Transcription
Hello everyone and welcome to the vibe coding video. Today I will share my experience. I have quite a lot of it. And based on what I generally decided to do this broadcast with, yes, it's quite simple. I hear everywhere that vibe coding is like the internet 2.0. And at the moment of its inception, when the internet was just beginning, whoever mastered it, and the better and higher quality they did it, the more they won. It's enough to recall such an interesting example. When SEO optimization was just happening, at least in Russia, I know cases where people knew that if you take keywords, hide them on the site, but make them invisible so that they don't interfere with the UI or spoil it, then by doing so, you significantly move up in the search results for practically 0 rubles. And by doing so, they, uh, accepted clients, improved their optimization, so-called, and by doing so, people occupied the first lines in their niche. They made very good money on this, but then the sites went into a ban after about 3-4 months. But the people, those who promoted with this method, and those whose sites, rather, were promoted, they earned a lot. Well, here's such a nuance and example. And why, as I think, it's still worth investing your time not only in vibe coding, but in general in learning, in everything related to this, i.e., agents, LLMs, vibe, but first of all, let's understand who this video is for, who it will suit. First of all, it's for developers, to speed up routine tasks, like some boilerplate, documentation, prototyping, and so on. Product managers and designers, who, by the way, don't know how to code, don't know how to write, they might now be able to implement something, yes, accordingly, entrepreneurs and startup founders, yes, i.e., MVP without a team, yes, and not just MVP, but even a bit further than MVP, but we will talk about this today too. Uh, meaning they can do it without a budget, without large teams. So, testers are not needed, designers are not needed, well, not in the way they were before, yes, so designers are not needed, meaning now everything, uh, uh, agents will be able to implement, yes, accordingly, freelancers, curious students. Well, in general, I would, honestly, if I were asked on the street, I don't know, maybe every second passerby, I would say that at least everyone should try to understand in what endeavors and in what matters it can be useful. Personally, I have chosen for myself that if I can delegate some routine task and get a quality result that is acceptable to me, then why not try, at least, yes? So, accordingly, who it won't suit, if you're looking for a magic button, yes, meaning that AI will do everything itself. This is it. Who, if they have heard of and used, uh, any dev tools, where you literally accept everything, the same cursor, yes, you just detach, detach, detach, and you're amazed, and it writes, writes, and writes, yes? So, accordingly, not ready to practice, to watch little, you need to open the editor, you think, will they replace developers, or do you want perfect code from the first try, not ready to check the result. Well, essentially, in general, as I said, yes, it's still worth trying rather than not trying. Also, yes, what will we talk about? If you want to know ahead of time, I will attach the timestamps after the end of this broadcast. Accordingly, look in the timestamps for the moments that interest you. But before we start, a small disclaimer, it's important to understand that this is my personal experience. This is not some kind of scam. It's actually more than 400 hours, I don't know, maybe 500, I just, uh, stopped counting at some point. I counted 400 and that was it. Yes, accordingly, this is not the ultimate truth. I may be mistaken somewhere, I may interpret something not quite correctly, so I will be glad if someone corrects me in the comments, for example. So, yes, another very important nuance to always keep in mind is that AIs change very quickly. Today one is an underdog, in six months it will be a winner. Accordingly, about the video, if by chance, I wouldn't really recommend it, but if by chance you absorb it, you can listen at 1.5-2x speed. But in general, I wouldn't recommend it. Why? Because I will be saying certain things based on my experience, and by doing so, it will just fly in one ear and out the other in the background. Well, in principle, something might stick, but in general, I wouldn't recommend it because, well, a lot of useful information can simply be lost. About responsibility, yes? So, don't copy one-to-one, but this is more about the second video. Yes, this is the first video, and there will be a second one, just so you know. About practice, watching little, actually. It's worth noting, taking notes, pausing. Well, not now during the live broadcast, of course, but in general, so as not to lose it later and study it. Literally, you open some Telegram, go to favorites, write that, ah, I need to research what a design system is, for example. Materials, all links will be in the description, some are already there, the presentation will also be there. And feedback via direct message is always welcome. You can write in the comments what you want, what desires you have, and so on. And of course, channel support. So, I'll be glad if you give a like now, or if you're watching the recording. No need for the bell. In principle, subscriptions will be enough. This also motivates me to make such videos, to try to describe them in such detail and show them. And accordingly, I have created a special Telegram channel, here it is on the screen, Code on Vibes. And it's also in the pinned messages, oh damn, in the description, I've attached it. There I will generally write about AI, share thoughts, because such videos are made when a certain amount of material has accumulated and a video can be released. But here it will be in the form of short notes, also based on my experience, what works there, ideas, not ideas. The next post, for example, will be about my stack, what I chose, why I chose certain things, and what helped me not step on rakes in later versions. So, this is what I will talk about based on my experience. The mini-app, it's the third version. Before that, there were two versions. Two versions went into oblivion because one broke, and it became very bad. But the stack didn't change, so, well, for those who are interested, yes, in general about AI, I recommend it. Accordingly, let's begin. Vibe coding, yes, 400 hours. I started learning when ChatGPT from OpenAI just came out, but I managed to try a lot. I spent a total of 400 hours, not on complete learning, but only on creating a Telegram mini-app. This is an application that helps find jobs, compare salaries with the market, and even prepares for interviews. Under the hood, there's an LLM that learns, plus original high-quality questions. And I didn't write, well, almost, yes, I'm exaggerating, I didn't write the code by hand. But in general, 90% of it was done by the AI. This is how it looks. There's no point in showing it. It's quite large in the description. You can play with it yourself, so to speak. But in general, how it looks. Here's such an interesting UI that expands. So, the mini-app opens, whether on a phone, MacBook, or tablet. You can stretch it in width and length in various ways. Here's how the UI looks. Here's how interesting it can be. This is how the questions look. There's a timer, bonus points, uh, accordingly, some keywords, four answers, they change. So, the stack is very interesting, actually, and some things under the hood. Here are the vacancies themselves, yes, so here's their list. Accordingly, there's already a certain salary range that is compared with the market. This is what is not parsed from the API, but what I set up myself manually so that the candidate seems interesting enough to me, because many job boards hide the benefits for the candidate, yes, meaning they don't tell them that, listen, this vacancy is beneficial, yes, because the employer will say: "Listen, how will we hire someone then?" Yes? And, accordingly, there is statistics here, actually, 150 active users, don't confuse it with users in general, yes, an active user is someone who has gone through these questions from start to finish more than three times, yes. Accordingly, there are already people, they are using it. I'm glad that it's beneficial for them, that they get benefits from it, at least they prepare, at most they find vacancies there and so on. Yes. Accordingly, further, according to the statistics here, this all personalizes over time. The more interviews a person goes through, well, by answering questions, the better the statistics are shown to them and more detailed, yes. Here are such insights and achievements. I've only shown you a part of it. Actually, there's more and more, even if you're not an iOS developer, you can still open it, see how I implemented it, get some ideas for yourself, why not. Yes, accordingly, I have demonstrated what I implemented all this on. You can try it yourself, but a little about me, yes, so, yes, my path was two versions, two of them broke, there will be about them too, but not in this video. But about me, yes, I am a mobile developer, mainly for iOS, I had experience with Android. 