Transcription
[Music] Claude, are you ready to build? Ready to help. What do you want to build? All right, so let's start simple. Claude, go ahead and create a simple Hello World starter coding examples for the six most popular programming languages. Let's go ahead and create them inside of a directory called starter coding. Hello world examples are in the starter coding folder for Python, JavaScript, Java, C++, Go, and Ruby. Fantastic.
Okay, Claude, go ahead take those examples and showcase how to make an HTTP request, make a um go ahead and pass in the URL as a CLI parameter and comment every line of code. updated six programming examples to make HTTP requests using URL parameters with added detailed comments.
Nice. Okay, as you guys can see here, we have real time speech to text running and it's getting fed into our agentic coding tool. And you know, this is really cool, right? Real time speech to text coming in. I love to see this in real time. Inside of the cloud code is programmable codebase. I've got everything dialed into a single file within, you know, 700 lines of code. We have our ears, we have our brain, and we have the voice of our personal AI assistant. Now, we're going to break this down, but let me go ahead and and cancel this large request. This is really cool. We have a personal AI assistant that we can talk to to make changes for us. We're going to talk about how Cloud Code completely changes the game for personal AI assistants in this video. Let's go ahead and continue making useful changes.
So, inside of the script, we have an issue. You can see we have the default list of cloud code tools. And if I search for this, um, I'm not actually using this yet. This is a real problem that I just kind of left in here to showcase in this video. And what I want to do here is have our personal AI assistant fueled by cloud code go ahead and make this change for us. Fire this up again with that same ID. A cool feature is that we can reference previous conversations. I'll show that off in a second. Let's just go ahead and kick this off. And now you can see our personal AI assistant listening to us, right? This is an always on assistant. You can see there it's autocorrecting things as it proceeds. And I'm going to go ahead and pause here. And you'll notice nothing will happen. Okay. So, nothing happened there because the trigger word wasn't detected. My trigger words here are one of these four. And as soon as I say them, the assistant will actually act. Right. So let's go ahead and make this change. Sonnet, go ahead and update our allowed tools CLI parameters. Update these to use our constant at the top of the file. We have a constant called default cloud code tools. Go ahead and use this instead of basically duplicating those items. We should get this picked up here. There we go. So sonnet is one of my trigger words. So it's going to go ahead and actually run this command. So you can see cloud code got kicked off. We should use the spread operator here so that we can reuse that constant that we had at the top of the file. Right? So we had our default cloud code tools. We want to see this get used here on the left. As we'll talk about in a second, you know, the system is not perfect. You know, the first cloud code is expensive. And then the second problem is that we have uh updated the process message method to use the default claw tools constant from lines 85 to 93 for CLI parameters avoiding duplicate tool lists.
Fantastic. And so you know the second issue here is that it does take some time. You can see there audio itself playing took 9 seconds. That's fine. But running our agent coding tool did take a decent amount of time. So you can see that change got rolled in there. That looks fantastic.
So, Anthropic just released the uh web search tool, and I want to play with this a little bit. So, I already have this documentation inside of our AI docs, one of the three essential folders. Highly recommend you set up this directory inside your codebase. Our assistant has access to this. So, let's go ahead and create a brand new plan that'll combine this and our UV single file script so that we can get a concrete demo of how the web search tool looks. So, I'm just going to use my assistant to build out this plan and then implement it for us. So, I'm going to go ahead and just quickly reset here. Sonnet, read a couple files in our AI docs directory. I want you to read the UV single file script and I also want you to read the uh Anthropic web search tool documentation. So, put these together into a single spec inside of our specs directory. This is going to detail how we can build out a minimal version of the new Anthropic web search as a UV single file script. I just want you to create the plan for us here. Write a brand new plan in the specs directory.
Okay. So, as you can see there, I think one of the problems with natural language is that, you know, it just takes some time to like really communicate everything you want. I've created a detailed spec for a self-contained Python script that takes search queries from the command line, uses Anthropic's web search with authentication, formats results with citations, and supports options like domain filtering and location context. The spec covers script structure, dependencies, CLI, authentication, usage examples, output format, error handling, and future improvements.
