YouTube Deep Summary

I have now for the last several weeks switched over from cursors agent to claude code and I'm not looking back at all. Here's how I use cloud code and my best tips. First, I install the extension which works with VS Code, cursor, and probably other forks like Windsurf. It doesn't do a lot, but it makes it really easy to launch Cloud Code right in your IDE. I still use Cursor by default because every once in a while it's nice to use command K and their tab completions, but the only time I've touched the agent sidebar is when Cloud was down. What the extension does do is it makes it really easy to open cloud code. I often have it in multiple panes at a time. So I can run things in parallel as long as they're working on different parts of the codebase and different files. And if I have a file open, it'll automatically pull that into the context. Now Claude uses a terminal UI which I was very hesitant about at first, but they actually do a really good job with it. You can tag files easily and choose what you want to include. They have slash commands which are awesome. Speaking of, I use the model command a lot and usually work with opus unless opus is having issues which happens and then switch to sonnet. A lot of people should probably just use the defaults. It'll use opus until you're at 50% of your usage limits and then switch to sonnet which is more costefficient. I found opus isn't slow like 3.5 used to be compared to sonnet at least not noticeably. And both models are very good but opus is just a little bit better. Other commands I use a lot I use clear a lot in my opinion. every time you're doing something new, clear. You don't need all that chat history in your tokens every time. And you don't need Cloud always trying to compact it either because compaction basically runs another LLM call to output a bunch of tokens, which takes time to summarize the conversation history. Just clear clear every time you're doing something new. The up arrow key will go back to past chats, including chats from prior sessions. So if you close out of Claude and open it again, for instance, another day, you can still go back to prior sessions. Speaking of opening Claude, one thing it does that's really annoying is after you type a prompt, it'll start working. Agents take a while. So, I'll go about my business. I'll check Slack. I'll check email. I might code something manually. But then here's the problem. I come back and I see it's asking me, can I edit this file. It's really annoying. Yes, you can edit files. It's the point of being an agent. Like, edit the files. And there's no way I found to globally say just edit files. It's fine. And then you'll go about your business and come back and it's asking if it can run a basic bash command. Can I run lint. Yes. Oh my god. Yes. So, here's what I actually do. Every time I open claude code, I actually quickly hit command C and then I run claude dangerously skip permissions and enter. It's not necessarily as dangerous as it sounds. It's akin to what cursor used to call yolo mode. And while it runs a minor risk that a rogue agent could run a command you didn't expect that's destructive, I've never seen that happen in my life. So, up to you if you want to take the risk. I have for weeks and weeks and I've never run into a problem whatsoever. Now, speaking of slash commands, Claude has a lot. One really cool one is installing the GitHub app. This makes it so when you submit a PR, Claude will automatically do a code review. This is pretty awesome because as you use more AI tools, your volume of pull requests might increase. And I found in certain cases the AI models are better at finding bugs than humans because they frankly put more effort into it in some ways. While I've seen humans are really common to nitpick at, oh, this could be named differently and stuff like that, I've seen Claude actually find real bugs that our humans missed in a good chunk of cases. The main tip I have for this is Claude will add a claude code review.yaml. It'll have a prompt in it already. Here's the prompt I use. The original issue we found with this tool is it was really verbose. It would comment on all kinds of like nuanced, unimportant things and write a whole essay on every PR. What we really care about most for the AI to review is bugs and potential vulnerabilities. So we tell it look for bugs and security issues only report on bugs and potential vulnerabilities and be concise. The cool part is when you run this command and edit that one line, you have a pretty awesome new addition to your workflow. There's a lot of other really cool stuff it can do like pull comments from a get a pull request and address them, review a pull request and do things like set up your terminal because out of the box shift enter will not work for adding new lines. But we can just hit enter and tell it to do it for us. And there we go. Shift enter adds new lines. Beautiful. Speaking of quirks with using a terminal interface with cloud code, you might be surprised, but you actually can drag files in. Though in tools like cursor and probably VS Code, it opens it in a new tab. If you drag in and hold shift, it'll actually pop it in and reference it like you need. Now, one thing that doesn't work is it doesn't paste images from your clipboard. So, if I do the Mac thing where I take a screenshot of this and command V, nothing happens. A special trick for this is control +v actually will work there. It pasted the image. That one took me a long time to figure out. Another thing that took me way too long to figure out because this is not a normal UI interface is when cloud is working. I always thought to hit C to tell it to stop. That doesn't do it. Hitting control C twice just exit entirely. Oops. To actually stop Cloud, you need to escape. And if you want to jump to any previous message, you can hit escape twice and see a list of previous messages and pop back to them. There's a lot of invisible features like this in Cloud Code. If you want to be hardcore, Cloud also has a Vim mode, but I'm not a Vim user, so I do not use it. Now, let's talk a little bit more about why Cloud Code is so good. In builder, we have one React component that is so large, I can barely even scroll to the bottom. It is 18,000 lines of code. There's never been an AI agent that can reliably update this file and tell Cloud Code. When using Cursor, I have still found a lot of little hiccups. It has trouble resolving patches and has to rewrite files often. It really struggles to update extremely large files. Cloud Code has no issue updating this file, like not even remotely. Cloud Code works great with large code bases and complex tasks. And I find it just getting stuck incredibly rarely. I'm not even sure if I've noticed that at all. Whereas with cursor, I feel like I have to babysit it more and when it gets stuck, stop it and realize maybe this was just not a good task to ask. Cloud is also exceptionally good at navigating a large codebase, searching for patterns, understanding relationships between different parts of the code, components, shared state, stuff like that. It's honestly kind of incredible. If you think about it, cursor built a product as a general purpose agent intended to support a variety of models. In order to do this, you have a whole additional team involved. They trained custom models as well and there's just more layers and requirements and things going on here as well as cursor doesn't build or control the core models that do the core AI. Compare that to anthropic. They definitively make the best coding models and so they make cloud code the best at using the model. And when they hit challenges with cloud code, they go and make the model better. They only