Anticodeguy

<written by a human being>

The architecture of an application being developed for desktop depends on the key functions of this application. For example, my application will use ML functions. And for ML there are different programming languages that are tailored for it to a greater or lesser extent. For example, everyone knows that Python is very well suited for ML, so if the main functions of the application involve working with ML specifically, then it's logical to choose Python libraries and compilers for the main language, which will then assemble it into the final program.

There are only 7 musical notes in existence. But those 7 notes have created millions of unique songs Read more about how to Red Pill Your Career: From Replaceable Employee To Irreplaceable Creator

<written by a human being>

A whole new brave world opens up when you start working with desktop software. Here are programming languages that can directly work with PC memory, with the processor, a variety of compilers. There are also databases here and, naturally, an interface part which, surprisingly enough, is very connected to web, because in modern desktop applications the front - that is, the external visual shell - can be written in TypeScript or using JS libraries. But of course, there are also a bunch of native frameworks to choose from.

<written by a human being>

Choosing a technology stack when writing a desktop application is no less important than when developing web applications. But since a desktop application runs in an operating system, first and foremost you need to think about who will be using this application, on which operating system it will be installed. This also imposes limitations on the technology stack that will be used. However, here, as in web and mobile applications, there are frameworks that allow you, using a single codebase, to compile the final application for different operating systems. For example, the one I'm using can be compiled for both Windows and macOS. Since my application runs on Python, PySide6 was chosen as the UI framework, and Nuitka as the compiler.

The only way to avoid being replaced by AI is to become someone who can't be replaced Read more about how to Red Pill Your Career: From Replaceable Employee To Irreplaceable Creator

When we talk about vibe code, we all by default imagine a web application, a SaaS website, or a mobile app. I thought, what if we use vibe coding to make a full-fledged desktop application? Why doesn't anyone write about this - has everyone really stopped using good old software that you install on your PC and use peacefully? This is exactly what I've started doing - a completely offline application, fully autonomous, not requiring internet or a monthly subscription, working on Windows or macOS. Let's see how vibe coding handles this. I've never seriously dove into the world of desktop software. At university we learned old Delphi, on which we wrote small desktop applications with databases and other delights. But of course, this never went beyond academic projects. Since then, I've been completely absorbed for more than a decade in web development - various frameworks, network technologies, data transfer protocols, and so on. Desktop software is a whole new world that has suddenly opened up to me.

When will crypto go up again? The next day after you sell all your holdings.

<written by a human being>

The final episode of my course on creating content with artificial intelligence came out on YouTube yesterday. This is a series of short video tutorials where I show how various tools can be used when creating content, including writing articles, transcribing notes, creating reader avatars, your authorial voice, formatting content for different social networks, and many other interesting things. All of this is also accompanied by written lessons in my newsletter, links to which are also in the video description. And in the text lessons, there are already all the prompts that can be copied, detailed instructions, and a detailed description of how to use all of this. There are still several more lessons ahead that will come out specifically in newsletter format, because there were initially more of them than the videos themselves. There are lessons that are without video, text-only, so all the videos came out faster. But if you want the complete set, then head to the newsletter.

<written by a human being>

In November of last year I paid for a Cursor subscription for the first time, and since then my timid steps into getting to know what has become so popular today - vibe coding - began. And since then I’ve already changed my approach to development several times: right now I mostly use Claude Code, but that’s not what this is about. What real things have I managed to do since then, things that are actually the result of my work with AI tools? The first thing I did using vibe coding was writing scripts that allow me to transcribe video and audio notes that I regularly record, and from which I then create text content. They make it possible to recognize speech in recordings and even separate it by participants if it’s a dialogue. That is, for example, when I record a meeting, afterward I immediately get a transcript of the dialogue, which I can then place into a knowledge base and extract from it later when needed. Next, I rebuilt the studio’s website and stopped paying for a site builder and for hosting, because it turned out to be essentially unnecessary. I picked a technical stack that lets me do this for free and on a modern framework. Moreover, I found a way - which I was embarrassingly unaware of before - to host websites for free, which is exactly what I did. And on top of everything else, the site also scores a 10 out of 10 in Lighthouse performance tests. There's a lot more, but I know you don't like a ton of text, so let's stop here for now.

