Transcription
In the month of May, I did 318 commits with cloud code. Now, I did have some with some of the other ones, like I used a little bit of codeex, root code, and augment code, but the vast majority of my commits and the vast majority of the number of lines of code that I produced or refactored came from cloud code. It is phenomenal.
When I first started using cloud code back when it came out, I actually liked it back then, too. It was early, but it had cost way too much money for me to be able to justify using this on a daily basis. Now, Claude Code, it's a CLI-based coding assistant that is really locked into the Anthropic ecosystems. That's important to note. I'm a little bit nervous about being locked into ecosystems, which is why I still really use and really love Root Code. And I think open source really is where I wish things like this would actually be. But I think there's something to argue with how good Cloud Code actually works and the fact that they can give you this monthly fee to allow you basically unlimited unlimited use.
So for OS support, this is what's a big problem for me and still is today. So they out of the gate support Linux and Mac. Now the reason that this is a problem for me is I am primarily a Windows user, and I know a lot of you would be like, "Why? Why are you using Windows?" Part of the problem with that is when I started programming years ago, I started in the game industry, and everything you do in the game industry was Windows-based; that's just become kind of my native operating system, and on top of that, I actually do game dev, so there is a dual purpose there. So I do have a Mac. I do dual-boot Linux. I have a Linux computer to my right, but I prefer working on a Windows computer. I know that is not normal or popular, but that is what I prefer.
So, what I end up having to run for Windows in particular is I actually have to run WSL. Now, WSL is actually gotten a lot better, and I'm actually using Ubuntu with WSL. I know some people use Debian or some other flavors of it, but Ubuntu has been really good to me, and I can do 80 to 90% of everything that I need to with WSL, and I have enough RAM and system resources to where that doesn't actually matter. So, I'm just calling that out. If you are a Windows user, you are going to probably have issues unless you are willing to set up WSL. And I know some people are resistant to that. So, that's just worth noting.
Now, on the pricing side, as I was putting together this video, I actually saw that Catwoo posted this: You can now actually use Claude Code a little bit on the pro plan. So, what does that actually mean? We've got the pro plan at $20 a month. Now, this plan is what I was on for the longest time, and this is typically for users that are just working in the web UI, but now you can actually come in and at least try out Cloud Code. I'd recommend doing this. This is a massive, massive thing for them. What kind of limit you get? I bet they're going to be very tight, to be honest with you. So, I wouldn't expect to be able to do a lot, but you can at least see how it works for you. Then they have a $100 a month plan, and this is the Max 5X plan. I've actually just switched to this for this upcoming month because I really do want to be able to use some of my other coding assistants a bit more and not feel like I'm paying $200 a month and I have to use Cloud Code. I do love it though, and if I wasn't in the position that I'm at where I want to test a bunch of different coding tools and things like that, the $200 max—I think it's what, a 20x or 10x? I can't remember exactly, but it's some X plan here—I never hit a single limit on it, and I literally was running multiple instances of cloud code.
So, just something to note here. These plans are phenomenally priced because when I was early on using Cloud Code, I would burn $20-$30 in a day. And that wasn't even me like aggressively using it, just using the API. So, just noting that that it's probably worthwhile considering jumping up to the $100 a month plan if you find yourself liking Claude Code, but try it out on the $20 a month plan because this is an exciting, exciting update for the people that haven't had a chance to test it out yet.
All right, so jumping over to the next topic here that I want to talk about. And this is one that kind of threw me off a bit when I first getting going, but I've come to really appreciate it. And that's really that prompting matters a ton. So you can actually guide Claude to ideate with you. You can guide Claude to think very hard—Claude code to think very hard—about something or work very hard about it or make a checklist, and it's just all about how you prompt it.
