Transcription
Hello and welcome to this video. I'm going to be discussing three game-changing things that I've learned about Claude Code this week that have massively improved my coding. I will probably start doing these more frequently because I do learn an absolute metric ton about this thing because I'm using it constantly.
Now, the first thing, and this might seem obvious to some people, but people who are on the max plan probably haven't actually realized this. You can actually use the 1 million token model with API key usage. And effectively, if you don't do that much coding, you can spend less than $200 or even less than $100 a month for a decent amount of coding. So, if you didn't know, Sonnet 4.5 actually does have a 1 million token version. So, you can see here, for example, if we just look at the models here, Claude Sonnet 4.5, scroll down a bit. Context window 200k/1 million tokens, right? So you can actually use the 1 million token model with Sonnet 4.5, which is extremely interesting. So you can see here, Claude Sonnet 4.5 supports a 1 million context window when using this context 1 million 2025807 beta header long context pricing, blah blah blah, nobody cares about that, right? So how to do this, right? I'm currently on API key, you just use /model and put sonnet 1 million context. Now, this might not seem like a big change, but let me just talk a little bit about this. When you are using this in conjunction with sub-agents, you can do things like have this entire conversation that I've had so far with Claude Code without running a single, single compact, right? So, I'm currently porting Grove over to Next.js. If you don't know, Grove is one of my SaaS projects that I have built, and you know, it's all sitting on HTML, CSS, and JavaScript. I'm just porting the entire thing over, and hopefully, I'm going to do that entire project in one Claude Code conversation because I'm using sub-agents at the same time.
Before continuing, guys, just a quick word from our sponsor, me. We're currently looking for a few projects in the first part of 2026 all the way up to 2027 even, to build applications, AI applications for people. I've been there. I've done that. I currently have three live SaaS platforms, as well as many other client platforms that we're working on in the background as well. If you've always had a dream and you've always wanted to build something, but you don't, you've never really been able to justify investing $100k, $200k, even $50k into an MVP, now's the time, guys. We are offering our development services. We're extremely cost-effective compared to other companies because obviously, we're AI-augmented. There is a link in the description and in the pinned comment of this video. Thank you for your attention, and let's jump back into the main content.
This is an extremely powerful method, guys. But not only that, because Sonnet is actually not that expensive, if you don't want to pay $100, $200 bucks a month, just add a limit of $50. So, you don't even need to necessarily add a limit, right? Just go to billing on console.anthropic.com, add the amount of money you are comfortable spending in a month, turn off auto-reload, right? So let's say you want to spend $50, add $50, and then when it's coding with sub-agents or whatever, give it a little prompt inside its user memory or whatever saying, "Use Haiku 4.5 for a lot of the heavy lifting." That will probably mean that you can get entire projects done spending £50 a month instead of £200 a month on the max plan. Now, this is obviously not necessarily cost-effective, but if you're, if you don't have $200 or $100 and you still want to use Claude Code, this is the way that I would go.
Okay, number two is you can spawn unlimited amounts of sub-agents in this context window, right? So, with the 1 million, without running a slash compact, and the way that you do this, and this is really, really important to know, when you give the prompt, let me just explain a little bit how this works, right? So normally, you have a 200,000 context window, right? And you spawn sub-agents inside this context window, and then each sub-agent has its own context window. So you effectively get more use out of it, right? But you still have to run /compact because the overall, like orchestrating context window of 200,000 will eventually run out. But with 1 million, obviously, the original one is much, much larger, which means you can fit more sub-agents into this conversation, right? And the really cool thing about this is you don't have to repeat yourself. You don't have to keep mentioning things over and over because it's the same context window. You've never lost the context. And the really cool thing is you can async these agents, right? And if you're using Haiku, it's not even that expensive. It's less expensive than you might think. Yesterday, I spawned, I think it was, 178 agents at the same time. Now, what it did was it worked, right? I ported the entire project, but I had to go back because, you know, I, we didn't like the design that it created, blah blah blah blah blah. But basically, it spawned 178 agents at the same time within Sonnet, right? Didn't even hit a rate limit. We hit the Opus rate limit, but that's fine. We just moved over to Sonnet afterwards. And it did the entire project in like 10 minutes. Now I'm spending hours and hours and hours doing the same thing, but I could have, if we liked the design that it came up with here, I would have been done in 10 minutes instead of, you know, 8 hours or whatever it might be. I did not know that there was basically no limit. And with the 1 million Sonnet context window, and by the way, Sonnet is actually a pretty good model right now. So, just like these things together really, really make this an effective method. The 1 million context window with the ability to spawn unlimited agents within that 1 million context window is proving to be insanely good for, yeah, for everything basically. So what I'm doing now is I'm doing a similar thing, but I'm not actually creating hundreds and hundreds of agents at the same time. I'm just creating a few agents at the same time. I'm letting it just complete things bit by bit. I would probably recommend this if you're doing a large project and you want a bit more granular control over it. Obviously, spawning 178 sub-agents at the same time is interesting, but it doesn't always work. So, just be wary of that. And I mean, this is a much better method, just doing everything kind of more slowly, not just spawning 100 agents at the same time. And another thing as well is this concept of making sure every sub-agent gets the same prompt, right? Then you have some consistency. The other option is to do what I did, which was I created an HTML, CSS, and JavaScript version of what I wanted to create first, and then I moved that over. Now, we didn't go with this design in the end, but I could have gone all the way with what we had here. I actually liked the design, but yeah. Any.
Okay, and then the final thing that I've learned this week is the Chrome DevTools MCP is way better than Playwright's MCP. And not only that, it can get some extremely useful information if you're making an SEO tool, for example, Core Speed Vitals, whatever they're called, Core Web Vitals, which basically shows people the speed of their website. So, the plan is to put that inside Grove and then probably have some kind of upsell to redesign people's websites because one of the, one of the silent killers of Shopify websites is actually having a heavy theme. But not only that, just specifically using it inside Claude Code. Playwright doesn't work that well. There's some issues where it tries to open too many pages at once, and it just doesn't work. However, Chrome DevTools just works instantly every single time. I'm not really surprised by that information, to be honest with you guys. It makes sense because it's, it's built by Google, obviously. So, of course, it's going to be good.
These are the kind of three things that I've learned this week. There are probably other things as well. If there are, I'll continue making videos like this. But I just, I kind of just want to brain dump, um, with important information that I think people can genuinely get use from and apply to their coding immediately. So yeah, if you want to know where I've been the last couple of days, guys, I've been porting SEO Grove over to Next.js and then redesigning the entire backend on Convex with whole new AI agents because I've just learned so much in the last few months. Um, and yeah, we are actually looking to do the same kind of thing for other people. So, if you do have a project, I left a little ad in the video. Um, but yeah, definitely check out the link if you're curious about working with us on building an MVP for an AI project or otherwise. I'll leave the video there, guys. I really hope this was useful. This has got a lot of good information in it. Um, just one tiny thing to mention is you can actually put MCPs into your agents as well. It's not just for Claude Code. So, what I'm actually going to be doing is I'm going to be doing LCP tests on people's websites using the Chrome DevTools MCP inside SEO Grove. Right? So, I just wanted to mention that quickly. Thank you so much for watching, guys. If you're watching all the way to the end of the video, you're an absolute legend. I'll see you very, very soon with some more content. Peace out.