7 years in programming. From the first days, when OpenAI released it, I started studying this whole topic. Yes, this is how my GitHub looks. If anything, you can find me by the nickname 07. Yes, accordingly, I am currently working on a Telegram mini-app project. Yes, I've already explained how it helps, but let's move on. What is vibe coding? Yes, so we need to start from the basics. Here I will interpret Andrey Karpatov. If anyone has seen his two interesting videos, I will also attach them, then. Hello. For those who haven't seen them, I recommend watching them. Why? Because in the second video, where he talks about prompt engineering, about what it looks like, how to communicate, what it looks like under the hood, how to communicate correctly, what to share in a prompt, what to consider. So, very technical and detailed things. I definitely recommend watching it if you haven't, yes, but in general, vibe coding is interpreted differently. In general, as he said, surrender to the vibrations, accept exponential coefficients, and forget the existence of code. But in reality, it's just sitting there, pressing Enter, accepting what the machine does for you, and enjoying the result, yes? So, if before we wrote code, now we manage an AI team, yes? Not just one, actually, but a whole team. This will also be revealed, why and how. Which, as it happens, do everything for us, test, program, design, yes, yes, don't be surprised. And accordingly, what do we have on the market now, what is worth being interested in. Well, actually, without beating around the bush, it's Claude. This is my choice. In general, it has a free version, but the main cost is $20-200. At one point, I had two Claude accounts. By the way, in the second video, it's a similar case, if I'm not mistaken, SEO Chernov or whoever he is, Anthropic's hundred, he still does that. But at one point, I just connected two and worked with two $200 accounts. In general, what feedback can I share, that $200 suits me better, because it works for me non-stop. Here, I would probably even say that those who will work non-stop, the $200 option will likely suit you. But don't rush to buy it immediately. Try it for free. Many models are available for free. The same GPT, GRK, Gini, and so on. Deepseek is completely free. Yes. But in general, $20 is not quite enough, around $100 and $200 is enough. What else is an advantage of $200, that, uh, the first some, uh, the first some innovations with Claude, they first come to us, who are on the $200 plan, and then to other users. But in general, what's here? It's skills, yes, we'll talk about them in the second video. MCP is fast, yes, so very good. High quality, I would also note. Well, in general, excellent code and context. Accordingly, limits. Well, since it's not consumed by rate, yes, meaning how many tokens we output, that's how much is debited, since it's a subscription, there are updates, there are updates every 6 hours during the day and there are updates every week. So keep this in mind in the settings, you can see everything there too. Yes, by the way, I'll add that I won't be very detailed in some moments, why so? Because this video would last, I don't know, two hours for just one part, yes? But therefore, I will just voice some ideas, but you can ask the same ChatGPT or Claude, even for yourself, yes, for example, not to go into a search, as if it's a greeting, yes, and it will answer. Accordingly, ChatGPT from OpenAI, in general, the experience is acceptable, prices are also similar. What I would like to say here is that, uh, here, our code-savvy model, but there are also nuances here that ChatGPT has in-depth study available. What experience do I have with this? Quite negative. Why? Because in-depth study implies that it's given, say, nine per month, if I'm not mistaken, or how much it was given to me before, I don't know what is given now, because I don't use it for these purposes. And in-depth study means that it will take about 30-40 minutes, well, a thinking mode, yes, it will study for about 30-40 minutes, sometimes longer, sometimes a little faster. And after it studies, it will give a fantastic result. I had a video that I jokingly called Venom. I uploaded it there and wrote a fairly simple prompt, saying: "Listen, I have a video, can you transcribe it into text?" It said: "Yes, okay, I can." Accordingly, it thought for 50 minutes. What it gave me at the end, after 50 minutes, it implemented it quite interestingly. It simply wrote me a description of the movie Venom and Venom 2. It took the title, thought it was the title of a movie. It didn't even... This, by the way, was in September 2025 for context. Yes, accordingly, not so long ago, not so long ago, in fact, and this is the kind of experience I had. If anyone else has... So, are there any lags? Wait, is everything okay now or what? Let's move on then, with your permission. Sorry if anything, there's this. Uh, if anything, I'm now checking from my phone, it seems like it's not reaching you. Uh, no, I'll open the chat now, if anything, just write in the chat. I'll open it on my phone now. If there's any fuss again, tag me or send me a direct message. Yes. Well, that's it, let's go. Sorry for what happened. Let's continue. Yes, if anything, yes, write, if there are any problems. I'll try to look at it again after this. Honestly, I don't know, I already, uh, looked, damn it, to be better prepared. Well, okay, let's go then. Uh, so, yes, so this is such an interesting moment I had with ChatGPT. Yes. But in general, it's not quite suitable for coding. Here, jumping ahead, what I would like to say, I had such an experiment that I, uh, wrote prompts through ChatGPT and gave them to Claude. Claude executed, gave feedback to ChatGPT, then again, so there was this idea on the internet that if one AI, in particular, in this case, ChatGPT, writes prompts, then it will supposedly do everything better. In reality, everything was worse for me, meaning, on the contrary, everything was longer for me, and the code quality started to drop. And in general, I, meaning, it was obvious to the eye that, no, sorry, this combination works poorly, it's much better to just tell Claude everything to do it. What can I say about Gemini? It's quite cheap, 3,000 a year, if I'm not mistaken. Yes, accordingly, it's also very, uh, lately it's been taking up a lot of my attention and everything like that. Uh, accordingly, yes, meaning it's becoming more and more powerful. What through it? It's quite interesting for me to analyze information, to think through any, uh, there, plans, for it to conduct some analytics, and all that sort of thing. Yes, accordingly, yes, meaning Claude is the main one. I will have MCP skills in the second video. This is a game-changer, this is on steroids, on boosts, and on everything. Uh, meaning it improves in every way. It doesn't matter if it's, for example, code or cursor and some LLM under the hood. But MCP skills, jumping ahead, will be a real must-have. We'll talk about them in more detail in the second video, if anything. Well. And here, what I would like to, uh, uh, fix here. We have no favorites, we have winners. Loyalty, efficiency, not brand or name. So, remember about improvement. Today one is on top, tomorrow another. We all remember about hype. Well, and where is Deepseek now, yes? There is also Grok. I was not satisfied with the result. When the third version just came out, I tried it back in June. I wrote an application, asked it to make a dark and light theme. It did, but on the dark theme, part of the UI looked bad. Accordingly, I asked it to adapt it. In the end, with 10 prompts, it couldn't. I don't know what it was related to, but that's how it was. Yes. Accordingly, yes, so regarding Grok, Deepseek, uh, sorry, yes, Deepseek, Syn Mod, and so on. With Deepseek too, when there was hype, I tried it a little. Yes, it's free, yes, it's cheap, yes, it even performed some tasks then, but in general, I didn't quite like it at that particular time. I haven't re-examined it now, but I want to re-examine all this in the near future. And joke number one: vibe coding is when you do in an hour what you used to do in a week, and then spend a week fixing what you broke in that hour. Hope it's funny. Well, here's the terminology too. If someone is familiar, then great, but some guys might not be familiar or might forget, so let's quickly go over it. Accordingly, AI can do everything for us, but if we don't understand some terms, then we don't control the process. And, accordingly, we stumble around like a blind kitten. And like in my second version, when the AI squashed the commits, meaning I had changes, there were a lot of them, about 200, I mean commits. And then it squashed these commits into one result, in which it broke the UI. In the end, I tried to fix it for three evenings. This UI, and the codebase based on it, on which the UI depends. I couldn't. That's how I got version 3.0. But, but you need to know and understand the terminology quite simply, that, well, now we'll get to commits, some GitHubs, and so on. The first thing I'd like to share is that it's worth writing down. Well, before that, if there were any valuable thoughts, of course, write them down. But this is a must, it's Visper Flow, it's an application, it's available for various MacBooks, phones, iOS. For Android, by the way, I don't know if it exists, but what is it? You open the program, press some key. When you press the key, you start dictating, accordingly, what it is. Meaning, you could write prompts with your fingers, if, God willing, you have them. And, accordingly, you write something, and you think: "Damn, there's so much to write, I'm lazy." But with Visper, you can open this program, dictate, and, accordingly, it all gets transcribed into text. But what's worth remembering here is that when you dictate to it, then you release the key, your voice is converted to text. And if you suddenly switch windows, meaning you were typing, let's say, in the Claude program, you get a message in Telegram, you open Telegram, and it will be inserted there. Accordingly, keep this in mind when you work with Visper Flow. Yes, there's a free version of 2,000 words per week. If you share your link with friends, well, make such a referral scheme, then you'll get more. Therefore, if by chance you have a friend, comrade, or acquaintance who wants to use it too, then it's better to exchange referral links like this to increase your word count. But in general, even 2,000 words is enough for me personally. Yes. And here, I hope it's also clear. Yes. By the way, after the first block, I said, are there any questions, clarifications, guys. Accordingly, if anything, write in Telegram, oh, in Telegram, sorry, in this chat about what was said, of course. Meaning, don't jump ahead. Well, even if you jump ahead, don't overdo it. I'll just say, well, we'll explain this in detail now, either at the end or in the second video. Yes. If anything, yes, remember about the lags. If anything, just write in direct messages, not in some chats, because I've opened it and I'm keeping an eye on it. I'll monitor it now. Everything. Say the quality, the seeds, the hair is visible. Well, that's it. I hope it continues like this. Well, okay, since there are no questions yet, let's move on. Uh, accordingly, how is the quality, how is the sound, please give a plus if everything is good, feedback is received, because who knows, maybe it froze. I'm monitoring the chat, if anything. This is how it is. It's not lagging. Great, it's not lagging. Voice plus-plus. Here come the pluses. Thank you very much. So, let's move on. Uh, what's next? It's tool