So a few things that really changed it for me, and a lot of you probably know this already, but I literally will type in, "Ultra think about this thing," whatever it is, "and implement the following." You can also do things like, "Architect, plan, and implement the following." Things like that make a huge difference in like getting it to do what you want it to do. And you don't have to use "ultra think" because sometimes you may just want something basic to be done. So you just say, you know, "do this thing." You can also say, "Don't write any code. I want to ideate with you." And this is important because, uh, what you want to do in certain cases is think through a plan and kind of figure things out. I typically use augment code for this, just to be totally honest. But I have found myself doing this more and more with cloud code because I find that that is just—it's actually pretty good at it at times.
The other thing that's really interesting is ripple. And so basically what you can do is you can do something like, um, "Tell me about my codebase," and it will create like a read-evaluate-print loop—is what ripple stands for—and it does a really, really good job answering kind of context-specific questions. Try it out, especially if you're using the pro plan. I was very impressed with its ability to pick up context of your codebase. And it gets better if you run the init command and create a claude.md file. And this is something that you can keep up to date. But basically what Claude will do is go through your codebase and build itself sort of like information—kind of like a map of your codebase—that it can then use. What's also interesting about this is it will search for other AI tools and see how you prompted those. So it'll say like, "Do you have GitHub Copilot instructor?" because it found mine when I was doing it, and it kind of built—use some of that stuff. It searches for like cursor rules and so on. So it knows about the other files that it's looking for there when it when it's kind of building out that rule set. So I would—I would learn to talk to it in the way that works for you. I started out thinking that I needed to be like very specific. So I was like, you know, you basically you can use "@" to go like chat.ts.view, and then you can tab into to kind of complete that so that it picks the reference of that file. And I would say, "Change this specific thing," and it's really, really good when you do this. And if you got a high enough level or a strong enough level understanding of your codebase, this is the quickest way to get things done in your code. But you can also be very vague. And this is the thing that surprised me the most because it's ripple loop does pretty dang good at figuring out what it needs to do. So I can say something like, "In, you know, in my orchestrator mode, I want to add a new tool blah blah blah." So basically I'm not giving it files. I'm just saying, "Go do it," and it'll, you know, 80% of the time it'll find what it needs to and make some good, good changes there.
The other thing that I would say is it's important on the prompting side to not be afraid to, uh, to kind of start over because sometimes you prompt it and you realize when you reread it that you were too vague or actually it answered the question, but you had meant something different because a lot of times I get busy and I'm like, "Oh no, right?" And I look back over there, I'm like, "Oh god, I would have had no idea; I'm surprised it actually got as far as it did." So don't be afraid to start over and just revert and do it again. Uh, because a lot of times I find that people feel like they have to keep grinding in the same direction, but don't do that. Like get rid of it and and start over.
So now what I want to do is I want to go through some tips that I've learned that have actually really helped me a lot with using cloud code. The first thing is if I want to do something like I kick off a task, let me do that right now. So in this one, I'm actually going to ask it to ultra think through refactoring a file. I can also type in stuff while it's syncing. "Please make sure all functionality remains intact. Also make sure that uh you break out components when possible." So you can send it messages while it's thinking. This is something that's very—that's kind of an awesome feature that I didn't know about, um, until a few days after they released it. It's a fairly new feature, but it's worthwhile knowing that this exists because it will allow you to think about, "Oh, I should have told it this," and now you can actually send it, and it'll do a good job kind of pushing that in.
The other thing that I would say is don't immediately just go to "yes" and don't ask again in the session. Make sure you're comfortable with the direction that it's going before you kick that off. Um, and then use this "no" and tell Claude what to do differently a lot. This has saved me so many times when it is unclear or going a way that I don't like. I'll say something like, "No, I'd rather you focus on a different part." You know, this—this is all too vague right now, but you get what I'm saying here. It's like you can actually redirect it if it's going—going astray there. Then I just want to be like, "All right, I'm going to clear this. Let's get—get rid of that. We're starting over with a new context here." So you can also use "@" to get a list of files. So I can see if you hit "@" here and I start typing chat.ts.view, I can hit enter, and then the path will be, uh, be filled in there. So in this—in this query here, I'm telling it in a particular file, I'm telling it the problem. In this case, I don't have an idea what the problem—like how to actually fix it. So I'm asking it to give me ideas on how to fix it. And because I've directed it into that file, it's going to be a lot better at helping track down that problem there. So that's a something that I highly, highly recommend, and it's really, really helpful if you know specifically the area of the code that you want to work in. But you don't have to be that specific on it. You can be very vague. But you should also be kind of specific too, if possible. So you can be—you can target functions, you can target files, you use things that are greppable or searchable. It's really good at that. Very similar to the way Codeex works.
