Yup, Claude Code Plan Mode is here: The Senior Engineer's Workflow

Key Takeaways

Let's open claw code. Shift tab. Shift tab. Bam. We're in plan mode. The debate is over. Claw Code is taking over the engineering world. The Claw Code team continues their red hot streak as the most cracked engineering team. How are they able to do this? The creator of Claw Code said it himself. Generally at Anthropic, we have this product principle of do the simple thing first. The team has stealth launched plan mode. It's simple and powerful. How does plan mode work and why does it exist? There are two powerful ways to use plan mode. We're going to break down right now. One of these techniques takes a little more time but pays you dividends. You want to make sure you're using the right method so you don't waste your valuable engineering time. Let's master claw code plan mode in 20 minutes. No one knows clawed code better than claw code. Let's ask it to tell us about plan mode. Reflect on your system prompt. Hi, can you explain plan mode? So, with certain features, you can be sure the system prompt has information about it. Here, we're quite literally asking Claude code to tell us about its new plan mode. You can see incoming tokens streaming in. And we look at that. Here is the details on plan mode right out of the system prompt. Plan mode is a special operating mode where I research and analyze, create a plan, present the plan for approval, wait for confirmation. So, this is really important and this is interesting. I didn't know it had an exit plan mode tool. This might be a fresh feature that just rolled out. The cloud code team's engineering velocity is mind-blowing. I'm restricted from making any file edits, running commands, creating or deleting, or making commits. And here we get a nice breakdown of when this is useful. We're working with AI here, so we can just ask stupid zero judgment questions that might have already been explained in the previous prompt, right? Ask stupid questions, validate your assumptions, learn fast, and move even faster. Can plan mode create or modify any files? Let's say like plan MD, can you modify that? So, I'm just being super super clear with cloud code. I want it to repeat things back to me. After all, we are operating in plan mode, so this is what it's for. No, plan mode cannot create or modify any files, including plan.MD. So, this is really about building up the context window inside of cloud code. There's no actions. There's no mutations happening here, right? So, let's be totally clear here. So I'll say so the purpose of plan mode read files understand the codebase by loading your context window with important information before we start working then switch out and edit correct me if I'm wrong. I like to use this correct me if I'm wrong to decrease sickle fancy basically decrease the bias in the model and present a more neutral question format. These models love to tell you exactly what you want to hear because it's good for business for these AI labs. So, you have to be really clear. So, I like to use this. Correct me if I'm wrong. It's just a little bit more neutral. You can see cloud code validating exactly what we were saying. Yes, you're exactly right. That's a perfect understanding of plan mode's purpose. It's research and planning phase where I load, analyze, design, present, then after you approve, exit, execute, and then we run all the commands so on and so forth. Right? This is really interesting. Notice how claw code is converging on the senior engineers developer workflow. First, we understand the feature by gathering the right information. We build up the state space of the problem and the solution. And only then do we start building. Before plan mode, we'd have to do something like this, right? You have your long prompt and then we say something like this, right? Don't write, update, or delete anything. Just read and plan. So with two shift tabs, we can immediately just enable that prompt modification. Planning mode means you can focus on priming your agent with the right context before getting work done. Great. So we have this new plan mode. Let's put it to work. I'm going to open up cursor here and let's enhance our previous proof of concept codebase, the infinite agentic loop. In this codebase, you can have claw code generate infinite versions of a solution to your problem until it breaks. Let's go ahead open up a new terminal window here. Fire up clog. We'll make sure we're running the state-of-the-art model opus 4. I'm going to run a prime command to get our context window loaded with key information about this codebase. You can check out the prime command inside of the custom slash command. This is a really simple one. We're just running get ls files and we're having cloud code read a couple of files. So you can also see we have claw.md added to this directory. So claw code is now primed. I'll hit shift tab tab. And now we're in plan mode. In the beginning I mentioned there are two ways to use plan mode. You can plan then write code or you can plan write the plan to a spec file and then have cloud code execute on the spec that you created. Four out of five times. I recommend the second technique specifically for mid to largesized features which you should be aiming for. We have more than enough compute and intelligent models in great tools. So aim big, hand