<written by a human being>

In November of last year I paid for a Cursor subscription for the first time, and since then my timid steps into getting to know what has become so popular today - vibe coding - began. And since then I’ve already changed my approach to development several times: right now I mostly use Claude Code, but that’s not what this is about. What real things have I managed to do since then, things that are actually the result of my work with AI tools? The first thing I did using vibe coding was writing scripts that allow me to transcribe video and audio notes that I regularly record, and from which I then create text content. They make it possible to recognize speech in recordings and even separate it by participants if it’s a dialogue. That is, for example, when I record a meeting, afterward I immediately get a transcript of the dialogue, which I can then place into a knowledge base and extract from it later when needed. Next, I rebuilt the studio’s website and stopped paying for a site builder and for hosting, because it turned out to be essentially unnecessary. I picked a technical stack that lets me do this for free and on a modern framework. Moreover, I found a way - which I was embarrassingly unaware of before - to host websites for free, which is exactly what I did. And on top of everything else, the site also scores a 10 out of 10 in Lighthouse performance tests. I run streams of my work sessions on several channels, and on weekends I stream gaming sessions in order to improve my spoken English. Broadcasting video streams to multiple channels at the same time puts quite a heavy load on my work machine, and to avoid that, the solution has of course long been invented - restream services, where you send one signal and they distribute it to multiple channels. This way you offload your machine and load the servers of the restreaming service, which of course has to pay for that, so there are no free services like this. I thought, what if I can build my own restreaming service based on a file server, which, by the way, I also set up with the help of AI. But that wasn’t exactly vibe coding, because at that time I was still doing everything myself. And ChatGPT was simply prompting me on what actions to take to get the desired result. In the end, it turned out that setting up your own restreaming server is a fairly trivial task. And with the help of vibe coding we wrote what was needed. We tested everything, configured everything. And now I have my own restreaming service and I don’t have to pay Restream to solve this task. I deployed a free project management tool on my own server, which will now replace ClickUp, which I had also been paying for before. That is, now before paying for some SaaS, first of all I will always ask myself whether I can just develop an analogous service. Because in most cases I don’t need 100500 features of that SaaS - in reality only the basic core functions are used. In this case, there’s most likely an opportunity to just quickly develop such a service, deploy it for yourself, and calmly use it without any limitations. And if needed, improve it later again through subsequent vibe coding sessions. So the project management tool I also replaced with a free one. Next, I deployed a knowledge base for my client’s business, which is designed to be read first and foremost by artificial intelligence, not even by a human. For a human it’s also readable, because in the end a nice documentation site is formed, which you can of course read in a human way. But the key point is that at its core it is easily scanned and read specifically by AI agents, which makes it possible to turn it into something like a RAG system. This is, of course, not a full-fledged RAG, but nevertheless. In my diary there are still many notes for the future that I plan to implement using vibe coding. And the most important thing is that I understand how feasible all of this is, because earlier for all these tasks I would have needed either years, without exaggeration, or to lay out decent sums of money in order t

I like working with AI agents. It’s like having an employee with infinitely strong nerves, who never refuses to work. And who accepts all your requests and revisions unquestioningly. After another prompt, you sit there, watching how it tries to solve your task. If something doesn’t work out, you sometimes have to interrupt it or, when it starts drifting in the wrong direction, guide it, explain things again, give hints. And there’s always this interactive element. Of course, it’s great when everything works on the first try and you don’t even need to intervene, but often the right context is missing, or it gets lost during the session, so you have to stop it. But nevertheless, it brings real pleasure.