infrastructure. Let's go through it quickly, briefly. Git, GitHub is a must-have. You need to know this, study it, if you don't know. Accordingly, GitHub is like some kind of cloud in cyberspace where our code is stored. It will be stored there, we will modify it there, possibly delete it, possibly add it, God willing. Accordingly, what's here? Commit, push, pull. More than before, we don't do this manually. Now we automate everything. But AI can delete a file or push. AI can break history, as happened to me in the second version. Although, it would seem, 7 years, but not because I messed up or was careless, but because it simply did it itself at some point. Here's a quite interesting moment that I encountered. Based on Claude, by the way, remember everything I'm telling you now. If something is different, I will emphasize it additionally. Claude at one point tried to push, but encountered some problem, I honestly don't remember, the accelerator failed, or there were lags like now. But it got stuck like this. It tried to push in a loop, couldn't push. In the end, for some reason, I swear, I used to hear that, yes, my AI itself took and deleted the project. I thought, well, maybe for effect, or maybe to create hype. But when I encountered it myself, I even wrote in the chat, tagged the guys, and said: "Well, imagine if I personally encountered this, I would never have expected it." That, yes, it was spinning like this, couldn't, it squashed all the commits. Meaning, what does this mean? That there were 200 changes, now it's one change, because it squashed it into one. And thereby, I simply couldn't revert to the moment I needed. But, but I did learn that such a thing can happen. And the most important thing, public repositories. In no case should they be public. They should be private. Why? Because if some ill-wisher wants to study your code, to cause trouble or gain benefits for themselves. Therefore, remember this 100%: repositories must be private, always. What's next? Environment files, i.e., tokens, passwords, keys, i.e., what you have as sensitive data, accordingly, that's what I was talking about. Repositories must be private. Here I decided to emphasize this and even highlight it, because it's a real must-have, it's definite. Yes, accordingly, the terminal CLI, meaning here we have some commands, you don't need to memorize them. Meaning.
In general, here's how to visually remember that something can be corrupted, completely deleted, gain admin rights for oneself, or be given away somewhere. Therefore, to prevent such negative moments, simply understand that "aha, I understand this command, it's important, he might want to delete something, so I'll ask him again, or go to another source on the internet, or clarify with another AI, and so on."
Hosting is also detailed in the second video, but what I'd like to mention is that we have the wonderful Railway. Ah, well, Vercel and Render, I'll even talk about them. There are people who use Vercel, but I chose Railway, it's top. Literally 5 dollars and a lot, a lot of everything, so while you're waiting for the second video, by the way, it will be next week, on Tuesday, most likely. Railway is my must-have and top. Why? Because, ah, it costs 5 dollars. Lots of perks. For example, if some change you make breaks the codebase, it won't be deployed, meaning users won't see it. It will simply appear on your dashboard as red, indicating it failed. You'll have logs there, there are build logs, there are, ah, deploy logs, all sorts of them. Plus, they recently released MCP, you can also connect it with our worker, let's call it that, yes, with Claude. Accordingly, it will start knocking there itself. Plus, there are templates. What is a template? It's a pre-prepared solution, meaning a database, some additional programs.
Here, by the way, I'd like to recommend something else, you know, there's a service called Centry, if anyone is not familiar with it, write it down, it will help you a lot with errors, but even without it, because it costs money, and after certain limits, Railway helps very well. It also has limits, actually, on errors, but in general, in my experience, they are enough. Well, for error logging. So, set up logs so that it sees them in Railway during the build. And what I even do, банально, is if there are a lot of them, I check it, meaning I write to Claude and say, "Please, you have access to Railway. Can you look at the logs to understand what the problem is?" It goes, starts looking, and then starts solving them, and so on.
What else? Databases, where we will store, like I have, questions, user ratings, users themselves, and some other nuances. So, here's what to remember: they can be deleted. They can be deleted without any conditions. We condemn this. Well. And it can, so to speak, optimize. And here, the most important thing is backups. Backups and backups again.
An endpoint, yes, also, what we do is банально write some code that will act as our server. The analogy is that we say we have a kitchen, and we say we come, we are a waiter, we come to the kitchen. Well, yes. A is the restaurant menu, an endpoint is a specific dish. We, like waiters, go to the kitchen and say that a specific dish has been ordered, bring it to us. And, accordingly, the waiter brings it. I hope this is clear too.
The project architecture is quite interesting. Modularity equals speed of work. Version one, when I told you I had two versions that were deleted. Version one, what was its Achilles' heel? I was sitting one evening, at night, and thought, "Okay, I need to set up the architecture." I sat there, setting it up, and at some point, I reached a monolith, right? Meaning, not even the monolith as we all know it, but in the form where, грубо говоря, everything was in one file. I don't know why it turned out that way in the end, I don't remember it doing that to me, but I remember that a monolith figured there. Perhaps it misinterpreted something. But what Claude did for me? It was all on Claude, if that matters. It took everything, saved it all, the front, the backend, and everything was in one file. That is, one file, it literally grew to 10,000 lines for me. And when it became 10,000 lines, how did I notice? I just needed to add a button. Usually, adding a button took me literally, well, 10, 15 seconds, but it started taking 10 minutes. I thought, "What's happening?" I gave the task to the worker, meaning Claude. I said, "Listen, the button is taking a long time to add." It studied the codebase and said, "Chef, all is lost." I said, "What's lost?" Well, so that I, well, it's clear that I gave him instructions, like, what's the context, but банально, when we get to the context, what it is, when it needs to understand the context, what it's working with, where it's working, what its toolkit is, and all that, with which any chatbot or application starts a conversation, right? So, there's always a chat with which we interact. But when we start a new chat, so that it simply understands what this project is about, it would take, ah, 3 million tokens. Remember this number. If anyone is already working and knows the price of 1,000 tokens, you can understand how much it costs just to understand the context. And then I tell him, "Listen, okay, now we have 10,000 lines of code, I've just started the application, and it's already 3 million. What will happen when we have, for example, 20,000 lines of code?" Well, банально, it will double. It thought about it and said, "Listen, it will cost 13 million tokens just to understand the context." I'm not even talking about any changes. And here, so you understand, this is an example. Well, for comparison, the chat context limit for Claude is currently 200,000 tokens. So, in total, it would require about 65 simultaneous chats to be launched, merged, and for them to simply understand the context of what this application is about, what's going on, and so on. So, keep this in mind. Therefore, of course, for some it may be банально, but it's better to quickly mention, right, that it's bad when there are 500+ components in one folder, 200+ services without boundaries, everything mixed up in one little folder. And it's good when we have a modular structure, well, by features, here I've highlighted a clear division of responsibility, that this service is responsible for this, and here we have folders and then subfolders, many files, meaning we decompose. Feature-based architecture, right? Well, for a normal setup, this is more than enough. And a bottleneck is a bad architecture, that's self-evident. Accordingly, here's a project without modularity. For example, this is how it looks architecturally, well, not architecturally, but hierarchically, right? This is what I actually did myself, right? I took a lot of screenshots on my journey for this presentation and for future posts. Well. Accordingly, it looked something like this for me later, beautiful and all that. But when I had problems in the first version, it actually wrote me this, right, some results, what negative things will happen to us and so on, right, that search takes 15-30 minutes just to find a file, understanding takes an hour to figure out, and so on, right? This is the real venom. Everything is very bad. Accordingly, if the architecture is bad, then Claude spends half the context. Well, half is just for effect. Like my 3 million tokens just to understand the context. And in this regard, also looking ahead, what I would recommend as a must-have is to connect MCP in the plugins. And the must-have plugin is Siren. Yes, yes, don't be surprised. In general, you should have documentation. In general, you should have a cloud.md file, agents.md. If anything, don't confuse it with the men's movement. This is just an MD extension. And, accordingly, you should have Siren. In the second video, we will talk about it in detail, but now, so as not to keep you waiting, I advise, if anyone is working, connect it. It will definitely help you. This is a solution, a plugin that helps maintain context, maintain documentation. And also, by the way, from Dima, I hope you're watching, I learned something interesting for myself: he created a command. That is, skill is update, when at the end of the dialogue we simply write update, and based on the dialogue, our worker Claude, or whatever you'll be using, will update cloud.md and understand more about our project. By the way, I would recommend writing all this down, or rewinding and listening to it all again.
