to the agent. So we have our list of top artists, but maybe we want it to look more like the Spotify uh wrapped. So let's copy this image. We can go back to here and paste it in so we can see what the image looks like here. And we're going to say make the top artists and music page look closer to the image with uh artist photos please. Now this is not scientific but I think if you add please and thank you the models perform better. You are burning more energy in the data centers. So, you know, there's some trade-offs here, but we're going to see if please and thank you help us get a better response.
So, it's going to go and make some changes to our Spotify code to how it's fetching the data for the artist. So, we need to get the images as well. And then it's also going to update the top artist. So, this component that displays each kind of element in the list. That makes sense. So, it's kind of cooking away here. And it probably also needs to update the configuration to handle these images from a remote URL. So, we've got the Spotify images here. So, it's going to go and update that. That seems right. Next config. Uh, okay. I'll keep that. Seems good. And I think that's everything. Let's go check it out. Uh, reload the page.
Okay. So, it was very honest to the image uh where it's got these big numbers kind of like the wrapped and we do have images. Um, so, that's not bad actually. Um, yeah. Yeah, I can now tweak it from here if I want to make it look even better, but it's a good start.
The fifth agent pro tip is that you can fork or duplicate chats. So, let's say I like what I have here, but I want to try a couple different options. I can keep all of the context I have so far and I can click on this dot dot dot and go to duplicate chat and then make some changes from here. That's one option. In general, I try to recommend using the smallest number of tokens in the context as possible. So maybe you want to do that. Maybe you want to start from scratch. Either one are valid options.
The sixth pro tip is you're going to want to make sure you're keeping an eye on this context window gauge. So this gauge shows that for the 200k context window of Claude Sonnet 4.5, we've used about 12%. Now, as this conversation grows and grows and grows and grows, you also have the option to do a slash command in here, like summarize, which is going to take your current conversation and compact it down to the smallest, smallest bit. And this is kind of a pro tip 6.5, but you want to be very intentional about how you use that.
The seventh agent pro tip is that if you want visibility into your usage in settings, you can toggle usage summary from auto, which will only show up when you're close to your limits to always. And you can see it down at the bottom. I've used 38% of my usage limit. And you can also see when that resets. So, if you're a person who's trying to be very cost-conscious, you can see that here.
The eighth agent pro tip is keyboard shortcuts. So you can use command I to open up the agent window and then command slash to change the model that you've selected. So maybe I want GPT-5 high-fast or maybe I want Claude 4.5 Sonnet. I use these two shortcuts all the time, but there's a bunch of shortcuts you can configure inside of the settings in any way you want.
The ninth agent pro tip is to start new conversations frequently. So, in this past conversation around the artist's photos, I used about 12% of the context window. Now, something I see a lot of beginners do is they keep this conversation going and going and going and adding feature after feature after feature here to the point where they're running up against the limits of the context window and the quality of the model degrades over time as you add more stuff into the context. So, I generally try to recommend starting new conversations for new features. And as a reminder, you can do command N keyboard shortcut to start a new chat.
The 10th agent pro tip is that you can actually go back in time in a conversation with an agent. So we have our top artists here and maybe I decide actually I don't know if I like those numbers that big. So let's make the numbers for the rankings smaller. Sure. And we're going to go and kick off this change and we'll see what it ends up making. Now, if I decide, for example, that I don't like that change, I can just go back here and see. They're a little bit smaller. Maybe that's better. Maybe that's not. Uh, if I decide that I don't want that change, we'll just let this finish. Great. It's done. I can go back to here. So, I can click this button to go back to this checkpoint. And it says, do you want to discard all changes up to this? Yes. Let's go back here. And now we have our big numbers back. So this can be small changes or you can go very far up in a conversation and kind of roll back. Now this is very helpful and I still also recommend using and learning about Git. Git helps you save those changes uh in a more uh persistent way. So you probably want a little bit of both.
Okay, I know I only said 10, but we're going to do a lightning round with some other really fun pro tips. Did you know in settings you can have the agent play a completion sound when it's done? You can also have it show system notifications when it finishes or if there's some command that you need to give permission to run. Pretty nice. Did you also know you can ask the agent to generate visualizations for your codebase? So I asked it to generate a mermaid diagram of the entire Spotify flow. And look at this. Look how detailed this is. The user to app to the API endpoints. How the OAuth flow works. What the handoff is for access tokens and refresh tokens. Pretty nice. Makes it easy for documentation.
All right, the final final one in our bonus lightning round of agent pro tips is the new agent layout. So, you can try this out in beta today. It's a new way of working and managing agents in your codebase with all of your agents on the left sidebar, your main conversation in the middle and then the diffs on the right. This can integrate with your browser, so it can read network logs, it can look at the DOM, you can run terminals from here. Everything that you love about the cursor editor, now redesigned for working with agents. We'd love your feedback if you try this out and stay tuned for the next video.