care about supporting their model. They know everything about how the model works, how it's trained, and how to use it in depth. And they continue to train the model to work well with what they need for cloud code. It also means that Anthropic can give you the most possible value for the least possible price because you only have to worry about paying them. So they could compete on giving you maximum access to models like Opus without situations like cursor has where cursor has to make money too. So cursor needs to make a margin and anthropic which fundamentally means you need to pay some degree more with cursor or they need to lose more money which isn't sustainable. So, I commend the Anthropic team for making such a good tool with cloud code because they can give you the maximum performance at the lowest price directly from the experts and make their models and products better altogether. It's really smart and it's where I am betting my time and money right now today given the incredible results I've been seeing. Now, speaking of pricing, I pay for the max mode. Now, if you're used to cloud code previously being based on API pricing, cloud code now supports the standard pricing plans. Max mode, in my opinion, is an absolute steal. If you feel like a shockingly intelligent coder working for you 247 is not worth $100 a month, you need to look hard at what a human costs per hour for engineering, regardless of where you look in the world, and it's orders of magnitude more than that. So, if you're having trouble justifying it for yourself or from your work, always remember the gains this has given you compared to actual manual human work. One other feature of Cloud Code that I absolutely swear by is queuing. So we could type a message say add more comments. This is a thread about adding comments to some code. And often I think about oh the next thing I want to do. What I used to do is create a notepad and start drafting other prompts that I want to do. And then when I see one is done I'll go and paste the next one and enter. That's what I did with cursor which is really annoying because again I'll usually go about my day and answer Slack messages, answer email, do something else and come back and see the agents been idle for who knows how long. Type the next prompt go away. It was not very time efficient. But now with Claude, you can just cue more messages. So if I think of another one, actually also add comments to the main chunks of the JSX. I can cue them and to the computed values. And what's great is Cloud is really smart about knowing when it should actually run those things. If it needs feedback from you, it's not going to automatically run the cued messages. It's a pretty smart system, but when it's wrapped up something, it will start addressing them when it makes sense. So you could queue up a lot, go about your day, and in a lot of cases just come back to a ton of work done in a good and smart way. But again, check it from time to time because it might need your input. A couple other cool power features you could do with Claude is add custom hooks and custom slash commands. And the coolest part is you can have Claude build those for you. In this case, I asked for Claude to add a couple default hooks, commands, and settings and created a settings file that I can easily edit. It added a Claude MD which gives a bit of project overview and some key commands that it should know about. This prevents it from having to figure that out each time and scan the codebase for is there a build command or a lint command. It always has awareness of that. And it added some hooks for what code should run before edits are accepted, such as run prettier on a specific file or after edits, like run a type check on a specific file to make sure that it only accepts good and correct files. Another really cool one you could do is add custom slash commands. Like if I want to output a test, I can describe my tests. To add commands, just add acloud/comands folder. add its name and then MD. You just write these in natural language and you can use this arguments string to place an argument into the prompt and the cloud will do exactly what you asked just by typing slash and whatever it is. You can even have subfolders and those we can access like this like builder colon plugin matches the builder folder plugin.mmd and that's how we can create a new builder plugin super easily. Another cool one is we can use the pound sign and add memory super fast like always use mui components for new stuff and then it'll automatically save that to the most relevant file. Cloud.md files can be hierarchical. So you can have one project level and you can have one in nested directories. It'll look at them all and prioritize the most specific the most nested when relevant. You could also save this to global user memory preferences you want to apply everywhere or local project memory that's memory specific to you and gets get ignored. add it to any of these files and then it'll write it for you. Now, the main problem I sometimes feel and I see people point out the most is sometimes you really just want a normal UI. It is kind of annoying that typing out long responses have these weird escape keys. I can't just click on what I want to edit and highlight and edit. It's still very clear you're working with a terminal and that comes with some inherent limitations. For those times where you want a UI or you prefer UI interface, but you still want the mechanics of cloud code, one trick is to use the builder.io VS Code cursor windsurf extension. At any point you can just launch into a visual UI from your IDE sidebar. What you get is a much more visual interface with a typical chat dynamics that you're used to and a live preview. And under the hood, we built the system to match almost exactly how cloud code works. This tool uses effectively the same agent and approach as cloud code down to the tiniest details. We reverse engineered it as closely as humanly possible. And then you can preview your changes quickly. Like I actually just moved items from here to right here. Perfect. And what's kind of cool is at any time we can switch to design mode and we get a Figma style interface where we can change up any of the styling we want and edit it just like a design tool. It's kind of a cool way to bridge the worlds of visual editing and code and get a bit of the best of them both at the same time. We can explore different UI patterns quickly and make precision edits and then just apply our changes to the code. You can do all this stuff from a browser interface, too, which can be really cool for letting people on your team create prototypes right with your design systems and applications. And then when you're happy with your changes and you're ready, you can just fire off a pull request. Here's our PR with a title and description. We can look what files changed. Looks like it's using our design system correctly as hoped. And I can leave a comment anywhere and say, "Hey, build a bot. Move this to be in its own component in a new file." In this world now, we're effectively just communicating with Claude code, but through the PR directly. The agent will then reply and push up commits addressing the feedback. And here we go. The builder bot address the feedback and describe the changes. Now, it looks like we have our customers table in its own file. And we are using it as expected in the original, we can merge away. So whether you want to use cloud code in a visual way or in a terminal, I hope these tips were useful for you to be able to build awesome stuff like never before. Check out my full list of tips over on the builder.io blog. If you want to try the visual interface for cloud code, go to fusion.builder.io.