Claude Code, oh my goodness. I've been building in tech for 20 years, both as a founder and as a VC, but nothing has ever captivated me the way Claude Code has. Between building my own AI startup and even starting as a guest judge for some pretty major hackathons, I've spent over a thousand hours figuring out what really works. So, here are my exact plugins and step-by-step system I'm using to build my startup using Claude Code. All right, so to build properly with Claude Code, the first thing we need to fix is Claude Code's strange tendency to go from super genius to complete idiot in the middle of a session. I mean, I'm sure if you've built anything with Claude Code, you've seen this happen. >> >> And you might think it's because Anthropic is dumbing down the model or you did something wrong, but it's not actually your fault. It's a specific problem with the way that these LLMs are built. Let me pull up Claude Code and I'll show you exactly what I mean. So, when you start up Claude Code, when it first boots up, it has no understanding of your code, your program, your project, anything. It's completely dumb starting from scratch. >> >> And the way that it learns about your project is it builds what's called context. But there's a huge problem here, and I'll explain it to you with some random numbers. So, let's say I said, "Generate me five random numbers." If I gave you these five random numbers and asked you to remember these, you could probably spend 10 seconds looking at them. And if I ask you in 30 seconds what they are, you can probably recite them back to me. But if I give you five more random numbers, >> >> you might start to forget the first five. And if I keep doing that over and over, where I keep asking for more, then eventually you're going to start forgetting those first numbers cuz I'm giving you too much context. >> >> Turns out Claude works the same way. The more that you work with Claude, the more your context window fills up, and it starts to become ridiculously dumb, forgetting things that you just told it a few minutes ago. And that might be fine for, you know, forgetting silly random numbers, but imagine if you're building a project and you start writing code, and five minutes later you've already forgotten the code that you wrote, and you start writing duplicate code, making a mess, breaking things. That's where the things really fall apart. So, to fix that, we need to manage our context window. And thankfully, there's a super simple plugin that's going to allow you to do this. So, in your terminal, before you enter Claude, all you're going to type is NPX CC status line >> >> at latest. And when you run this command, it's going to install a custom status line for your Claude code. And from here, you can actually go in and edit your lines. Uh and I will show you my exact setup. So, in line one, I have the model, a separator, context percent, separator, session cost, separator, session clock. And then if I go back, line two, I just have the Git branch, separator, and Git work tree. You don't have to understand all of this now, but if you don't know what you're doing, I would recommend you set it up like mine, and I will show you why in a second. So, now you can see when I boot up Claude, we have this new status bar that shows model, context, and so on. This context percentage is going to be the most critical. What you want to do is as you give Claude any commands, even simple stuff like what is 2 + 2, you will see the context window starts filling up quite quickly. My rule of thumb is anytime I see that creep up to around 50% or so, I start to get suspicious because I know Claude is already forgetting things that I've told it, and it's getting dumb. You can't get to 100% and then start forgetting stuff. It gets dumb way sooner. And so, if you want your Claude to stop hallucinating and start writing real production level code, first things first, don't let your context go much past 50%. And yes, you know, Claude has these commands like you can do compact, things like this. Well, one thing to remember, friends don't let friends compact. Worst of both worlds. You lose what you're working on in terms of context. You remember a little bit, but then you still have that context poisoning from before. But there's a way to make this even more powerful and ensure you almost never run into context issues again. Because even if you're managing context with the latest models, they burn through contacts really fast. So, the real solution is to launch sub agents, where your main Claude terminal acts like an orchestrator, dispatching dozens of many sub agents to do things like write code, review code, test bugs, etc. And critically, they do it all in their own context window, so you never run into these forgetting issues. And then, when they're done, they report back to your main Claude code and saying, "Here's what I did. Here's the update." So, with this technique, your main Claude code instance just gets little updates to hold in context instead of the full workload, and it can actually be way more efficient and effective. But, as you can imagine, that is a nightmare to run and maintain as a human. And that's where my favorite plugin of all time comes in, Superpowers. Superpowers, like the name implies, literally gives your Claude code superpowers, >> >> so you can orchestrate dozens of sub agents in real time to build anything from simple features to incredibly complex apps. It's even officially endorsed by Anthropic. It's just not included out of the box because it wasn't actually built by them. So, here's how you install it. So, this is the second critical step in our Claude code setup. You're going to go to {slash} plugin, and you're going to see here, you have a list of discovered. We're going to want to add >> >> some plugins. The first being Superpowers. Superpowers is going to allow us >> >> to essentially move from just randomly blowing up our context window and building whatever we want to using sub agent driven development with built-in code review, systematic debugging, and much, much more. >> >> It actually takes your sort of vibe coding setup and automatically builds in the workflow that the world's best companies are using today to build software. >> >> So, it'll ask you how you want to install. I always just install for user scope. Uh but, we're not done yet. Let's go back to plugin. And then, while we're here, go ahead and add Code Simplifier. With that, we can exit Claude and go back in, and you will