And so in this case here, uh, it's adding some debugging logs because it doesn't really know. It's trying to get some information on it. I think this is fine. I'd probably accept this and kind of continue on there. But if I—if I want to be super vague on it. So here I'm going to ask it, "Can you walk me through writing code about how to add a new tool like multi-image?" But this time I want it to be search and replace for brief chat. Now what this is going to do, I didn't give it a file to start in. I did give it some things that are greppable, but I wouldn't even have to do that. It usually will do a pretty good job finding it. And I've actually tested this prompt a few times, and this one actually will return really, really good results.
So, if I take this same prompt here, and I'm going to open up augment code over here on the right, and you can see some of the stuff that I do in there. I'm going to put it in chat mode, and I'm just going to say, uh, basically the same question, and we're going to take a look at kind of what it comes up with. So this one—it said to create a new search/replace tool handler, which is awesome, and yep, that's exactly what it's talking about here. It's a very, very similar—similar response. It's very, very snappy for each.
So for this longest time, I actually just—kind of the first couple of weeks I used cloud code, I kind of ignored its ability to pick up context of my codebase. Don't do that. Like if you're not paying for augment code, cloud code does a really, really good job, uh, getting context of your codebase, and whatever they're doing under the hood here is incredible. So that—that's kind of just some of the tips that I picked up kind of running through this.
Now on the command side, there's only a few that I think are like very important. The "emit" one we talked about. I'd recommend trying this. It's impressive—the what it puts together, and the cloud.md file does help. Uh, in my experience on an existing codebase, it does help. Now, what this will do is this will actually generate a cloud.md file. And I can show you some examples of that here in a minute. I'll actually pull over to VS Code and just run this in my one of my repos. I've cleared it out already. I'll show you what that looks like.
Now, the other command is "clear." This is one that's very important because there's sometimes you're just done with the—you're done with the context that you're in. So, like wipe it out and start a new context. Don't keep it going forever. Don't be like one of the people that keep a chat window going forever and don't realize that you should just clear your context. So clear it when you're done. Uh, there's also "models." So you can actually switch—switch what models are active. This can be things like Opus or Sonnet 4. Sonnet 4 is what I run on a daily basis. I very rarely switch to Opus because that eats into my limits, uh, very, very highly. Uh, the other thing that they have is they have a "compact" command. This is one where it will—it will auto-run for you. You don't really need to worry about this a lot of times. So you can just let this kind of do its thing, right? If you are running via the API, there's like—there's a "cost" command and there's a "status" command, and then all the normal ones like "log in" and "log out." These are the ones if you're just getting started that you should just get comfortable with—specifically "init," "clear," checking your or changing your model, compacting. Beyond that, everything else becomes kind of advanced and awesome, right? Like things that you can run cloud code via a deploy pipeline. So you can—or you could have it set up so that it runs on—where with no input. So it just auto-runs for you. So there's a lot of like really advanced stuff you can do which I've been playing around with, and I think it's important just to keep that all that in mind.