off more work to your agent. Do more in single swings to save time and understand how far you can really push these models and these tools. So right now our infinite agentic loop set up in this codebase. If we open up source or source infinite, you can see they're generating single files. And in our previous video, we looked at these UIs, we can quickly just take a look at one of these. If I open up UI hybrid one, copy the path to this, and I just type Chrome this inside of bash mode, which you can trigger inside of cloud code with the bang or the exclamation. You know, you can take a look at this. Basically, we're just generating infinite UIs that could do many different things. If I go to two, you can see here's another one. Here's another one. Right? So, we just generated all of these interesting user interfaces. Right? Here's another kind of wild one. Here's another. Right? We can go, you know, we had tons and tons of these generated. If if you open up these directories, you can see all of them. You can check these out yourself. Okay. So, we ran a prompt that generated a ton of UIs inside of a single HTML file. Okay. So, very cool. What we're going to do here is update our lower level prompt that we passed into our infinite agentic loop prompt and we're going to have cloud code plan out exactly what we want done here. So what we're doing here is effectively metaprompting. I'm writing a prompt that's going to write a prompt that is a plan aka a spec. What I want to do here is just something simple. I want to showcase the power of plan mode. And instead of writing single self-contained HTML files, what we'll do is we'll break them up into their own isolated directory with a CSS.js and HTML file. Okay? And then we're going to embed all that in a new plan. So let's go ahead and get to work there. And all we need to do is be concise and write a great prompt. So I'll say let's plan specs V3. And I'll say V4. So I'm going to update that. That's what we want to generate. We want to build on specs v3. I'm using clause autocomplete here with the at@ symbol, but instead of writing a singlehtml file split into CSS.js.html, we also want to store these three files unique directory. All right. And that's it. So in plan mode, Claude Code is now going to get to work filling up its own context window with the information it needs to plan this workout. Okay, so we're going to kick this off. I'm going to leave open the specs directory. Nothing is going to get created here, right? This is all going into Cloud Codes context window. It's not going to create or modify anything, right? Just like it told us here, just like we learned by prompting and asking questions against Cloud Code. So let's go ahead and take a look at this. It's reading context. It's figuring things out. It's understanding what it needs to do to build out this V4 spec, right? So, it has a clear understanding of requirements and now it's going to present a plan. And so, this is quite interesting, right? It's planning inside of its own context window. It's just looking at your codebase, gathering information, reading files, right? It's all read. There's no mutations, there's no create, there's no update, there's no deletes. It's a very safe mode. And very importantly, this mirrors the beginning of a senior engineer's process, right? We plan, work first, we architect, we think, we plan, and only then do we build. So let's see what we have here. It's walking through based on my analysis. Here's a comprehensive plan. So this looks decent. You can see it has this key changes from V3 to V4. Each iteration creates its own all CSS, alljs. Perfect. So this looks good. Now I want to ask a question here about the implementation. There's one key detail. I want to make sure every version that we're going to generate with the infinite agentic loop gets their own directory with those three files. So I'm just going to be super clear here and showcase you know how you can use plan mode. So we're basically talking with the cloud 4 opus model running on claw code to be clear. showcase what the output durest mark directory looks like after running three agents. And so output directory is going to be part of our arguments, part of our variables inside of our infinite agentic loop prompt. And so I'm referencing this variable here, which is where all the iterations will be saved. So I'm going to kick this off. I want to be super clear about what's going on. And importantly, this conversation we're having is getting added to the context window. Right? So, check this out. This is exactly what I wanted to see. Right? Scroll up. You can see output dur right there. However, that gets defined in the infinite prompt, right? As a variable. We then are getting our own separate versions. And this is exactly what I want to see. I'm just proving things out uh with cloud code, right? I'm planning with cloud code with powerful intelligence. all my abilities, all your abilities, they're now augmented by cloud code, by Agenta coding tools, and by these models. So, make sure you can tap into it, right? Planning mode really helps us tap into this. So, let's go ahead and kick this off. I'm going to get out of plan mode. So, I'm going to run shift tab, shift tab. I want to get into autoaccept yolo mode. And now we'll let cloud code rip build the v4 spec mirror v3. and be clear about the