My client's information system operates through a reverse proxy. But some users are encountering errors when loading certain content elements. That is, they simply don't display. After the classic investigation of the possible cause through the browser developer console, where you can see that requests are failing and content simply isn't getting through, an assumption was made that some requests for certain content - which works not through static code but is generated dynamically from JavaScript - aren't passing through the reverse proxy, which means it's necessary to update its configuration to allow all requests through it. I described the task point by point - what isn't working and for what presumed reason, provided the initial conditions, collected a series of screenshots with errors from the console in which the request and its parameters are clearly visible, and also provided SSH access to the server where the reverse proxy is located. I sent all this to Claude Code, it analyzed all the materials, connected to the server and looked at its settings, proposed how the configuration could be updated. After receiving my approval, it did this, and after literally two iterations with verification on my part, everything worked. A task that previously would have required paying an expensive specialist - a network administrator or DevOps engineer - many hours of time, possibly even several days, took me literally one hour and was included in my regular Claude subscription.

A few days after I asked the question about how it's more profitable (cheaper) to use AI, I got an answer to it. It turns out that paying for a subscription to an application, specifically Anthropic Claude, is much more profitable than buying API capacity, because a lot is included, including dialogue cache reads, which are essentially free. Here's an article with detailed calculations https://she-llac.com/claude-limits on how this works. It turns out that with a $100 subscription, we're buying the equivalent of $1,354 in API cost! And thus we understand that for vibe coding, it's much more profitable to use the application subscription, which is much more generous, rather than buying an on-demand API subscription where you pay per volume. The calculations were done only for Claude, but it would be interesting to find out for other models as well - for example, whether this math applies to ChatGPT or not. Waiting for the continuation.

By the way, my YouTube channel have auto-dubbing function and you can watch my videos not only with Russian subtitles, but but also with Russian voice-over! It sounds a bit creepy, but still watchable.

I'm sure open-source software is getting its second wind nowadays. Why? Because there's no point in not using it at scale: you just ask AI what free solutions you could use for a certain task, it gives you options. All you have to do is choose one that suits your preferences. Heck, you can even evaluate these options using your own criteria automatically with the help of AI, which will gather all the information for you. And then the fun part: you can ask an AI agent to deploy everything for you, locally, or on production hosting. And it's very fun to watch how it's going around, making mistakes, fixing them, asking for your help, receiving your feedback and getting back to work again without hesitation. And a couple of hours (or days) later you get your final product, without paying anyone except your AI employee. Now back to the open-source solutions. Recently I discovered so many of them I didn't even know existed! All because of asking myself, can I have the same tool on my own server for free? And more often than not the answer is yes.

We live in an amazing time when you don't have to pay for what was previously impossible to imagine for free. For example, I conduct streams and broadcast them simultaneously to several channels. To do this, naturally, you need a powerful computer. My computer's capacity is enough to provide broadcasting simultaneously to three channels. However, if I'm simultaneously editing video with this, the process begins to noticeably slow down, since the graphics card needs to process three video streams for streaming, two video streams for both my monitors, and another one that's rendering video in the editor. This is quite a heavy load that can be reduced, and a solution for this has long been invented. There are services, such as, for example, Restream, which give you one channel to which you send your video stream, and they already restream it to the channels you need—to YouTube, to X, to Twitch, and so on. But here's the problem. These services are paid, because obviously they require server capacity that costs the company money, and it's clear what they earn money from. And since I'm not yet doing streaming professionally and regularly, it doesn't make sense for me to allocate my personal budget to this yet. Moreover, we live in a time when there's an opportunity to make such a service yourself. Of course, with the help of artificial intelligence. And very conveniently here came my file server, which I assembled using components from my old computers that I gradually upgraded. I thought, what if I could set up my own service that would restream one stream to different channels based on my home file server. I think its capacity is quite sufficient for this. And as it turned out, indeed, I can do this. After consulting with Claude, we developed a plan and architecture for what this would look like. Then I gave it access to my local file server. And in a few hours with debugging and testing (some may have noticed test streams on my channels) we did it. Now I have my own restreaming service that only consumes electricity, which was being consumed anyway because I use this server every day regardless. It's interesting what else can be implemented completely on your own side without additional costs. What a wonderful time to be alive.