see now Claude has Superpowers. So, the way you're going to use this is there's three core commands: superpowers brainstorm, superpowers write plan, and superpowers execute plan. So, superpowers brainstorm is actually where you're going to start any project. In this case, I've set up L7V Claude setup. It's a brand new untouched repo. So, I'm going to say, "I want to build a webpage for my VC fund that showcases our investments and also the companies we are directly incubating." >> >> Um I'll say, "I want it to look super cool and have cool effects." So, that is not a very specific prompt, but as you can see, superpowers knows how to actually do this properly. So, it's going to explore the context, ask clarifying questions, propose two to three approaches, and then present a design doc to me for approval. So, it's already running an amazing workflow that you can just copy and paste into your system. And as you can see, Claude code is running through, asking me context, asking me questions, presenting different approaches, and ultimately building a design spec for me. So, I'm going to say approach A, approach one for the CSS first. I'll review the architecture. Now, critically, your job here is to review this. Don't just click yes to everything. I want to look and say, "Hmm, do I want home, portfolio, about, contact? Does this make sense? For the data structure, let's see, I want a one-liner, sector, type is investment or incubated, logo." Oh, yeah, that sounds good. Yes. And so, it's presenting ideas to me, but I am still in charge, and that's very critical when you're building. You need to be in the driver's seat. So, I'm just going to tell it to stop here and transition to the design doc. Getting way too in-depth for this fake project for my tutorial here for you. Uh so, we're going to make the directory. And now, what Superpowers is going to do is it's going to write a spec plan >> >> to your folder called docs/plans. And it's going to go ahead and build a get repo for me, which is amazing, like Superpowers does so much out of the box. So, even if you are not technical, it'll do all the kind of techy, nerdy stuff for you and just make it really clean and able to actually build production-level code. And just like that, we have our spec plan here, and it's automatically transitioning now to implementation planning. >> >> So, in implementation planning, it's either auto triggers or if I had typed Superpowers write plan, it would have started the implementation plan process. What this does is it's actually going to go through and build a full plan for turning your spec into a real-life product. But critically, it's going to do it in a way where you can actually launch sub-agents. So, you don't want any of these poisonous context windows and things like that. So now, with these Superpowers, you have a new three-step workflow. First, Superpowers brainstorm. This is how you're going to build any feature that you want. >> >> It's going to ask you clarifying questions, explore different approaches, and ultimately turn your idea into a spec document. Once you have the spec document, you're going to use Superpowers write plan. This will take your spec and turn it into a real implementation plan with line-by-line code changes and all the little details that make things production-ready. And then, Superpowers execute plan. This is where the magic happens. This will automatically dispatch sub-agents, >> >> code reviewers, testers, all to automatically turn your implementation plan >> >> into live code. Superpowers ships you from this sort of ad-hoc vibe coding to real spec-driven development. It's the same way the biggest tech companies on the planet, like Netflix, Spotify, etc., are actually using AI coding. But there are three more things we need to really make it shine. First things first, we need to upgrade our thinking powers. Because now we're relying on Superpowers to really do all the brainstorming. So, the better it can think, the better your product will ultimately be. And the best way to do that is to install something called sequential thinking. What sequential thinking does is it uses chain-of-thought reasoning to actually allow Claude to think deeper, longer, and ultimately come up with better insights about what it wants to build. The install for this is a little more technical than Superpowers, but thanks to Claude Code, we can actually make it pretty simple. Just spin up your Claude Code instance and say, "Please install sequential thinking MCP server." And that's it. Tell it enter, and Claude will actually go figure out what sequential thinking is, how to find the MCP server, and set it up for your computer. Once it's done installing, just go ahead and restart your Claude Code instance, and now your Superpowers have been upgraded to also use sequential thinking. But all that thinking is useless if Claude doesn't actually know what to think about. I mean, one thing that happens a lot with Claude Code is it will just hallucinate either outdated APIs or outdated information from its memory, which generally lags 6 to 12 months behind. And so, even though it's using Superpowers, sequential thinking, it can still build things that just break in actual live use cases. So, in order to fix this, you need to give it a data source of constantly up-to-date knowledge about all the other systems it might be using to build. And that's where the next plugin comes into play. It's called Context Seven, and all you got to do is go to plugin, and it's an official Claude plugin. You click enter install for everyone, and once you restart your Claude Code terminal, you will have Context Seven. What Context Seven does is it instantly gives your Claude Code up-to-date knowledge on every single API, service, and library it could ever try to build with. And by just adding these simple and following this workflow, you've already turbocharged your Claude workflows. But there are a few ways to make it even better. See, Claude Code is great, but it's kind of like having a junior dev that says they're doing everything right, but they won't actually let you see their screen, so you kind of just have to close your eyes, hope they're writing things properly. I mean, yes, you could go and double-check their work, open up the files that they're writing, and things like that, but let's be honest, one, we need to move fast, and two, it's pretty clumsy. Your files aren't really linked to Claude. It's kind of just hard to navigate between the two, and this is