improvements we are making. And I'll throw in a think hard to trigger the reasoning model here. This is awesome. So, we're going to get our new V4 spec out of this. Okay. So, the important thing here is that we're having Cloud Code write a new plan. There's a big problem with if we just said, you know, build out write the code for this, right? We can't see the full plan without consistently prompting Claude code. And remember, agents are ephemeral. What does that mean? They're temporary. They're stateless. Their context window will be blown away. So, we can improve on plan mode by having it write out the full plan into the specs directory, one of the three essential directories for agentic coding. Okay. And then we can once this plan is written here, we can review it, we can iterate on it, we can tweak it, and we can commit it to our repository, which has many downstream positive side effects. We'll dig into future videos. Make sure you're subscribed so you don't miss out on actionable agentic coding ideas like this. As you can see, week after week, there's a lot of value we can create with powerful agentic coding tools like cloud code. So, we're going to get the spec here written. It's activated the reasoning model. We use the think hard information dense keyword encoded by the cloud code team. Where was that? Let me search for that. Up here. Yeah, right here. Right. I said think hard and this triggers the reasoning model. Right now, Claude 4 Opus is running its thinking process, making sure that everything's super clear. You can see very importantly, we're mirroring again another IDK right there. We're mirroring the V4 structure, but adapt it for the multifile approach. This is key. And while it's loading here, we can open up, you know, UI V3. Again, definitely check out the previous video to really understand the infinite agentic prompt and what's really going on here. But, you know, you can see just at a high level what exactly is going on, right? We're generating uniquely themed UI components that combine multiple UIs into one elegant solution. So that's what we're doing there. And you can see we just got V4 created for us here. 360 lines. Let's go ahead and open it up and make sure it makes sense. Okay. So themed hybrid UI specification V4 looks awesome. Builds on V3 with a critical architectural improvement, separation of concerns, new modular file structure. Great. And so I just want to quickly look through this plan and I want to see this key detail right here. output requirements directory structure UI hybrid iteration number very important right this differentiates it from v3 where we just use the iteration number from the infinite agentic loop and then we just create thehtml here we're creating the directory fantastic looks great I'm catching cloud code here before it makes a mistake I'm going to hit escape uh this is one of the you know great reasons to see the thinking of your model it was about to create an updated infinite command which we don't need at all this is good I'm going to stop it right here then I'm going to boot up another agent here. Open up a new cloud code agent and we're just going to kick this off. Right? So, this plan looks good. It has all the details from V3, right? We just built on the V3 spec. And so, now we're going to kick off this infinite agentic workflow with our new plan. Just to kind of hop back into that idea from last week, this is the power of having higher order prompts. What does that mean? It's just like higher order functions where in certain programming languages, you can pass a function into a function. We're doing that same thing here with auntic coding and with prompt engineer. We're passing a prompt into a prompt, right? The infinite prompt takes a prompt as a parameter that it can operate in. Okay, there's a ton of value here. We're just scratching the surface of what can be done with higher order prompts. More on that later. Again, make sure to subscribe so you don't miss that. Let's go ahead and kick this off. Right, so there's V3. Let's close that and let's kick off V4 in our new Cloud Code instance. Check the model. We always like that big beefy opus model. And let's go ahead and kick this off, right? So slash um infinite. This is a custom slash command inside of the directory. We'll hit tab. And now we need a couple of arguments. So we can go ahead and sync these up. We need the spec file v4. We need our output directory source group. And then we need count. How many iterations are we going to run? Uh let's just fire off 10. We're going to have 10 new versions here created. Let's go ahead and fire this off. Now our infinite agentic loop is going to read the spec output to the source group based on whatever is inside of our spec and it's going to create 10 versions. Let's go ahead and let this rip. And we're doing all this thanks to planning, right? This is all thanks to plans. We have plans on plans, right? Custom/comands. Then we have our specs, right? Another prompt. We had, you know, cloud code in plan mode, right? Thinking, building up context. So much of what we do as engineers in the generative AI age in this kind of second phase is all about building up information sets and organizing information for our agents and really understanding what our agents can do. There we go. You can see we have two waves. We're going to run in two batches. Wave