Planning In Claude Code Just Got a Huge Upgrade

So right here I have regular Cloud Code building me a plan and it's been running for almost 4 minutes now. And in this terminal I have it doing the exact same plan, but now it's doing it with Ultra Plan mode. You can see it says monitor progress in Cloud Code on the web, and right here it says Ultra Plan ready. So if I hover on that, click enter, I'm able to review this on the web and this finished in about 1 minute, whereas this one is still running and it's been over 4 minutes now. And when I go to that web session, we can see the actual plan. I have context, I have what already exists, the the new approach with new files to be made, modifications, as well as just a final verification. This gives me the plan in a lot more structured way. I also sometimes get diagrams based on what it's actually planning. And in here, I can go ahead and leave emoji reactions on certain places, which I think is a nice touch. Or I could go ahead and comment on specific elements. As you can see right here, I could add a comment. And then when I want to send these comments off, Claude will go back and do another iteration of the plan. Or I could click on this button, which is approve plan and teleport back to the terminal. So I do that and now it's going to send it back. And if I jump back into that session right here, we can now see the plan has been received and we can either implement it here or we could start a new conversation and implement it there. So if I just click on this, now we have that full plan that Claude Code on the web actually built for us and it's being worked on right here locally. And the whole time this original section is still planning, it's still asking me questions and it's just moving very slow. Now I'm not saying that speed is always the most important thing, but if you're getting speed and higher quality, then that is a good combination. Now what else you'll notice is in this session of the cloud where it sends off my prompt to be ultra planned, it's still able to look what's inside of our directory because it syncs to Git. So right here it starts off by exploring what's in the directory to see what it can build on, and it was able to find existing skills and scripts and just able to leverage those and use those in the plan. So basically this is all thanks to the new kind of secret feature called Ultra Plan. Plan in the cloud with Ultraplan. You start a plan from your CLI, you draft it on the web, and then you send it back to the terminal, or you can just execute it remotely as well. And it's really interesting. I've ran a ton of different tests between the same prompts locally and the same prompts to be Ultraplan. And not only is Ultraplan faster at planning, but for some reason it's much faster at executing too. And I think it's just because the plan is so clear that the agents have a much easier time locally just executing on it and building everything out. And I'll show you guys an example of that in a bit. This is useful when you want a richer review surface than what the terminal offers as far as targeted feedback, because you can comment on individual sections. You get a nice sort of like doc style plan as well, rather than in the terminal. Sometimes it can just be like lines of text and it's kind of messy to see like the structure. You also have hands-off drafting and you have flexible execution. So it's really simple. You just do a /ultraplan and then your prompt, or if you just have the word ultra plan in there, it will ask you if you want to ultra plan. So if I just come back in here and you see I literally just type the word Ultra Plan, it lights up the same way that it would light up if you do something like Ultra Think. So maybe you Ultra Think and Ultra Plan, you burn through your entire session in one prompt because obviously the Ultra Plan is using a little bit more tokens. I'll show you guys an example of that later. We don't know exactly how many, but the way that it's spinning up and it's using a lot more compute resources in the cloud, it's going to use more tokens than if you just do a regular plan over here. But once again, planning is so, so important. It sets the whole structure and path of your project on the right path. So sometimes it really is worth spending more tokens upfront in order to shoot it off. I always think of this quote by Abraham Lincoln, give me 6 hours to chop down a tree and I will spend the first 4 sharpening the ax. So that's basically our planning. Also important to know that this has to be done in the CLI. If we are here in the desktop app and we tried to do something like Ultra Plan, build me a 5-day workout guide. If we shoot this off, we're not gonna get that whole Ultra Plan in the cloud functionality. In fact, it actually just doesn't do anything. And same exact thing if we hop over to VS Code, if we're doing it in the extension rather than in the, uh, terminal Ultra Plan, and then we just, let's just copy in the exact same thing. It's basically just going to do nothing. So you have to use it in the CLI. Okay. So quick explanation of how this actually works, and then we're gonna jump into an actual example that I ran and show you guys what it looks like. So, so we start talking Cloud Code on our machine. Push an ultra plan, it sends that to the cloud where that's where Claude will spin up, you know, different agents there using Opus and it will do a really in-depth plan. Once that's ready, we're able to review and we have this kind of feedback loop, but then we can teleport it all the way back to our machine if we want that to happen there. Like I said, you can launch it by doing /ultraplan. You can just have the keyword in there, or you could, you know, from an existing plan, like you already have one written up in locally, you could say, actually, I wanna revise this in the cloud with UltraPlan. And then of course you have that review loop. You can do comments and you can do emojis, which is a nice touch. All right, so let me show you guys an example of this. So I have UltraPlan on the right, I have a regular plan on the left, and I gave both of these a huge spec about building us a dashboard. So I used UltraPlan over here, pasted it in. I said, do not use superpower skill because I didn't want the local version to use that. I wanted to just get raw. You know, keep the variables the same. And then I wanted to give it to us on a localhost. So I'm gonna shoot this off. You see on the right-hand side it says, do you wanna do this in