Okay, so this looks good overall. It looks like we are actually missing our uh code examples. So, Sonnet, update that file. Uh we are missing a concrete code example. Make sure you pull, you know, real code examples from Anthropic web search tool markdown file from our AI docs and make sure that that's added in there. We need concrete examples because we're going to use this as the kind of framework for actually writing this code. So, we have the update coming in here. And so, you know, this is just continuing down that trend of great planning is great prompting. This is the key principle in lesson five and I've mentioned it on the channel a million times now. I'm going to keep mentioning it because you know successful engineering really being successful at anything um it's not always about new different ideas. It's really about doing the same correct thing over and over and over. updated the Anthropic web search UV spec code examples covering tool setup API requests citation extraction and response formatting. It includes a full working example with error handling specific Anthropic error codes and improved citation formatting. The spec is now ready for implementation with real world code from Anthropic's docs.
Okay, so if we take a look at this documentation here, the updated version looks a lot better. You can see our command line interface, web search capabilities. This looks good. We have block domains, there's tools, there's max use. Let me just go ahead look at the documentation again at a high level. Type name, max uses. Okay, great. So this looks right. We just had our personal AI tool build out this plan for us. Read some documentation and now I think it's time to go ahead and implement this.
So the great part about this, you know, this is a really important part of whatever system you're building out if you're, you know, integrating with LLMs and agentic systems. You really want context reusability. This is something that uh I hope the cloud code team spends some more time building out. They just recently launched this feature where you can do claude continue and this will pick up on your previous session. They also just added claude resume. So if we click cloud resume you can see different sessions pop up and so this is a really great way to hop back in specific sessions. But in the system I've built out a simple conversation tracker which you can see here in the output directory. If we click CC you can see all I'm doing is tracking you know basic conversation history right? So there's that intro. Cloud code, are you ready to build? So on and so forth. This is a great way to reload the conversation. Obviously, we're using cloud code here as a programmable tool. That means we need to do a little bit more managing of context. Go ahead and fire this off and let's get this feature implemented. Claude, go ahead and read the Anthropic web search UV script spec and let's go ahead and implement this into a brand new UV single file script.
All right. So, this should be good enough. The big unlock here is that cloud code is smart enough. It's intelligent enough. It has enough searching capabilities, right? It's agentic enough to take this request and it has the abilities to look through whatever it needs to inside this codebase, right? The specs and the AI docs in particular. This should be enough for cloud code to get started, right? It knows to read this spec and of course it can see that specs directory if it does any file searching and then we say let's go ahead and implement this into a new UV single file script. So let's see what cloud code comes up with here. Let's yeah there we go. Nice. That looks great. Um, let's uh see what we got here. I've created a self-contained Python script for the Anthropic web search UV with command line support, error handling, API key checks, citation formatting, and colored output. It includes time zone support, usage tracking, domain filtering, and robust API error handling. There's also a readme with install steps, usage examples, output details, and pricing notes. The script runs via doshanthropic search.py, your search query.
Awesome. Okay. So, you know, really what I want to do here is showcase the uh capabilities of a personal AI assistant. So, I don't want to harp too much on this script. Um, but you know, since it's here, let's go ahead and run this. So, let's open up the readme that was created. And there's our search query. So, uh, we probably need to run CXO on this. Probably give permissions here. CX enthropic search Python file. And then let's go ahead and run a search. anthropic claw code and let's see what we get here. So, I'll be impressed if this uh works in one shot. I feel like that happens more and more. You know, the more you kind of learn to write concise prompts, you know, natural language, written language, the easier it gets to actually clean this code up. So, um looks like there was an issue here. Okay, interesting. So, we got one search and response. I'll search these capabilities for you. So, that worked. This is here. I'm actually going to go ahead and commit this into this codebase. All of this, by the way, is available to you. Link in the description. This is going to be in the cloud code is programmable codebase. We're continuing to explore how we can use cloud code as a programmable agentic coding tool. This makes personal AI assistance not only better, but more performant than ever. As you saw right in the terminal, we can just talk, we can communicate, and we can have Cloud Code do arbitrary work for us thanks to all of its incredible tool calling capabilities. Not only did it write a bunch of cool starter code for us, it also uh was able to write a plan. Okay, so it merged a couple pieces of documentation, we wrote a plan and then that plan contained the kind of you know key framework for our actual code. Okay, so we're using AI plan drafting and then we use that to um actually generate a new piece of code.