Security, right? So, modular architecture, feature-based. I've talked about all of that. Accordingly, this is roughly what a good structure looks like, right? Well, this is roughly authorization, then vacancies, payment, shared, and so on. Accordingly, the bottleneck is, of course, payments and finances, because you'll want to, well, somehow earn money. And here, the bottleneck is the AI itself, the payment system itself, right? Especially if we test it with our own money, and especially when we connect some first users to whom we will sell something. There's a real story, not made up, which cannot be silent about, right? That, hello, your application debited money in the amount of 12,000 rubles instead of the intended 255 per month. What happened? The AI wrote the payment logic, the code passed without thorough verification, meaning either tests were missing, or even manually, we, well, банально took it, looked at it, retested it, will everything be exactly okay? And in production, the wrong amount was debited. Just imagine how negative it will be for the user, what negative feedback they will have, that they might even tell their friends, if it's some blogger, or if they have some communications, like Telegram groups, or elsewhere, they will also go and write, "You know, this application is stealing." So, in general, this is a very negative case, right? You can get the money back and all that, but what if the person needed that money at that particular minute? So, definitely, everything related to finances, payments, and especially with users, right? With yourself, it might have been possible, but with users, it's definitely necessary to cover it with tests, retests, some sandbox tools. That is, we have a sandbox mode, we enter it, click through it, that everything is okay. So, it's definitely necessary to check all of this thoroughly.
Accordingly, how can we protect ourselves here? Well, don't do auto-approval for payment code. Also, I would add manual payment processing, monitoring of abnormal amounts. That is, if we have some kind of dashboard in our payment system, a test mode, which I was talking about, Sandbox, transaction limits, meaning even in the code, you can protect yourself by saying, "Here's a subscription, and just in case, we're not sure, let's add some check, not more than 1,000 rubles." Yes. And accordingly, double-checking, both by AI, meaning tests, and by a human, and by other specialized agents. Well, and what rule can we highlight here? That AI can offer code, but we must check it properly, because, well, it's a very negative case.
Accordingly, what rules can be highlighted for "wipe coding"? That we have security, we have secrets, let's remember them, let's close them. That is, close them so they are not visible. This is a private repo. Remember to familiarize yourself with keywords like Git, GitHub, Commit, Deploy, and so on, right? Here I would even highlight that if we don't understand how some things happen, meaning we see that it works, and we need to do something, meaning to send, to send the application somewhere to the cloud, right? So, about GitHub, what I said. Everything around that, for example, ask, "Listen, I'm new to this field, please tell me what I need to study for this, briefly, in general." Well, and infrastructure, backups, modular architecture, we've also covered all of that.
Accordingly, if there are any questions about this, comrades, you can just, let's, a standby plus, if everything is excellent for you. I'm looking, by the way, frame drops are practically zero, 0.1%. Damn, if only it were always like this. Beautiful. If you have questions, please ask, I'll be happy to answer. If there are no questions, then give a plus, if everything is going well. Now, let me read, maybe something from the stories. Thank you very much. Do you process all prompts through Whisper? Not all. This is when it's банально lazy or tedious, or something long. I would even recommend this here, you know? I wrote about this recently in the chat. If we're too lazy to write a detailed prompt, we remember all of this, it will be below, but looking ahead, if we're lazy, we can ask the LLM itself to write a prompt for itself. Yes, yes, don't be surprised. So, I'm too lazy to write: "You are a developer, you have 10 years of experience, you are technically advanced for your age, you must understand this and that architecture, you must be able to do this and that." No, I just tell it: "This scope of work, this description. Make a detailed prompt for yourself and show it to me after my, ah, after my approval. Let's continue." A very interesting approach, in my opinion.
Accordingly, this is how it looks on an update. Well, this is what I was talking about, right? So, skill is what I was talking about, there's MCP and there are Skills. This is exactly what makes our Google on steroids, these agents, these workers. So, Skills are actual abilities. That is, грубо говоря, there's a file where we've written specific prompts for specific abilities. And the update is about what Dima inspired me to do. It's банально. You just write to your LLM, your worker, "Listen, you need this skill from me: when I write 'update', or just 'update', or something else, as you find convenient," that history of the conversation, which will be in each chat, I want you to, based on this history, this current conversation, to update, for example, Siren and, for example, cloud.md. So, it's something like that, so that it learns about our project better and better each time.
Here, by the way, what I'd like to say. I was at a workshop from Cursor, from the creators of Cursor, and I asked them a question: "Is there a difference if I use my native language, and I specified Russian, or is it better in English?" That is, is there any error in understanding my prompts? I mean, well, when it's text, of course, because some understand by voice too. To which they told me: "The error is ultra, absolutely ultra small." That is, грубо говоря, 0.001%. So, you just think, is it worth it for me to write all this in English, or translate it into English from my language and copy it, so that it's 0.001% better? Well, I would say no.
Accordingly, it's roughly like this. Are there ready-made solutions to connect Yookassa for monthly payments? I haven't worked with Yookassa. I don't have any payment systems at all yet, and I'm not sure if there will be any in this specific case, but regarding Yookassa, what I've heard is that it's very difficult to work with. Or it was Yoomoney, or possibly Yoomoney, I don't remember exactly. So, sorry, I don't have that kind of experience. Well, okay, since there are no questions, let's move on. Thank you for the questions. AI and prompts, the third section. Let's speed up a bit. The golden rule of wipe coding. The quality of generation depends on the quality of task definition. Here's such a wonderful tweet: programmers won't even take on a job without a proper technical specification. "What did you want?" You need to set a proper task. And also programmers with LLMs: "Kid, make it beautiful." Well, I think, I hope the idea is clear, right? That here you can also understand it this way: if the person doesn't understand how you describe it, then the neural network, well, not that it won't understand, it will understand it poorly.