Now orchestrating tasks. This is something that I actually really am excited about because we have a CLI tool. If you think about it, it becomes very easy to specifically take and send out byte-size jobs. And so in root code, what I actually did is I created a, uh, a cloud code mode that literally would break out tasks and it would send CLI commands to cloud. And this is great, and it works great some of the time. I honestly found it more kind of a pain than anything trying to—it go between root code and, uh, cloud code. But there are tools starting—starting to pop up where you can manage multiple cloud code instances. One of the problems I run into a lot of times—a lot of them don't support WSL very well. So again, when I have to go over to Linux to try them out, but I do locally—the way I do a lot of orchestrating of tasks, I do that with Git worktrees. And Git worktrees is something that's actually really, really, really powerful. Or I mean, simply you can actually just, uh, run multiple instances. I've never ran into an issue having two or three instances running at a time, and I typically run them in the terminal window of my VS Code. So it's just all right there. I know some people might prefer to run that outside of VS Code, but I—I'm a traditional IDE guy. I like it being in VS Code. It just—it's so nice having it integrated in there. Very well.
On the Git worktree side, what it typically does is you can actually set up sort of like feature branches in sort of this tree structure. So there's these things like "git worktree add," right? So you can do "git worktree," you can list them, you can do "git worktree," I think it's "remove." Um, there's some other commands too. I think "prune" is one. I'd have to go through and and look at all of them. So what you do here is you go "git worktree add." Here's your path to where you want to add it to have the—the repository sort of sync to. And then this would actually be your branch. And now you can just "cd" into that path and run cloud. And I can run that and have multiple instances working on different feature branches very, very easily, and then just switch over to them, run them. I still run into the same issues that I have locally with my other type of multi-tenant systems, which is I have a—I have a server that needs to be booted up, and this is kind of, uh, probably kind of whining more than anything, but basically I have to stop and start the server when I switch between those instances. So, it's not the most ideal thing, uh, where ideally I had it that could be a little bit quicker for me to actually be able to go in and test it. It's again—it's just me kind of whining about it more than anything.
So on the impression side, I just want to say it's very positive. I have very few negative things that happen. What I will say on the negative side is there are times that you end up with situations where, um, spacing or merging or weird states you get into it where you can't get Claude code to fix it. You're better off just nuking it or going in and manually fixing it. Trust me, I've spent enough time with cloud code now to basically be able to pick up instantly when it's in a position where it's not going to get out of it. Sometimes it's—it's just simple as—like I know this is going to sound dumb, but tabs versus spaces, it will, uh, get into a state where it just can't merge things because it's trying to do one or the other, and it just is really weird. It's very rare that these things happen, but again, it's just—you'll—you'll pick it up as you use it—that if as long as you're committing and you're keeping the progress of things you're working on well enough, you're not going to feel bad reverting whatever change it's got or go in and help it sometimes. Uh, but on the positive side, it is just fun. Like I'm not kidding you about how much fun it is to use. I—it's everything about it just exudes fun. And—and I know it's—it's coding, and a lot of people would be like, "Why is that fun?" But it really is. I—I feel like it's so good at what it does, and there's so much nuance to how you kind of guide it that you feel like you have control of it. Whereas some of the other AI agents, you have some control, but not to—
The point is that Claw code is. I don't need to change modes. I basically just guide it based on how I prompt it to behave, from this ultra-hardcore, like really deep thinker, taskless creator, all the way down to "I just want to talk to you about an idea that I've got," and there's all the variance between there that just makes you feel like you have control over this agent, unlike other agents. You may find that you can just reward a problem that other agents have trouble with, and you word it in such a way with Claw code, and it'll freaking nail it.
The other thing that I want to say from the impression side is that it's fast; like it, it is never, it never feels to me like it's a slog. So, a lot of times, you know, with some of the other agents, especially GitHub Copilot, one of my big complaints about that is how slow it is. It, it literally takes like triple the time of some of these other agents. But Claw code is fast, but you can make it faster and you can make it slower by how you prompt it, which is also what I love about it, which goes back to sort of the control that you have over it. Um, until you use it, you really would have a hard time understanding what I'm saying about that. But it is, it is a real thing that I feel.