While we're talking about this, I'm going to fire this up. And I don't want this additional readme here. So, um, I'm just going to have our assistant clean this up. I want to merge this. So, uh, Sonnet, can you go ahead and take the Anthropic search readme and merge it into our base level readme? I don't want that duplicate readme. Please go ahead, clean that up, and then delete the Anthropic search readme when you finish. This is going to happen automatically in the background for us. So, that's great. While you're working with these assistants, you know, and doing scary things like merging and deleting, I always like to uh just commit whatever code I have. Um, you know, I highly recommend you get this on some type of short alias. I'm just going to run that there. Why is this important? Why is having a incredible 700 line personal AI assistant valuable? Right? Why is this valuable? It's valuable because of one simple idea. You know it. We talk about on the channel all the time. Readme is merged, the duplicate deleted, and Anthropic search docs are now fully included with sections on prerequisites, setup, usage, output format, and keynotes. Fantastic.
Okay, so sorry. Yeah, we got interrupted there, guys. Sorry. Uh, what was I saying? So, the key idea here is compute equals success. Okay, the more you scale your compute, the more success you will have as an engineer in the generative AI age. If you understand this one idea, you are going to win. You're setting yourselves up to win. So, how were we able to tap into more compute here with our personal AI assistant plugged into Claude Code? Cloud Code is a programmable agentic coding tool. And not only is it just programmable, it's infinitely programmable. So, you can do stuff like this, right? Um, you can write entire workflows, right? Here's a super simple one we've looked at, right? You have a prompt that uh creates a new branch, creates a to-do, to-do.ts, ts minimal C cli application and then it commits okay so there are tools embedded in this prompt right and and that's the big idea if you get that you'll get a lot of things there are tool calls embedded in this prompt and it unlocks all types of craziness right reusable ADWs principal a coding members know how powerful that can be we can call any tool natural language we can embed inside scripts and we can stack multiple cloud code instances we can scale our compute with cloud code in agents. You're seeing this, you know, in the flesh. You know, I I don't just talk on this channel. I don't just, you know, hype up things. I always aim to deliver real engineering value to you. We got to talk about pricing as well in a second. But the personal AI assistant is a concrete example of that. Right? We have a cloud code call here. You can see in our process message, we're setting up that cloud code command just as we have in our previous more simplistic examples. Right? So in this codebase, you know, whenever you're ready for it, we have underscore 2 um cloud code is programmable and you can see this exact simple example boiled down to its atoms. Okay, you can see there's a prompt. We're doing a few things there and then we're running cloud code as a programmable agent encoding tool.
There is some confusion in the industry right now. Some engineers are wondering what is cloud code for? How is it different? You know, is cursor better? Is better? A couple thoughts there. First, don't think in ors, think in ands. Use these tools together. Use different combinations. Don't limit yourself with the ore mindset. That will set you back. Point number two here is there is no other tool right now that is an agentic programmable tool. Okay, we've gone into detail on this in a couple previous videos. You know, we know that AI coding is a small subset of agentic coding, but the the the true impact of this we're going to be unpacking on the channel. Make sure you're subscribed. Make sure you're part of the journey because this is going to get really interesting, right? You you can't do this with Windsurf. You can't do this with Cursor. You can't do this with Klein. Codeex is the only tool uh close to accomplishing this. It has full auto mode and then you can write a prompt and then your assistant will do work for you. Okay, so this is close but codeex is is you know to be fully honest it's nothing special. It's basically a clone of cloud code. The only thing they have going for them is the fact that they enable open models. Okay, so you can use whatever model um you want, right? They have custom configuration option for uh different models, right? So, of course, everyone's running toward the new Gemini models. They just released the new Pro model. That's all fantastic. If Claude Code opens up their models, they just win this game immediately. Okay. Now, I doubt they'll do that. It's not a good strategic decision to make from a business strategy perspective. They are making so much money off this tool. And uh you know, maybe that's a good place to to go here now, right? This personal AI assistant, cloud code, this new powerful agentic tool. Um, it's not all roses, right? This then costs. You can see here um you know, what but before we get to this before we dive into this uh I just really want to I want to stop and uh I always forget to do this. I want to stop and say uh just a huge thank you. Uh we're about to hit 50k subs and we were never supposed to get this big. I imagine we would flatline around 10 or 20k subs and we would just go sideways for basically ever, right? Like I imagine that this was it. Okay, but here we are almost at 50K subs. It's been a really long drive. Every week I show up four, you know, mid and senior plus engineers working in the field with their boots on the ground every single day, right? It's about building real valuable software. Engineers are skeptical by nature. It's like a huge percentage of the audience that watches the channel, they're not subscribed. And that's fine. That's fine. Whenever you're ready, I'm here. I'll be here every single week. I don't know if you can tell yet. If you can't, this is not a scam. This is not a griff. The one product that I do sell on the channel, Principal AI Coding, has been immensely valuable to every engineer that's taken it. The reviews have been insane and it's setting up for what we're going to do next. You can already kind of feel what's coming next. If you've been watching the channel, everything we're doing is about scaling our compute so that we can be successful, right? We're talking LLMs, AI coding, agentic coding, AI agents, and whatever comes next. I'm I'm really proud to say there are some big big hitting engineers and high performers that watch this channel. We're talking about your favorite engineers favorite engineer and um you know I just want to stop and just say thank you. You know it's been a obviously a ton of hard work doing this every single week, but it is always worth it to help you know you to help other engineers. There's so much noise and and the the noise to value ratio I feel like is very low. um just across the ecosystem.