Accordingly, what we need to understand here is that prompts can be different. In informatics. What example of a metaprompt can be given here? You are an experienced iOS developer. Your goal is to create a plan for implementing a task description, breaking it down into sequential steps, and so on. Describe the task, some artifacts. That is, brief, informative, contains role, goal, and context. Where does this go next? It goes into a Task Prompt. This is already AI plus human. That is, this is a prompt for solving tasks. It already contains the implementation plan, links to some components, atomic steps, that is, step-by-step. What do we need to do? Also, as I ask Claude, for example, "Here, I've described this whole task, and at the end, I also specify: 'Break it all down into a To-do'." It also, it understands perfectly: "Break it down into a to-do," so that I can simply see what it will do and why. From this, I will understand.
Accordingly, Task Prompt is to move resources from IP to Zip on the backend, load on first launch, add a feature flag for remote resources, some action, implement changes, possibly even files. You don't need to get too bogged down in this, but what I'm showing structurally, just try it, see the difference between how you started doing it and the result, and the result when you didn't do it. Accordingly, here's to design a protocol, implement it, and so on, right? So, here are atomic steps, links to files, or even names. You can also have clear actions, but this is really for an ultra-large project. Personally, I've tried both naming the file directly, and it works equally well.
And, accordingly, the third is our goal - code. Working code. That is, specific project files, components, logic, and results. Also, what else is there? System prompt - this is what sets the role and rules for, rather, sets the role and rules for the AI before the chat begins. Example: You are an expert in React, answer briefly. Here's what I'd like to say: there's this link, I'll attach it too. And this is a collection of system prompts for all LLMs. What will you gain from studying this if you go to this GitHub and read a bit? It will give you an understanding of system prompts that will help you understand why the AI behaves in a certain way and how we can influence it, right?
Accordingly, LLMs, let's recall, are ChatGPT, Claude, Gini, there are also code solutions, there are AIs, what I was talking about, that is, programs with which we interact. I would also add Claude here, because it was released recently. There's Cursor, many people love Win, Cline, and so on. Well, and agents, like Manus, Devine, Replit. I haven't looked at them in detail, to be honest, but I've listed them so they're there. Meaning, I haven't tested them myself. Well. Accordingly, bad, good prompt. Well, I hope this is clear to everyone. What's bad is "make a button." What's the result? An unclear button. Not the style we expected, rather. No behavior, will have to redo. Good: "Create a submit button for registration, define the style on click, and so on." But here I'll give a small life hack so you don't have to do these things constantly. What I did: since I have programming experience, at the very beginning, when I set up the project, I implemented a design system. I even implemented a separate HTML page for it to see what I have. And, accordingly, for it, I then just, well, I implemented it. Meaning, I have buttons, I have some screens, I have some pop-ups, I have bottom sheets. I even have a lot of UI screens that already contain some screens in combination, some placeholders, and so on, but which are generic, meaning they are reused differently, right? In one context, in one way, in another, the button is slightly in the center, and so on.
Accordingly, I would also recommend implementing a design system here and taking ready-made options from it. So, in general, I would characterize it briefly like this: observe yourself, what will cause you the most routine. For example, do you often describe a button, or do you often run tests manually? So, just think in your free time, how can I optimize this? If you don't know, ask the LLM to tell you how you can optimize, right? So, tokens, contexts, that's critically important, of course, but in general, what is a token? A token is approximately 0.75 words, about four characters. Accordingly, we have a context window - this is how much the AI remembers. For Claude, as I said, it's 200,000 tokens. For ChatGPT too, I'm telling you, I made the presentation in December, and I don't remember how much for the fifth, but the fifth has a very interesting point. No matter how much you load into it, everything fits and fits. For Claude too, no matter how much you load, everything fits and fits. At least from the program, from the browser. It still has a limit. At some point, it gets angry, or at some point, rather, it doesn't. Before, it got angry, but then it does this: you ask it to do something, it starts doing it, then stops everything. That is, the result is empty. You think the internet is down or some other issues, you try to send it again, the result is simply empty. That is, nothing happens. So, everything, ah, the context window has ended.
What I'd like to say here is that even though we have these limits, I've tested it myself, my project is already quite medium-sized. And in general, if my context window holds about a million to 1,300,000 tokens, it's okay. That is, if you use the Claude application specifically, it's free, it's available, at the top where the arrow is, you can hover over where the project name is, it will show your branch based on Git branching. And it will also show that, look, we have 200,000 in total, but you've used this much, and so on. So, you can track it. And then what else? Real limits. Well, yes, in general, it's best not to go too far beyond them. That is, I wouldn't even say 150, but in general, 300-400 is also fine. But just keep in mind that the longer we go like this, the harder it will be for us to work with the context window in general. And I should have written this here, but if anything, I'll just say it. Let's move on. Ah, well, or okay, it's better to say it, that in general, just keep in mind that you have some task, for example, you want to implement, I don't know, autotests. Oh, not autotests, unit tests. Okay? You've created a context window, implemented some part of the unit tests. Okay, the task is done. You've done this update function. Well, you've closed our skill, reopened it. Then you continue doing it. And this way, we, ah, that is, the context window equals the task. Keep this in mind, roughly. You shouldn't think that, ah, now I've done the tests. Okay, we've done the tests, let's make buttons, let's make a multiplayer mode, let's make a PvP mode, let's make a multiplayer. Meaning, you understand, many, many tasks. And it will all, ah, get worse. Well, what will get worse? Claude's understanding of you. The context will grow, and we don't need that. In the end, we could have done it, but stopped halfway, deleted it. In the end, we'll have to ask in a new context window for it to go into the old one and pick out the important things. Meaning, you see how many nuances, right?