one is going to create five agents. Wave two is going to create five agents as well. You can see their themes. And now cloud code and opus mode is going to kick off five agents in parallel. Again, once you finish this video, check out the previous one to really understand the infinite agentic loop prompt. Here's the magical moment we've all been waiting for. We're going to get five parallel agents kicked off and they're going to start working for us. There we go. So, you can see we have that new directory and we also have those five new agents running in parallel here. Let's go and just open all these directories up preemptively and we're going to be able to watch all of these parallel cloud code agents work together to accomplish every version of this task. Okay? And so if somehow you've been watching this channel and you haven't been convinced of the importance of the fundamental principle of AI coding, great planning is great prompting. You should absolutely be convinced now. Okay. a leading AI lab ghost launched and used the tab key for the feature for quick access. It might seem like I'm micro analyzing, but this is what great engineering is, right? Every detail matters. Every product decision matters and every product decision says something. All right, so you can see five agents running in parallel doing their own isolated pieces of work based on our plans. um you know principal ad coding members and subscribers of the channel you've already known about planning and architect editor prompt chains and you know that key principle that the plan is the prompt right we've known this for some time now you already have this deployed into production you're already using this in cloud code every single day this is just a new you know mode of operating for us as engineers now right but if you're just hearing about this principle of a coding the plan is the prompt check out principled AI coding to learn the other Seven principles of AI coding link in the description. This is fantastic, right? This is working for us on our behalf. We have five agents here. Everyone's up to two tool uses. It's probably going to be three each and then we'll be able to dive in and actually look at what's going on here. Okay, so it looks like hybrid v2 is coming around the corner here. This is exactly what we've detailed in our plan. And thanks to Cloud Code in plan mode here, right, shift tab, shift tab, uh, we were able to very quickly load up its context window with a couple key details, right? Very concise prompts. You don't have to type a lot. You don't have to do a lot with these great tools, but you do need to know exactly what to say and what to ask. We focus on these key ideas of writing concise prompts on the channel all the time. All right, so it looks like all of our versions are getting filled out here. You can see task two complete, task three complete. Both agents had four tool uses each. They both consumed about, you know, 40k of Opus tokens. That means there's a lot of value in these tokens, right? Uh 3 minutes each. Not too bad. All right. And all of our other tasks are now complete. These should all be, you know, ready for us to operate in. Now, our agent, you know, running about 400 seconds here. 5.2K tokens on the top level agent. Um and now we're going to kick off wave two. So, I'm going to stop it here. I I want these tokens for some other engineering work. But you can see it's going to kick off wave two here. I'm going to go ahead and stop it here. And let's just go ahead and look at our results here. Right. So I can type um Chrome. Let's open up one of these files. Copy the reference with command shift R. And I'll just paste that here. This is going to open up our brand new index script style. Still with that key prompt of generating interesting new UI. So let's go ahead and kick this off and check this out. Right. Uh nature's answers. Uh plant your search seed. Very interesting. All right. So, we have this interesting search bar here. We have a filter. Interesting UI there. Let's go on to number two. And remember, these were all generated in parallel. So, this is another interesting one that's going to validate as we type. And phone number, credit card. You see we have complete progress at the bottom. This is a really clean minimalist UI. Looks great. Let's look at V3. Okay. And V4. Okay. Very cool. We have like a music player in this kind of retro neon style. Okay. So there's more versions here. You get the point. Um the prompt or should I say the prompts and this prompt system uh that we have with the infinite agentic loop. It's doing all this work for us. This is very powerful. This is all thanks to planning as a you know key engineering concept that's been embedded inside of Cloud Code, the best agentic coding tool in the game right now. This codebase is going to be available to you in the description just as last week. I'm going to add this new spec here. And I'll leave source group in here as well. Planning is a key element of software engineering. First, we think, then we plan, and then we build. All with our agentic coding tools. There's a tectonic shift happening in the engineering world right now. The winning engineers and teams are setting up their code bases and tools, not for themselves, but for their agents. You know where to find me every single Monday. Stay focused and keep building.