One last thing I'll say here is I have three priorities and no matter how big we get, I want you to know that you know what these priorities are because it differentiates what we do here on the Indie Devdan channel. Okay, so one build living software, two unlock your engineering potential and three make a living by creating value. Okay, the order is everything, right? Um, I I want you to know my priorities and I want to just really hint on why it's so important to have priorities yourself. I'm not here to make a living, right? That's not my first priority. That's my third priority. I'm not here to just unlock your engineering potential. Okay? My top priority is to build living software that works while I sleep. Okay? And by having this mission, right? By having that as my key cornerstone mission, I'm able to provide you with unique value that you can't find anywhere else. Because of my mission, because of my priorities, this isn't changing. It's not going to change. And I'm going to be here every single week, 50K, 60K, no matter how high or how low we go. We're going to have big videos. We're going to have crappy videos that don't do well. But in every single video, I'm going to be aiming to give you concrete value to help your engineering every single week, every single day. And you know, by the way, you know, I do focus a lot of my time and effort on senior plus engineers. But you know, for all the noobs out there, following learning, big shout out to you. Some of the content we look at is dense. Hats off to you for learning, for pushing above your weight. The best way to grow as a noob, you know, junior engineer is to spend time with other heavy hitters and, you know, listen to other engineers that know what they're doing and that, you know, have concrete takes. Anyway, that's a long rant. I just want to say thank you. We're almost at 50K. That's cool. In the end, it's just a number. What matters is what we do here on the channel, the technology we use, and the way we use it. Patterns, techniques, and most importantly, principles, right? Tools, great models, great applications, great. What matters is what we do with them. Back to Claw Code and back to the personal AI assistant. Back to scaling our compute.
So, um I just want to talk about costs. Okay. Uh we're getting deep fried here. This is my cost chart for Claude Code. I just took a quick image of this today. Um and you can see here in just 10 days I'm at $100. Okay. And I even have some off days on here, right, where I'm using other technology. So, I can guarantee you uh I'm not using enough compute. You're not using enough compute. We can get these numbers way higher. And we're going to. And you know, to be clear, it's not about spending more money. It's about spending more money and getting more value out. If you can spend $100 and get $200 worth of engineering work done, you should put in as much money as you can, right? You have a value generator, right? That's what a lot of these powerful tools are. You put 20 bucks in and 40 bucks worth of value comes out right now. Now, the trick obviously is when we're comparing our compute advantage across all these tools, the real question is, can you put a 100 bucks in and get 300 bucks out using a different tool? Okay. And then then and that's what the compute advantage equation is really all about. And that's what we're going to be exploring on this channel over time. Right now, there just is not a tool more important than Cloud Code. They're shipping the highest amount of value and the ROI you can get here. The exchange of time plus cost is actually quite low if you really really understand the value of your time. You know, on one hand, yes, very expensive. On the other hand, something is only expensive uh relative to what it gives you. You can see yesterday almost 50 bucks. Pretty gnarly. So, I love Cloud Code. I'm going to continue using it. I'm definitely not saying don't use this. You know, they just launched their Max plan. And I've been scratching my head on this one a little bit, as you may have as well. I can't really tell if this is actually going to save money or not. As you can see, I'm like a prime candidate for something like this. I'm already at 100. The big problem with a lot of AI labs right now that are providing compute is that there are still limits. Okay? Like it doesn't matter how much you're paying. Um you can see there's still limits. I don't know if this will actually be helpful. I'm going to wait to hear a little bit more before jumping in on this membership. I am not a clawed user. I think that if you're a clawed user, this subscription is a lot more interesting and enticing because it is unified. They say that somewhere here, right? This is one unified subscription. So, if you use Claude Desktop or, you know, web, I think this is a no-brainer. Like, pick this up, pick up their what was it, the $100 a month, you know, you're going to get some good usage out of this. Where it stops being useful is if you're not hitting your usage limits, right? And and that's where the math just it isn't really clear to me. Um, I wish that they would just come out and say, "Hey, this is going to be cheaper if you use the entire membership." This is going to be cheaper than paying via API key. um as as I am here, right? This got mentioned on our previous video. So, I just wanted to bring this up. This is available here. I think the giveaway here is if you use web and desktop and you use cloud code, this is a no-brainer.