Accordingly, what's the rule? 80%, right? That is, not more than 80% of the context, but in general, I'm saying, I'm around a million, and it's okay for me. But
I allow that some, for example, GI, it seems, it swore at me when I, well, not that it swore, worked poorly, when I went beyond the context. And new chats every 30-50 messages, in fact, this also depends on the application, on the task set, just to choose for yourself, how many messages, how many there can be. I will change it for myself in the context. It used to be in messages. Well, naturally, if there are glitches, then a new chat. Well, accordingly, MD needs to be set up at the very beginning. We create an update function to constantly update it and make it more detailed and better, better. As our vibe coding solution, applications, mini-apps, and so on expand, what you will do there, so that it constantly improves. And for larger tasks, separate chats. Well, and step by step, that is, it improves quality, that is, to think step by step. Well, accordingly, what kind of hallucinations can there be, when it confidently lies. Well, what I already told you about Venom, gave an example, non-existent entities and folders, give a like if you've encountered this. Made-up functions, APIs, outdated syntax, and so on. That is, they can actually make things up. And frontend and backend, these are parts of the application. Here, too, in brief, that the frontend, is what the user sees. This is exactly where we can click, enter something. These are buttons, forms, reactions, and so on. what is on the server, we move some calculations there, we don't keep it on the frontend, we just get the result. Accordingly, as an example, the same one. A store is a front, plus a factory is a backend. That is, the store sells, users see what products are available, and so on. And on the backend, all this is implemented, bread is baked, buns are made there, and so on. When chat GPT starts to talk nonsense, literally, how to understand this? Well, we do several iterations and look at the result, then, accordingly, it's not us who are stupid, it's the context that has overflowed. Start a new chat. Magic. No, it's knowledge. Well, and here's a joke. Do you know why vibe coders go gray earlier? Because once an AI wrote Mrf and said: "This will optimize performance. Just kidding, but make backups." Well, we remember this from the initial section. Accordingly, advanced prompting techniques. Metaprompting and writing a prompt for AI, and why not? I tried this, I talked about it at the beginning. I don't know. Well, I conditionally assume that what if I explain poorly, let chat GPT explain for Claude. Accordingly, I started making this kind of template for him. But, as I said at the very beginning, allow that at some point it might not have worked, but time passes, it evolves, improves, and at some point it starts to work. That is, the fact that it didn't work for me doesn't mean that, I don't know, in six months it won't work. Perhaps in six months it will work even better than if I write to Claude and so on. But, but, well, that is, yes, the template, remember, the presentation will remain, you can rewatch it, but what positive or negative results can we get, that we will have a good structure there, requirements will be highlighted, it will even come up with some examples, details, because remember, we will be too lazy to write all this out, too lazy to even talk about it on Vitsper. But what can be the downsides? We might not know the context at all. That is, we have a design system, it will tell him what kind of button should be beautiful. Although in fact, he should have said, listen, we have a positive button, let's take it. Or a destructive button, let's take it. The result of the work is deplorable, in particular, mine was like that. A longer path, yes, because we need to keep chat GPT in context. We remember that for Claude, the one who works, to work well, he needs some agents. Plugins, it turns out there are plugins for Claude MD, it turns out there are. In general, documentation needs to be kept. I would also recommend decomposing this documentation. That is, not one documentation where everything about the application is, but that there is a folder, here are services and files, that here, the goals of the services are a separate file, a separate file of assignments, what they can do. In the third, these are examples, and so on. That is, you see how many nuances. And all this needs to be maintained, right. Well, accordingly, what else? What techniques? Structure, well, instruction and context. The correct order, the correct result. Accordingly, the structure of the prompt, instruction at the beginning, goal, limitations, format, context at the end, that is, code, logs, data. So, the AI immediately knows what to look at. That is, it's bad. Well, audit the code for performance bottlenecks. You are a senior developer, iOS, backend, it doesn't matter, and so on, yes, the stack is indicated. Everything is in a mess. Instructions are mixed with context. How could we do well? Instructions. First, identify critical bottlenecks, do not offer solutions that will break our current diagram. If there is not enough data, explicitly tell me so that I understand. Project Navui plus Express graph. We observe problems and so on. Screenshots can also be attached. Here's what I'd like to add. If you work through the terminal or CLI, then those who don't know this, it means you won't work. Those who know will understand. Keep in mind that 4-5 MB screenshots break the context window, that is, it's gone, it can't be restored. This was a big pain for me. That's why I learned to specifically send him things. But in general, keep in mind that, firstly, you can send pictures to all IDEs and instead of describing them to you. You send it, and it will understand what it's about. Secondly, I would also say here, this is a real must-have. Specifically, what did this idiot do to me? When he couldn't find a solution. For example, I was editing questions for guys who were preparing for interviews, I needed to specify sources where they could prepare further. I wrote to him, that you should place these and these links in this and this question. For some reason, he couldn't find it in the database. Very often I reminded him that he has direct access in Railway, not that there is no access, a script needs to be made, and after that, SQL injection. This is complete venom. But I reminded him, although he knew this, it's in all the docs. You see, not just one doc, but all of them. And accordingly, I constantly reminded him that please, you. By the way, this is the second must-have knowledge. Just write it down everywhere, remember it everywhere, so that this rascal, this hard worker, never does anything with scripts. Well, of course, there are some scripts for searching, etc., etc., but where you are looking for quality, there should be nothing script-based. For example, there are 10 questions, and he says: "Well, let me update all the materials with a script, it will be faster." And he often does this because he considers it more profitable and better than doing it step by step. And what happens is that when nine out of ten scripts are complete nonsense, this rascal breaks it, and I have to redo it all over again, and redo what is already broken. That is, it's like, you know, an empty plot of land, you come there, you build a house from scratch, you know how to build it well, you have good builders. But when it's when he does it step by step, that is, he improved one, as I asked him and what I asked, okay, the next one. The next one, the next one. But a script is when you buy land, there's already a house there, and you want to either demolish it and build anew, or build something good based on it. You built something good, but it turns out that the foundation is rotten and so on. Well, I hope you understand the analogy. But this, believe me, this is a real must-have, so that it doesn't do scripts where you are aiming for quality. This is something to write down, learn, and always follow. Details 80% of the result. The fewer imprecise formulations, the better the result. Inefficient. I want a beautiful, engaging, interesting design for the user for my application. Effective. Create a design for the application in the style of, say, futuristic minimalism. Here are some design examples from Figma. Remember, you can use links, screenshots. Study best practices on the internet to implement a design system and reuse components, and so on. That is, you see how detailed we are about what we want. Because if the technical specification is unknown, then the result is unknown. Checklist of details, what exactly to do, according to what rules to do, where exactly, how it looks when ready. Example. User personal account. Bad. Create a personal account with an avatar and buttons for settings. The button to change the name should be able to notify the user. How could it be done better? We will have a personal account. The avatar will adapt to the sizes. Settings have buttons, well, we list them from the design system. The button to change the name validates the user. Create a validation service. It can even be reused. We will use it. Create it. Well, it will even understand on its own that it needs to be made Generic. That is, you don't need to go into such details here. Typical errors in prompts can also be. Coding without a plan. That is, add a button, then a form, then a popup, spaghetti code in 2 hours. That is, in general, I would characterize this not so much as without a plan, but without a certain vision. That is, well, it seems like it will be cool. That is, even if you don't have a plan, it's better to ask Claude, talk to him about how best to do it, so as not to have to redo it later. Putting everything into one prompt, well, as I said, so that it does both testing, authorization, catalog, and so on, yes, working with only one model. That is, in fact, I am currently working with one model, but I periodically review others. Like this, cascading. First one, then the second, can it code better and so on. But in general, Claude, as it seems to me now, is simply the best in coding. That is, Claude is an LLM, remember, right? AI, that is, there will be a cursor, and the terminal will be, that is, a string window, or it will be, I don't know, some other things. Maybe you yourself are on a vape, it doesn't matter at all. Well, accordingly, not testing the result. Well, yes, if the AI says it's ready, and you're like: "Well, since it said it's ready, then it's ready." I would recommend writing tests simply so that it can run them. And this way you will even catch bugs in production. What else would I recommend, as I do, when I catch a bug in production, I write to the workers, listen, we have such a bug in production. Why did this happen and can we solve it with testing? That is, every bug in production is a ready-made test for the future. Because imagine, when your