How can you work more optimally with AI coding agents? There are three payment methods for their work. The first is purchasing directly from the service provider itself. That is, for example, we pay Claude a monthly subscription, and we're given a certain quota for Claude Code. We can do this directly in the Claude application interface itself or connect it in an IDE. The second method is paying specialized IDEs for working with coding, for example, Cursor or Copilot, which already includes some quota for the same Anthropic models, and they independently track their consumption. This is the second method. And the third is purchasing, or more precisely, paying for the consumed capacity of one model or another, when we pay for API access - and here everything is quite simple: how much we spent is how much we paid. And this API can already be connected to your IDE application. For now, I'm using a combination of the first and second. That is, I pay Claude as an application and pay Cursor as an IDE and use them together: first Claude Code directly as a plugin in Cursor, and then, when I run out of limits for the current time period, I switch to Cursor's own quota. At the same time, I'm working on the same project in one environment, which is quite convenient. And I like Claude Code's limits because they reset every few hours, and here you don't need to wait for the next day, and you can't spend your entire monthly subscription in one day, as you can do in Cursor itself. Which of these methods or combinations of these methods is most optimal in terms of expenses, which one do you use? Share in the comments.

How can you work more optimally with AI coding agents? There are three payment methods for their work. The first is purchasing directly from the service provider itself. That is, for example, we pay Claude a monthly subscription, and we're given a certain quota for Claude Code. We can do this directly in the Claude application interface itself or connect it in an IDE. The second method is paying specialized IDEs for working with coding, for example, Cursor or Copilot, which already includes some quota for the same Anthropic models, and they independently track their consumption. This is the second method. And the third is purchasing, or more precisely, paying for the consumed capacity of one model or another, when we pay for API access - and here everything is quite simple: how much we spent is how much we paid. And this API can already be connected to your IDE application. For now, I'm using a combination of the first and second. That is, I pay Claude as an application and pay Cursor as an IDE and use them together: first Claude Code directly as a plugin in Cursor, and then, when I run out of limits for the current time period, I switch to Cursor's own quota. At the same time, I'm working on the same project in one environment, which is quite convenient. And I like Claude Code's limits because they reset every few hours, and here you don't need to wait for the next day, and you can't spend your entire monthly subscription in one day, as you can do in Cursor itself. Which of these methods or combinations of these methods is most optimal in terms of expenses, which one do you use? Share in the comments.

Did you know how to make Claude AI way smarter? The prompt is here: https://anticodeguy.substack.com/p/how-to-set-up-system-prompts-for?r=1m5hbt

Now you don't need to write prompts to artificial intelligence. If you've read my article about how to provide context to artificial intelligence and about prompt engineering, then you know how to do this most effectively. But not long ago, Claude released a new feature called Skills, and it allows you to create skills. A skill is essentially a predefined set of instructions that will be applied in one context or another. For example, I have a skill for editing legal documents, a skill for help with system architecture, or a skill for creating content in a specific format. Our prompt engineering, which we've just learned, helps tremendously with creating these skills. Creating skills is very simple: go to settings, the Capabilities tab, scroll down to the Skills section, click the add button, and here choose the simplest option - create with Claude's help. And then simply use that same prompt that we usually insert for a particular task. We ask Claude to create a skill based on this prompt and clarify if something needs to be added, changed, or if additional nuances need to be considered. After Claude thinks about this task and composes the corresponding skill, you'll just need to copy it and it will appear in those same settings. Now when creating a new chat, you'll simply need to provide context and Claude will automatically pull in this skill. If this doesn't happen - and from my observations, this is quite a common occurrence - just copy the skill name and explicitly ask in the prompt to use this particular skill. I think they'll fix this glitch over time. How do you determine that Claude is using a skill in its work? A line will appear in the chat that will indicate that Claude is currently studying this skill. In this case, you can be confident that the AI is following the given instructions. By the way, you can find the article I mentioned previously, in my newsletter.