I'm jumping over into VS Code because I want to just show you what this looks like in practice here. So, the first thing that I'm going to do is you can see here that I've been working with Claw code quite a bit today. Now, I have nuked my particular um claw.md file. And as you can see here, there's actually an update ready. So, I'm going to go ahead and restart Claude. That was weird. I actually got a weird error here at first, which I've never got before. So, that was kind of interesting. Regardless, I am in Claude.
The first thing that you see when you come in is this "What's new." This is such a, such a nice little touch. I know it's not much. Um, but you've got this: "Cloud code can now be used with Cloud Pro subscriptions." You also have this upgrade, which is pretty dang sweet. That is a cool feature that they've added, uh, where before I logged out, I had to log back in to kind of link my Max plan. And I can tell you right now, the to-do list stuff has actually been handled very well after compaction. Their compact feature is state-of-the-art. Like it's incredible.
So now let's just run a net here, and I'll kind of show you what this looks like and kind of talk through what it's doing. Now this actually may take a while, and I may have to speed up um some of this. It's looking at the entire codebase structure. It's reading some lines of code. You could see sort of the style that it adds. Like we've got this actualizing. It keeps track of the number of tokens there. It's really informative that things are happening. This usually takes a few minutes. So, what I'm going to do here is I'm just going to uh periodically talk about what's happening, but mostly skip over a lot of this.
So, now we're at the point where it has done the analysis of my codebase. And I'm going to go ahead and say yes, you can create that file. You can kind of see what it asks you. The second option is "yes and don't ask me again during this session." That's also a really nice thing because there are some commands when I'm working with it that I always want to approve. There are other commands that I never want to auto-run. So I like having that, that control there. And you can see here that it did 15 tool users, 61,300 tokens, 1 minute 36 seconds, and it completed the claw.md file.
And just to show you what that looks like here and how it actually comes together, uh, this is what it does. It kind of builds its own guide that it uses to help better kind of understand the structure of my codebase. It even picked up my old folder that I'm trying to phase out as I've kind of migrated up into a new version of uh, Vue. It's got all the technology that is used. We've got all the domain models. Does a really good job with local development. Like all of this is 100% correct. All of this is correct. So, it's, it's really, it's really useful. And I found that this kind of, just having this kind of does speed up the development efforts a bit. Um, in particular for me, we care a lot about our design stuff. So having this in here is actually really nice as well because it doesn't need to go look in other code to try to figure out that design. You can see our integrations that it's using, so on. Right?
So now I am in the Rue code repo. This is a public codebase that you can pull down from GitHub, and I basically asked it to diagram out how the orchestrator subtask version system works. So you can see here that it actually generates some mermaid diagram sequences. We need to load those into a mermaid renderer to kind of see how that actually works. But let's not focus on that right now. Let me focus on the flow that this actually went through here. So the first thing that I ended up doing here is I basically asked it to tell me about this codebase. So this is one of those generic queries that I was talking about. It read the readme, it read the package.json, and it got enough information from that to give me a definition of what the codebase is. It even picked up the, the main components on it. This was relatively fast. Let me see if it actually tells me the time there. I don't think it does when we're this far in.
Now, basically what I, I did is I ran a nit on it. So, let me show you what the uh claw.md file looks like here. And you can see that it actually did 26 tool users, 90.6k tokens, and ran for two and a half minutes to generate that. And I know a good amount about the Ruru code codebase, enough to say that the majority of this I can confirm is correct. But I'm not as familiar with every part of it. So I don't have the same ability that I do with my own codebase to be able to say, "Yeah, that's 100% all correct." But it did generate 149 lines of information here based on, you know, 90,000 tokens that it ran through. And you can kind of see some of the summarization there. Then I basically ask it about diagramming it out. I actually told it to ultra-think through this. And you can see very particular, like you can feel the difference that it needs to go through. Like look, it triggers this thinking portion in here. So you can start seeing "thinking" where if you go up through these because I didn't tell it to ultra-think, you're not actually getting to the thinking as part of the output here. So you could see that just carry through through all of this. Now it does slow down the process, but sometimes you really do want it to take its time and generate, you know, a more thoughtful response. 4 minutes and uh 9 seconds. So there we go. 32 tool uses, 88,000 tokens, 4 minutes and 9 seconds.