application grows, how many screens, how much functionality there will be, and you are simply doing something new, and somewhere deep down it broke. And how to understand that it broke? And what if it affects the user? And the user, well, won't tell you, because they are shy or something else, so, well, I would write tests in your place. Different LLMs for different tasks. You also need to understand that not chat GPT does everything, but chat GPT drew the design, came up with the naming, possibly came up with some tests, Claude implements all this, and Gemini, for example, conducted some research, made a PDF, and so on. That is, the essence is that do not skimp on looking at what solutions are available on the market and how they can help you. Well, accordingly, the rule: one chat, one task. Each block of code in a separate chat reduces the chance of hallucination. Stages of vibe coding. Everything is quite simple here. Development of a plan, that is, we must have a task, constraints, and definition of done. Definition of done is what we consider ready, right. Accordingly, we have iterations with agents. That is, we give a task, we get code for this task, we start to clarify something, if we don't like something, or some bug. We repeat until it's ready. But before it's ready, we can have tests and debugging. Accordingly, we run the tests, look at the logs, error screenshots, if we have such functionality. If everything is okay, we deploy. And as I said, with Railway, it's much better. Not that they even pay me, it's just a very good service. I like it so much that I always recommend it everywhere. Well. And we will analyze each stage in more detail in the second video. We've just gone through it now. Accordingly, here, please tell me, maybe someone has questions, guys, clarifications, I'll look now. So, why write the role of the executor in such detail at the beginning? You are a professional developer, you have such and such experience, and so on. How much is this introduction? Good question, actually. Great, Denis. This is so that he understands who you are, because based on the documentation, based on some previous prompts, based on some chats, he remembers something somewhere. But when you specify it, that is, you need to draw, and I can draw, for example, but I'm not a designer, I'm a programmer. Well, I'll draw something for you. And now imagine, you call a specialized designer who understands how design is done, what meanings are in design, and so on. That is, it's like a specialized person, that is, you don't have an any-key person who can do everything, but you have a specific person, for example, a coder codes, a designer draws. Specifically, I feel a striking difference in the quality of the output, that is, in the tests, in the functionality, in the code, in the pictures themselves, and how he does all this. That is, simply, I ask him, how I do it, that is, I write to him, listen, assemble a working team for me, I list who I need there, a project manager, two programmers. You can even, you know, do something very good, a practice, that you have users, but you still can't know completely. And you conduct research, right, some, to find out yourself. And then you tell him, listen, let's have three users, who, in fact, he will even understand that he will set up himself, one of the best buyers of your application in your world, the average is someone who came, is doubting something. And the third is, for example, a hater, and he says: "No, it's all crap." And imagine, when he runs this feature, before, you were like: "No, the feature is cool, because I'm cool, and I think this feature should be." And then, he set up these roles, and he goes through each one and tells you that look, this guy, who is a hater, thinks that this is a very cool feature, and it even makes him think, hmm, that is, the division into roles is even basic, that we need such and such roles for the implementation of this task. Here's what I'd like to add: if you start doing this, you'll encounter the fact that the agent might have sub-agents. And here I would like to say that my personal experience with sub-agents in specific executive activities, namely doing something in the end, that is, implementing code, design, and so on, I didn't like it much. I still liked it when Claude specifically accumulated all these roles in himself. He just went through each one in turn. Why I didn't like sub-agents much and they didn't suit me is quite simple, because they don't inherit context. They need to have prompts written for each one, and they don't know about each other because they are atomic. The first one only knows what he is doing, and so on. The second one also knows what he was told to do. But for the first one to know what the second one is doing, and the second one what the first one is doing, they need to look at each other. So, they are writing there, maybe everything is bad, but no, oops, oh my god, everything is fine. Well, accordingly, it's something like that. I hope I answered. Well, accordingly, yes, this is also interesting. Someone compared Vinsurf and Cursor specifically. I compared Cursor versus Claude. I like that Cursor has a UI on the left, shows folders, shows files, and so on, but in general, I don't need it at all. My project is medium-sized, so it's just distracting. I have Claude, I have a chat. Before that, if it weren't for the pictures, I would have stayed in the terminal, because it's just distracting. Why do I need to know this? The project is growing anyway, and the worker knows better what is where in the end, and so on. And even if I don't know a specific file, well, let's say I don't have this interface, I always think: "Damn, did we have an analytics service or not?" I would look for it on the left, go into folders, but instead I just say: "Listen, do we have it or not?" He will quickly look and so on. So, in Gemini, I only work through prompt engineering. I have chats that do tasks. There is a separate chat that explains to them what I want. So, metaprompting. Yes, thank you very much. See, I had specific experience with chat GPT. It gave me an irrelevant result. I don't like it at all. But it's good that you shared. Thank you very much, because, as I say, you need to look wider, you need to try other LLMs, other solutions, and try them out. Well, okay, what if chat GPT can't? What if, for example, this is a good example. I'll try this too. Thank you very much. Send the link to GitHub with prompts to the GT group. Yes, I'll send it now. It's very good that you said this. I'll send it to the chat now. Just a moment, a moment. It should be here now. So, now where I'm gray and literally, if anyone else has a question, ask, and I'll answer the last one, if there is one, and we'll move on. Yes, here it is. Yes, thank you, it's good that you said this. So that it can be studied. In fact, there are many different links. They just fall into different videos. Does anyone else have any questions? So, I've sent it now. Uh-huh, you've been on YouTube for 17 years. Choose me. No, this is Uh-huh, also last February. Prompt. Interesting. Enough. Share your results, what are you doing at all. You can write to me in private messages later, or in the chat, or both. Well, okay. Let's move on. Come on, guys. We've discussed the stages of vibe coding. Advantages, yes, accordingly, in fact, what are the advantages of vibe coding? Well, the first one is business. So, now I'll choose the laser pointer here. This is business. For business, speed equals money. That is, business can be different. This needs to be kept in mind. That is, business needs three things, in fact, for a product to earn, work stably, and be safe. What code quality, what picture, it doesn't matter at all. Here I would like to give you advice: don't, if you, my soul aches, I know exactly who I'm doing it for. I don't earn from this, I just deliver it as additional functionality for the guys, which is part of the overall functionality for which they have already paid. And therefore, quality is very important to me, so I really get into it and look for different hypotheses. But if your goal is to earn money, you don't need to sit there and write, do something, right, it doesn't matter at all. You came to vibe coding to earn money, to make some kind of rocket, right? Our task is to beat this damn rocket with different schemes. Therefore, we don't get into gambling, we don't write detailed quality code, we don't cover it with tests. Wrote it, tested it, formed your opinion, that's it. That is, here, too, what if what you came up with is not needed by the market at all, and you are connecting some tests, some Grafana, and so on. No, first of all, just understand why you are doing it. Maybe you want to get acquainted with vibe coding, okay? And so on. Accordingly, needs and how vibe coding solves them, this is understandable, but let's discuss that MVP in days now, not months, that we don't need a team, we have one worker, we do everything quickly like this, prototyping is also done quickly, and so on. But an important caveat is that the code needs to be checked in one way or another, because AI generates, but we control. And my project started generating income after 2 weeks of development. I didn't have perfect code, but it was working. That is, the guys could already pass tests, yes, maybe the UI didn't save something or something else, oh, the backend didn't save something and so on, but there was already something, and the guys were happy, and then we refine it further and further, that is, I started writing tests, improving the architecture, and so on. For the user, what is good about vibe coding? Problem solving, actually. That is, the user doesn't care what it's written in. Whether it's, I don't know, React, or Node.js, or, I don't know, some other W, and so on. No, the user needs a solution to the problem, convenience plus beauty. That is, the user comes in, wants to prepare for an interview, what's under the hood, what tests, units, integration, I don't care at all. Accordingly, a shift in focus. If before it was 80% code, 20% user, now it's 30% working with AI, 70% is the user. That is, I spent 2 days making a convenient salary comparison. I showed it above, but not because the code is complex, but I wrote it very quickly, in 30-40 minutes, but because I tested different UIs. I tested whether I needed this shield or not, but then I found the one that would be understandable, and everything was great. So, vibe coding has freed me up specifically, it frees up time to focus on what is truly important, which is understanding the user. That is, code is now a means for me, not an end. Here I missed something, but okay. That is, here, I hope, it is also clear that now everything is delivered much faster for the user, much clearer, yes, maybe we can complicate something, but we complicate it, not the user. And for the programmer, as many think, everything is bad, but the fear that AI will take my