the cloud? I said yes. On the left-hand side, it just goes straight into plan mode. And once again, they got the exact same prompts. So here is the link. I open this up and it takes me to the cloud where it's actually going to start building out the plan right here. As you can see, it's checking the directory to see what's available. And then over here it's just, you know, kind of doing the same thing, getting acclimated with the environment. Okay, so at this point we're about 2 minutes in, right? And you can see that I got the plan already on this right-hand side. So I click through it a bit. We have different tabs. We also have like a little diagram and it looks really good. So what I do is I end up sending this back to the terminal, as you can see right here, and then it gets brought back into our local session right here. Okay, so now we're at the 14-minute mark and the Ultra plan version's already done. The localhost is spun up. So not only did it create the plan, but it also built everything out. Whereas this regular plan mode has been going for 12 minutes and still hasn't even given me a, a plan to review yet. Then finally, a few minutes later, we're at the 15-minute mark. This is when we actually finally get the plan from the local session. But you can see at the 26, 27-minute mark, it is still being built out. So I end up stopping this recording to just wait until it's actually done. So now I'm gonna switch over to the actual live reaction of me looking at the differences in the two dashboards to see which one really built it out better. Okay, so just the regular plan mode is finally done. It's really weird. I, I normally stand in these videos, but I had to sit because this took honestly like 45 minutes. I'm not sure what was going on, but this thing on the right side, we did ultra plan. The plan finished in like 5 and then it executed everything and built everything in, you know, like another 5 minutes. So about 10 to 15 minutes total over here. And on this side, it literally took half the time to plan and then half the time to build. Like it took a long time. So anyways. We have the two dashboards here. Let's open up the, um, regular plan mode dashboard. So let me open this up. You can see we have our Pulse board. We have a nice little graphic here. We have some stats like MRR, annual run rate, total customers, churn rate. We can change this between all time, 30 days, 90 days, 6 months, 1 year. We have different tabs on the left. We can do revenue and see different tiers. We can also see Enterprise and Pro, top 20 accounts. Obviously all this is mock data. We've got our customers here with a nice chart and by tier, and then we have support. So resolution time, tickets by category, distribution. We can do light mode or dark mode, and that all looks pretty decent. These charts look a little bit weird in light mode, but anyways, that is the plan version. Let's open up the ultra plan version on a different localhost. This one, obviously they're gonna have very similar features as far as what we're looking at with the data. Similar statistics up here, similar charts, everything like that. If we go to 30 days, 90 days, 6 months, 1 year, that all loads in good. If we go down to revenue, we can see we also have a cohort retention. We have Enterprise and Pro, and we got some— our 20 accounts there. Customers, obviously we've got the same exact type of layout. And the support over here, if we go to light mode, let's see what these look like. These all still look very, very solid. So I mean, it's not a huge difference, right? And for something like a dashboard like this, from an aesthetic point of view, it's kind of hard to tell like which one objectively between these two did better. If we wanted to look at things like, um, how accurate the data's being pulled in or reflected, or how quick things load up or how buggy it is, we'd have to do some deeper investigating. But also take a look at this. In the local plan version, it took 131,000 tokens. The Ultra plan version took 82,000 tokens. Now, of course, the bulk of the planning work was done in the cloud, so those tokens don't count. You guys saw earlier when I was doing this, this is the section where we did our Ultra plan on the web, and I tried to run a /cost to see how much this took, but it wasn't working, so we don't know exactly. But I will say it's using more tokens than the local, so it looks like this is using 50,000 more tokens. I'm assuming that the planning section for the Ultra plan took at least 50K, but it's hard for me to say exactly. But I think that the plan was so much better because for some reason it was able to execute on the plan so much quicker and better. And they were both using obviously the exact same model. In all the other tests that I've done, Ultra plan has always been quicker. I mean, in the cloud there's so much more compute compared to what might be going on here on my device. But yeah, like this was a noticeable difference, not only with the plan, but when it came to the actual execution. Like I said, the execution over here was maybe, maybe 10 minutes. The execution over here was upwards of 30 minutes. Okay, so I tried running some experiments for you guys. I did a session with API billing and I ran a local plan, as you can see, and this ended up costing me about 1.5 million tokens in and 23,000 tokens out, and a mix between Haiku and Opus. Now unfortunately, it doesn't let you do the ultra plan with API billing because you need basically a Pro or Max subscription in order to do the Claude code on the web. I tried, it wasn't letting me do that. So anyways, I did do this though. When it came back, what I did is I just checked how much of my usage did it actually go up, and it went up 1%. So we don't have an explicit number of tokens that we get allotted within this max 20x plan, so I don't know how many exact tokens it was, but I am going to estimate that it was more than 1.5 in and 23,000 out, and it was about 1% of my 20x max plan. So hopefully that gives you guys a little bit of context and just at least gives you a nice baseline. Okay, so let's talk about a few other things that you need to know about how this works under the hood and what you need to know. So first, I'm going to start off with like your project. If you want to Ultra Plan, it has to have some sort of Git or GitHub repo synced to it because it has to be pushed to the cloud somehow. So if you try to run Ultra Plan in a project that's only local, it's going to say, hey, you can't do this on