So now that I've had that, got sort of this high-level understanding of the orchestrator subtask system, what I would typically do now is this is a, this is a feature I've been wanting in root code for a while where I want to be able to represent in root, in root code basically the orchestrator and all the subtasks, how much each of those use, how much they cost, how many tokens they use, and be able to visualize that and ideally click through them. I don't actually have an idea for, I haven't sat down and figured out the design on that. But basically, I could go and kick this off now and say, "Hey, now that you have that, go figure out a way to do that and get some ideas on it." Most likely, it's not going to be something that um we're going to, to actually be able to keep. But that's just an idea of kind of how you can work through this.
So, here you can see that it actually created its own to-do list, and it's going to work through it one by one. So if you watch my previous video where what I talked about a lot here, let me move my video over a little bit, is basically the rankings of each agent, and Claude 4 does really good. Claude 4, sonnet, very specifically sonnet, does really good in Cloud code, and it should because it is their tool that they're building, and they understand the inner workings of it very well. And it kind of goes to say like my personal ranking is this, and I don't see this changing drastically over the next month. It has historically changed a lot. Like one time I was using Cursor 100% of the time, and then I was using uh, I used Clin for quite a while, and then I used Augment for quite a while, and I still use Augment now. I think they're phenomenal for certain things, and that it just, you know, you kind of have felt so much sort of turbulence in this space because people are leapfrogging each other. But I think Cloud Code is just this really sweet spot for me. Just, it's my workhorse. So I think Cloud Code is just a sweet spot for me. I'm going to be paying for the foreseeable future the $100 a month until that ends up being like, I, if they start nerfing it, I'm going to um probably upgrade to the $200 a month plan. But I, I really do need to try to get my spend down.
So, my current stack that I'm going to be running is going to be uh $30 a month for Augment code because I'm on the grandfathered plan. I'm going, I currently have $200 a month for uh Open AI because of Codex. There's something really special about Codex, being able to be on a bike ride, get something popped into your head, pull over, get tick something off, bring it back, and work on it in Claw Code. Now, I think there is probably a way that I could probably downgrade this now that they've given you access on I think the pro plan. So, I'm going to look into that because I don't know what the limits are on that. But if the limits are where I can run five, six tasks a day on the pro plan, there's no reason for me to be on the $200 a month plan there. Um, and then the other thing would be like API cost. Now, my API costs are kind of crazy because eval, typically they've started to creep up in, in pricing. So, you know, I'll spend, you know, $200 eval. And that's me trying to use models that are built into these systems as well. So, this can actually run $200 to $1,000 a month just on trying to run um eval.
So, this is for Cloud code. I'm going to be having my $30 Augment code. I'm going to be keeping Codex for the foreseeable future. And then this is going to be a combination of root code um and any other API costs that I have. Of course, there's just going to be miscellaneous costs with, with testing other things. So, one thing that I'm just noting here is Google has kind of fallen way off. There is always a chance there's some upcoming rumors about them launching something that may bring me back into using Gemini or uh Google models again. But it's going to, it's going to be need to be like something pretty massively good because Cloud code right now is in an incredible spot.
Now I think that's going to about wrap it up. There's so much I could talk about about Cloud code. Is there anything you want me to go deeper into on Cloud code? Let me know. I've been getting so much feedback on all the stuff that I've been doing. Trying my best to take that in and just continue doing better and better. Um, anyway, I appreciate all of you. Let me know your thoughts below. If you've had a chance to try Cloud Code, if you're paying for the $100 or $200 a month plan, if you hit limits, because I've never hit a limit on the $200 a month plan, but I've heard people hitting it on the $100 a month plan. So, I'm curious what is going to happen as I switch to that. Anyway, till next time everyone. Have a wonderful, wonderful day. Peace out.