job. What is the reality? AI makes us more valuable now. Why? Because now we have a new role. If before we were executors, doing everything, now we are like architects plus controllers. Important skills now are architecture. I would add decomposition, code review, prompt engineering here. And thinking about business is not just that. And this doesn't apply to programming at all. This is something you need to grasp. These are just some nuances that I decided to highlight. Because in reality, who among you programmers has encountered this: you were lazy to do a task, you went to do it through an LLM, it did it, say, in 10 hours, it did it poorly with bugs, or didn't do it at all, and you were like: "Damn it, I'll just do it myself." This is not about, well, this is how it is now, but in 1 year, 2 years, 5 years, let's talk, because I want to remind you that OpenAI was released in December 2022, if I'm not mistaken. So, essentially, it's been about 3 years. And, accordingly, this, of course, does not mean that they will be stupid, because if you are on Twitter, you have seen Boris Cherny, that's his name, where he wrote that in the summer of 2026, let's synchronize our watches, and there will be no programmers, and AI will completely replace programmers, he wrote. I would say here: "Well, this is a person, this is the creator, this is a big shot at Anthropic, this is, well, not the last person, right?" So, okay, then what if I give you Elon Musk as an example, that is, can we say that Elon Musk is a bit more serious than Boris Cherny? Well, in a way, yes. But who remembers how Elon Musk told stories. It was around 2020, how in 2026, by the way, or in 2027, Teslas will fly, there will be robots, well, and who was there, there will be automatic Teslas driving without a driver, and everywhere. Well, synchronize your watches, ask people, or look at where this is being implemented, right? For example, in America. Accordingly, what is the metaphor here: a programmer with AI is like a pilot with autopilot. That is, autopilot does the routine, but the pilot makes the decision. And when something goes wrong, it is the pilot who saves the situation, not the autopilot. AI does not replace a programmer either. And it makes a good programmer even better or excellent. Accordingly, AI is a tool, and it will not replace a programmer. AI makes a good programmer excellent. This is actually a big holy war. I would just say this: try vibe coding and you'll understand. Literally a couple of days ago, I was vibe coding. I needed to change one field in the database that was not being read correctly. What did the worker do? I wish I had inserted screenshots here. I'll insert them in the second one. The worker did the following: he wrote a function on the frontend that would accept the incoming value, somehow modify it within this function, and output the desired result. And I write to him: "Are you crazy? You're using a crutch. We could have just not made a function that converts from an incorrect format to a correct one to show the user." We could have not made this function, gone into the database, changed this property there. Voila. We have no bugs, we have no unnecessary functions. And here, just imagine, when we, without understanding the context, that's what I'm talking about, click yes, apply everything, yes, apply everything, yes, apply everything. He keeps doing this, and he will do these kinds of things. And then the application will breathe. That's all the science. And here it's even interesting that when bugs appear in the application and you just, well, Claude, for example, says: "Chef, all is lost, we have bugs." You tell him: "Are you stupid? You do 95% of the code. How did you allow bugs? You're smart. I wrote such prompts for you." Well, that's how it is. And here's joke number two. Do you know the joke? A programmer goes to a psychologist. Doctor, I think AI wants to take my job. Psychologist. Have you tried asking AI to write a resume for a new job? Programmer. Brilliant. Moral. AI is a tool even for fighting AI. Are there any questions here? We are almost at the finish line. Yes, we're a little delayed. Unbelievable, I condemn. So, earlier Vinsurf had a better payment system. They give 500 credits per month, one request costs one credit, and Cursor, in turn, used to bill by tokens. Now it seems to be the same. Uh-huh. Anti-gravity, by the way, yes, also a topic, guys. Force is constantly a topic, in my opinion, now. Can I buy a year-long subscription through the OneDrive tariff? Yes, yes, that's right. Anti-gravity is very hyped right now. Well, no questions yet. Thank you. Ah, I understand, svyatik akrobatik. This is that rascal. Ah, agents. What are they? What are agents? That is, it's not just a chatbot, that is, not something we open, ask questions, it doesn't have access to files, it doesn't execute any commands, right? And an agent is such a beast, it's a program that can read our files, execute commands, fix errors, write and edit code, make decisions. Integration via MCP. Here I would also add skills, integration with other services and solutions. Well, that is, an agent is a whole autonomous program that acts, not just responds. That is, as I said, MCP and plugins associated with it, skills, this is exactly what clearly shows how all this is tuned, right? That is, clearly. Here's an analogy. An agent is like a junior developer working 24/7, doesn't get tired, doesn't get offended by criticism, right, when we scold him, learns instantly, but without a clear technical specification, he will do nonsense, like any junior. Types of agents. Well, here I've highlighted browser, new, lavable, replet. And I would also add Atlas from OpenAI here. Well, accordingly, for beginners, quick prototypes, nothing needs to be installed, less control. What do we have? IDEs, that's Cursor, Winerse, Codex, and, well, Claude, recently released. For whom? Developers in their usual IDE. Well, actually, for everyone, I would even say more. Familiar interface, tied to some program. Here you have it all well. Desktop application for whom? Universal option. Well, for everyone. This is my choice. I want to emphasize this again. Full control plus MCP. Works with any stack. Works very well. Recently, but for paid users, they added the ability for it to do something on the computer, go into folders, copy and paste somewhere, and so on. That is, there are three plugins. One is like a chat, the second is like a worker, that is, to build something, do something, and the third is to interact with the computer, well, with yours. Well, and terminal first, I've already talked about this, this is CLI, like Claude, GMI, GI, and so on. For whom, well, experienced developers, in fact, that's why I said, those who are in the know, are in the know. If you don't know, you don't even need to delve into this. Maximum flexibility, requires experience. Well, CLI, of course. Accordingly, why I chose it. Claude application, convenient interface plus control, I really like it, especially after the terminal. Behavior settings, skills, which are here. Well, they are also in the browser. You can put instructions in a folder, but this is just to note. Three modes, right? That is, remember that we can ask it to do, or first create a plan, then do it, or ask questions, and so on. It writes code better than analogs, but not entirely about the program, this is again I would emphasize about the LLM. MCP plugins, which we can use in CLI, that is, in the terminal, and in the program, but not in the browser. Well, and it installs tools itself. Yes, yes. Don't be surprised when you tell it, listen, let's set up this program. That's what I was talking about, that they added the third one, when it can rummage around on the computer. In the second video, we will analyze Claude in detail, namely MCPIs, skills, operating modes, and railway deployment, which is also now associated with it. Joke number three. Cursor is like an iPhone. Beautiful, convenient, everything works. Claude is like Android on steroids. First you configure it for yourself, then you don't understand how you lived without it, and then you ask it to install plugins for itself. Here we are approaching the conclusion. Maybe there are some questions here. Not yet, but okay, let's wrap up. Well, and the conclusion, what did we learn today? What is vibe coding, why is it the word of 2025, basic terms, what LLM models exist and how they differ, security we recalled or learned. Well, and the golden rule is that the quality of generation is the quality of task setting. Again, this is very important. Promtflow, metapomtask, prompt cд. Advantages for business, for users, for programmers. What are agents and what types are there? Why Claude is an excellent choice for vibe coding? Well, in my opinion, of course. This is, as it were, a fundamental video. Somewhere it was, perhaps, very simple for someone in some places, but in general, remember, we are all different people. I wanted to cater to everyone, so to speak, in this regard, to share my experience. And there will be two videos, because, you see, it's almost an hour and a half here. The second video, of course, will be shorter, and I will try to ensure that this doesn't happen, but it will contain. What will be? We will delve deeper into Claude, into MCPIs, into skills, railway deployment without pain. My mistakes over 400 hours with screenshots, of course. Well, and workflow, yes, Boris Cherny, that's what he was called, I wrote down how they work there. Well, and accordingly, a bonus, yes, reality versus expectations, roadmap for starting, where to start, what to focus on. I even mentioned this today, by the way. Well, looking ahead, again, to create a design system for yourself. Well, and looking ahead again, to implement some stack for yourself, that is, a design system, some real things like Railway, for example. That is, just to talk to Claude. What technology stack, solution stack can we create for ourselves that will accompany us from one IDE, from one of your vibe coding solutions, be it a website, an application, or so on, to another? Why? Because it's very good, as we developers do, that we set up navigation, we set up an authorization service, we set up a design system service, conditionally, well, not a service, but just, that is, we did these kinds of things. That is, when I do the next application, I don't need to do navigation anymore, I have it. I don't need a design system, I have it too. I just need to change some colors, the palette, and so on. I don't need to do authorization all over again. I have it implemented too. I just need to connect some specific service, that's all. Well, and accordingly, yes, I will be glad if someone likes this video, writes a comment, you can even send a greeting, if there's nothing to write. This will help me promote it. Well, and write your questions in the comments, if there are any, what to really ask. Well, and see you in the second part, comrades. Yeah.