So, something to mention here is, you know, we were one of the first channels to really start talking about cloud code. But you can see here inside of Hacker News, we might need to go to the second page now. You can see uh, you know, cloud code is getting a lot more attention. 246 comments on Hacker News for this post. Exactly. To me, this is a really great sign. Uh the industry is catching up to how important this tool is. Um even though there are still, you know, some questions. Everyone always, you know, mentions cursor when any AI coding tool is brought up as a comparison. As you've seen on the channel, there is no comparison to cloud code right now. Even when you exclude the fact that this is a programmable agent coding tool, I don't think other tools are close. And you know, I say that by the way, look at the bottom right corner here, right? Cursor tab. I use cursor. I pay for cursor. Um, I'm a huge fan of cursor. I just never use their uh agent capabilities because cloud codes is frankly better. Okay. And we can dig into what better means uh on the channel as we go on. But it's great to see this. The industry is catching on the the, you know, serious impact of cloud code. And after I finish this video up, I'm going to dive head first into podcast that was just released with two of the creators of Cloud Code. So, I'm really excited to check this out. I'll leave those links in the description for you as well to check this out. I already know this is going to be a banger video and they're going to be some really important ideas. I think you know big takeaway here. Uh cloud code isn't going anywhere. This is the best leading agentic coding tool. Um as you saw here we were able to build out a very very powerful voice to cloud code personal AI assistant in just 700 lines of code.
So what are the components of this? Personal AI assistants at their core are simple. You have the ears, you have the brain and you have the voice. So the ears of our application is something called the real time speech to text. So this is a great library. You saw we were using it here and this gets our input from speech into text so that we can run it into our brain. And so for our brain we are of course using claw code. This is what actually does all the work. This is what does the heavy lifting. This is what does the responding and this is how we get work done in natural language using our voice. Okay. Lastly, we have our voice. Right? So, our voice is fueled by OpenAI. And you can see here it's a simple request. You pass in your text. It gives you an output file. You play the output file. The only kind of important note here is that we're using compressed speech. Sometimes cloud code will return a bunch of details. So, we're using gbt 4.1 mini to just compress whatever was returned. We have this clean, concise compression prompt. Um, this is going to be here again in the codebase for you to check out. But basically, we just want to compress and make things really concise and put it into a natural language, you know, conversational response. And this is a great use for a smaller, faster model. You know, we can save some time. So, I'm throwing GPT4.1 Mini at this problem and having great success. Whenever you're thinking about building out your own personal AI assistant, this is how you can do it. I think this is going to be a really powerful tool. Not many engineers are going to unlock this. I want to give you valuable ideas like this so that you can tap into it if you so choose. And that's the three components, right? Ears, the brain, and the voice. This codebase is going to be available to you. Link's going to be in the description. This is inside of the cloud code is programmable codebase. And for anyone interested, if you want to accelerate your engineering, check out principled AI coding. This is the best way to learn AI coding in a principled way, giving you foundational skills to stay relevant with AI coding tools of today and tomorrow. This is an essential theme. We talk about principles, not tools. Principles not models. As many of you know, the next level course is in the works. I'm working on the course that comes after this. This is going to be a brand new course that builds on top of everything we've done here. Stay tuned for that. We're going to be using powerful next generation agentic coding to take everything we've done to the next level. We're going to multiply everything we've done. Enough about this link in the description as usual. You can also get voice to claude link in the description. There's a ton of amazing real time uh speech to text functionality to you know modify. There are many ways to make this faster to make it more accurate or you can trade off speed uh for performance and vice versa. I'll also leave a link to real-time speech to text. I'll also just go ahead and commit the anthropic search. Our assistant built this out for us. I basically talked to my computer and had it built out, you know, 230 lines of code, which is fantastic, really cool. As discussed, there are pros and cons to using natural language and speaking aloud, but I think it's another way we can scale our compute to scale our engineering impact. You know where to find me every single Monday. Stay focused and keep building.