where the next tool comes in. It's actually going to be an entirely new AI native terminal that makes it way easier to manage Claude code, get work done, and see what's actually going on. So, you can see here, I've opened up the terminal. It's called Warp. Now, Warp is an incredibly powerful tool, and it has a lot of different features. We are not going to use all of them. We're just going to use a few, and I'll show you which ones matter to us. So, when I open it, I'm going to open up my repository, and then it'll ask me for some permissions. I'm just going to skip all those for now, because we're not actually using the built-in AI that Warp comes with. We're going to be using Claude inside Warp. So, to do that, we can open our repo and just type Claude, and there we go. We're back to our very familiar fun Claude interface. However, unlike the regular terminal, which just allows you to run Claude, Warp has a few features built in that are going to make your life way better. So, the first thing is this toggle panel, where you can actually look at your repository. So, remember earlier we used Superpowers to write this plan. I can actually click into this >> >> and see the exact plan that Claude wrote. This is the implementation plan, and this is the spec design that I wrote. And so, as it's generating these different plans or different code, I can be talking to it in this terminal, look over and see the actual work that it's doing. And this is so critical, because if you want to actually build real things, you want to review these plans and say like, "Yeah, does this home page hero image sound like what I want? Does this portfolio toggle make sense to me?" And if it's not what you're looking for, you're going to tell Claude. So, just this ability to see the files and talk to Claude at the same time make this a game-changer. Let me show you two other features that I also really like. I'll close my panel. The other thing that I really like is if I press command D, at least on my Mac, I can have another Claude terminal instance, and I can keep doing this and spin up lots of different Claude code instances. So, if you've heard these things of like, "Oh, I have multiple Claude code terminals open at the same time." Warp makes that really easy. The other thing that it does is it also allows you to have new tabs. So, if instead of pressing command T, I press control T, I can have a tab here. So, I can have my multiple terminals here, and then also tabs. And so, as you can see, very powerful to manage lots of different Claude instances >> >> and check the actual work it's doing, look at the docs, look at the plans, review them yourself, and also just familiarize yourself with your code base. Warp is completely free to install. They're not sponsoring this video, but honestly, I should reach out to them and ask them to sponsor cuz it's so good and I use it every single day. Uh and it'll just make your life so much better. So, go and install Warp. And then, to make it even more powerful, what if we could use Claude code on the go? I mean, yes, of course, Claude code has an official mobile app, but here's the thing. It kind of sucks. I mean, it can't run local files. When you try to navigate to most web pages, most companies have blocked Anthropic from actually scraping them, so you can't access those. And it only has about 10% of the capabilities of a real Claude code terminal. The solution is another free tool called Happy Engineering. You just go to happy.engineering to check it out. And you can see it has Android, it has Apple, it has a web app, has everything you need, free, totally open source. The install process is super easy. You're going to copy this command, run it in your terminal, >> >> and then you'll do a little bit of setup on your phone. And the next thing you know, you have the ability to spin up real live terminals on your computer that you control from your phone. And the terminals that you control from your phone have 100% access to all the same tools you have when you're at home. And so, you can use the things that we talked about such as superpowers, context seven, sequential thinking, all things that you can't actually use from the official Claude code app. Because these terminals from your phone are actually running on your computer, when you get back home you can just resume the work on your laptop seamlessly like you never left. As long as your computer is plugged in, you can ship new features, new code, and actually build. Honestly, this has been a game-changer. And once you get this, I think the biggest thing will be not being socially awkward and whipping out your phone for that new idea and shipping it when you're hanging out with friends or, in my case, on a date with my wife. I have to kind of put the phone away and be present because I can always ship stuff at any time now. So, that's the full setup, but there's one more step to really go from beginner to pro. And that's to start building custom skills. For example, in my own startup, I'm building an AI game right now, and we need to add creatures to the game pretty regularly. >> >> Now, every creature needs an element, an archetype, a description, link to the art that we have for it, stats, attack styles, ultimate abilities, etc., etc. Instead of me manually typing that every time, I want to make this process of adding a creature a skill to Cloud Code. Here's all the specifications. Here's how I think about it. Here's the process. Here's what the files are. Here's everything you need to know. Build a skill. Superpowers does its thing, >> >> uses sequential thinking, and actually built me a custom skill in Cloud Code. So, whenever I want to add a creature, I just type in creature-forge. Cloud runs the full process, and I'll be honest, the first time I ran it, it had some issues, but I gave it feedback. Hey, here's what's wrong. Here's what's broken. Fix these things. And after a few runs, it was perfect and it's saving me hours every day. And so, as you're building your project, you want to ask what are the things that you're doing often and start building skills for them. I now have a dozen different skills that I'm using every single day that really save me time, make things more efficient, and more repeatable. So, there you have it. That is my full step-by-step workflow that I'm actually using to build my own startup. Now, you know how to build, and if you want to know what to build, check out this video where I share the best startup ideas. Hope it helps.