the web yet. It's as simple as just initializing that Git, pushing it super easy, and then you can do it because obviously in the cloud when it's doing the UltraPlan, it needs to look through everything that you have and it needs to be able to actually work on your code base. Otherwise it's gonna be a very, very generic plan. Now after that, sometimes I've noticed that it's not calling my skills even though it looks through the skills are things that get invoked. So if you're not specific enough on your plan for the UltraPlan, then it may not use the skills the first time. And actually all of these visualizations here that I made, I made using UltraPlan. So I'll put up a clip here where I'm showing how that worked. And what you'll notice is that I had to give it correction on using my visualization skill because it wasn't pulling that up for some reason. Okay, so here is that clip now. It says, I want you to UltraPlan and research Cloud Code's new feature UltraPlan. And I kept saying UltraPlan because I love when it, when it lights up like that. But anyways, we shot this off, right? And it goes to the web and now it's doing a fetch on our project and getting an understanding of what we want it to do. But what you'll notice here is that it doesn't actually create the Excalibur style diagrams using my visualization skill, which uses Nanobanana and it has like a prompting guide and a style guide. But it thought that I just wanted like markdown, um, you know, mermaid style diagrams, which is what it built as you can see here. So I realized that it was creating these not the right way. I added a comment right here that says, hey, I already have a skill set up that does this, so I want you to do it that way. And then when I shoot off that comment, it basically comes back and says, okay, I didn't see that skill. What's the name of that skill? So I had to tell it explicitly right here. Which is the skill that it's looking for? So it asked me this question and then I say it's called visualizations. And then once I said that, it was able to find it and it was able to use it, which is good. But just something to keep in mind, even though it can see your whole project, sometimes you still have to prompt it in the right way. And once again, this is just in research preview. But anyways, then it came back the right way. So I sent it back to the terminal and then I was able to actually have everything be built out from my session here because the plan had been approved. And then I just go ahead and implement it locally. Okay, so now let's talk about what is actually happening because we're sending it to the cloud and it only works in the cloud. So that makes you think, why? So obviously it's getting more compute, but I wanted to understand why and like what that even means. So I had this agent here do a deep dive into the Ultra Plan feature. It looked at source code, it looked at community forums, and it looked at the actual documentation. And here's what it came up with. Ultra Plan offloads your planning session to a cloud-hosted instance running Opus 4.6. It uses multi-agent exploration to build a deeper plan than the local mode can. Then it either executes in the cloud or goes back to your terminal. So what I thought was interesting is the multi-agent exploration. So here's a quick comparison table. When you do local plan, it runs in your terminal. When you do UltraPlan, it runs in Anthropic's cloud container runtime. The model is whatever session you're using. UltraPlan always uses Opus 4.6. The planning approach is a single agent and it's linear thinking, whereas UltraPlan is multi-agent. It is 3 parallel exploration agents plus 1 critique agent. So obviously this is gonna use more tokens, but like we talked about, if you use more tokens up front, it's definitely worth it if it's, you know, faster and higher quality. Now the max compute time over here is bounded by your session, but on the cloud it's 30 minutes as a cap. I've never hit that and I don't think I will because of how quick it's been, but there is a cap on it there. Now the other interesting thing is your terminal gets blocked while you're in plan mode because your session is focusing on that. But in Ultra plan, if you send your plan to the web, you can continue talking to that terminal session and wait for it to get teleported back. Now truthfully, I don't know exactly when I would do that. I'd probably just spin up a different session because I'd wanna keep that context clean so that I can just execute when it comes back. But that is something that is a difference. And then the review interface, you basically get text dialogue in your terminal. If you're using VS Code with Cloud Code and you're using like this extension rather than the terminal, then you do get a nice like, doc and it looks a little bit better. You can leave comments on it, but if you're not using that and you're using the terminal, then it's definitely going to be better in the Ultra plan version. And really the only requirement here is that you have to have it synced to some sort of repo so that Cloud Code with your web account can actually log in and look at your project before it starts planning things out. So I'm sure that this is going to evolve as more people use it, as more people start to give feedback about it. Um, when I was testing it out, I did run into a few issues where I was getting this authentication error. I'm not sure if that was something internal or something I did wrong, but it just seemed to pop up randomly and then I'd run it again and it'd be fine. But it also seems like we'll start to get more clarity around how exactly this works soon, why it's so much better at executing so much faster, and the fact that these are parallel exploration agents. So I'd be curious how they're prompted differently so they're not just duplicating work, or maybe they are, maybe that's the point. Um, I don't assume that this is some sort of agent team that they can talk to each other and share a task list. And what I'd be curious to know is if we could get more visibility on how many tokens are actually being used in that plan. But yeah, it would be nice to be able to see, hey, the Ultra plan is finished. And by the way, here's how many tokens it just ran. Anyways, that is going to do it for this one. So if you guys enjoyed the video or learned something new, please give it a like. It helps me out a ton. And as always, I appreciate you guys making it to the end of the video. I'll see you on the next one. Thanks everyone.