Transcription
Hello everyone. So, hello, hello, hello. I hope everything is well. I hope I am heard great. If you hear me great, hello. Write again that I am seen and heard. Write in the chat, and in the meantime, I will write in the wipecoders chat that our broadcast has started. So. We are live. Okay, super. Well, hello everyone, it's 2026, and today the stream will be about CLD-code. Probably the coolest tool that you can use right now. I have been using clodcode for, probably, 7 months already. And all this time I have been trying all the new tools that come out. And at the moment, clod-code is my favorite tool. And honestly, I think this is the very AGI that they promised us, because there hasn't been a single task yet that I couldn't live through with Clod Code. And, frankly, it gives me a feeling of superpowers. And and what else? Well, there are many thoughts on this. In this stream, I will talk about my journey. If you have never used, never tried clodcode, this stream is for you. If you have tried clodcode, if you are a developer, the stream is also for you, because here I will talk about my entire experience over 7 months, about life hacks, about some stories that might be useful to you. So let's start. And I have been using platkod for 7 months, paying for it, I'm on a $100 subscription. Let's start by writing down the stream plan. I expect this stream to take me about an hour, but we've all been there, we all know. So, lodkod, first of all, let's talk about, yes, where to start. Where was our list? Mhm. Checklist. Great. Where to start? What else? Tools. In my case, I use Super Powers, and this thing gives me direct PS. I've been talking about them for a long time, I promised. This stream will be about that. By the way, prompt in the chat. I'm reading the chat on a separate screen, and we can also do it together. I am in Argentina, so don't be surprised. This is maté. I will be drinking it. So, where to start? Hmm, where to launch? Ah, subscriptions, probably about subscriptions first. Max subscription. I pay $100 for it. Is it worth it? Is it not worth using it? Superpers It tools I use. What else do I use? Probably, probably that's all. Let's talk about MCP Skills at the end, something else in that vein. Something like that. This is maté. A South American drink. It's made from the leaves and branches of the maté tree. It tastes quite nasty, bitter, but it invigorates. It really invigorates and wakes you up. I switched completely from it to coffee. How did Open CД do it? Yes, a beautiful interface. I like how it looks. For those who haven't used it, it looks like this. Oh, what is this? Here it is. Here it is. It looks beautiful, but I haven't used it. In the end, I just realized that I don't need anything. I have I have code-code and I don't need anything else, because any any things, any things it solves. Well, where to start and what is clodcode? In general, code. It's not it. Yes, it is. And I and coding agent. In general, it's an infrastructure for launching, for launching artificial intelligence. In this case, in my case, it's Coop Opus 4, so-called wrapper. A wrapper for your agent, which you install on your computer and launch in this terrible terminal for many. But since it's the same as a chatbot, when you only have and you can write: "Hello, how are you?" And there's nothing else here. And we have completed a full cycle. The history of computers began how long ago, about 80 years ago, right, if I'm not mistaken? And it started precisely with terminals. So we have returned to all of this. Therefore, therefore, yes, you need to start by simply installing it, figuring it out. By the way, OpenCД can be a good alternative for those who are reluctant to spend money on clodcode. Therefore, therefore, yes, I have, I don't know, they are still giving away, code-code gifts. I can try to share. Or have they already run out? It's unclear, unclear, unclear. Okay. I saw it was fun, but now this inscription about sharing has somehow disappeared. Therefore, yes, OpenCД can be a good, good alternative. Therefore, it's worth starting by simply downloading this thing. There is an instruction here, launch it. And that's basically it. You can launch and work with clodcode in several ways. I would highlight, probably four, four options. The first and simplest. You don't even need to download anything. It's available to you right here from the browser window. For this, you need a GitHub repository. You add a GitHub repository here and then give tasks, and it connects to your repository and writes code there. Very convenient. I use it for some minor edits. A big plus, it's convenient that you can use it from your phone. This is a cool story. Convenient for some small tasks where you don't need to test anything. Maybe for the backend, maybe I use it for translation, honestly, most often to research something, translate, and, for example, update text on a website. Therefore, this is the first mode. The second format is for those who love the terminal very much, there is a DD-code application, and it has a few more features. You don't need a code editor, you don't need anything. Right here in the clod application, which you download on your system, there is a local history. You select a folder and, accordingly, give commands there and start working with them. And it will even support Work 3, meaning it goes into such tasks here. And you can even run multiple sessions in one folder in parallel. It might be convenient. The next stage, yes, the third, the third way. You can install clod-code in cursor or VS-code. I'll show it in cursor now. And in this case, it will work for you as an extension. Let's open some project, I'll show it on a project, for example. So, code it is, and accordingly, it will work for you here. It will also work the same way. And I don't really like this format, but I'm just a spoiled person. I have quite a lot of development experience, even if only frontend, but nevertheless, this thing, there are many advantages. Firstly, visually everything is nice, everything is cool, it looks like this, your usual work, it writes code. But when it comes to, when it comes to some workflows, in short, I don't, it didn't stick with me. I have many friends who use it, and they like it, it works for them. But, in general, I have completely switched to it and will show the fourth stage, which I use, which is to launch it in the terminal. I use ItM, and I write about all my findings in my Telegram channel, so the link is in the description of this video. You can join, you can watch. Here it is about, here it is about atyorms, and forecasts, and about skills. In general, various announcements fall, fall, fall, fall, in general, here. And I have a series of workshops starting on January 13th, so you can still join. There are, I think, five spots. There I talk about all these things in more detail, and we analyze practical things there. So, let's get back here. This is ITM. Am is a shell, for the terminal, which supports out of the box various things that I used to configure additionally in the terminal, now I don't configure them anymore, because everything is here. For example, multiplexing, such a thing that allows you to run several, several windows within one session. And this is very convenient. For example, like this, I have two windows open here at once. Let's do the second one. What I probably did is I set up an Alias. An alias is a shortcut for quickly launching a program. That is, if you downloaded something yourself, downloaded something to your system, it will most likely create some alias for quick launch of something. In this case, you can ask code to create a quick launch for this thing for you. Bypass Permissions, it seems to be called. I don't remember the exact name anymore. But the story is that it won't ask you questions, clod-code won't ask you questions. Yes, I have this mode enabled here. Bypass permissions on. This is this mode itself. And I have it configured, I asked for it that way. Prompt looks like this. Make me a quick launch for clod, in permissions mode. This is probably the first thing I do when I install clod on my system. It, in general, does this. Then hotkeys help a lot. The very first key you should learn is the Escape key. It cancels an action. That is, you launched it, it's doing something, you realized that no, this doesn't suit you, it's going somewhere wrong. For example, what's in the wipecoders, and it will start thinking for a long time. But at some point, if you realize that no, you don't want it to write code or do anything, you just press Escape to cancel, it interrupts. The second thing I often use is Shift + Tab. Shift + Tab changes the modes of clodcode operation. Probably one of the most useful is plan mode. I use it in Ask mode when I don't need it to write code immediately, but I need it to just answer me in chat, for example. For planning, I use something else, but it's fine. And for planning too, if you're just starting something, use, start with planning mode. Another thing, the third, the third very useful thing is, yes, by the way, questions, questions in the chat. Let's answer them. Mm, so, an example of using CPX. In what situations? I'll tell you about MCPX. Let me let me set this up for myself. So, MCPS skills XM. So, where to start? Let me start, then, by entering the extension and the fact that now this is called power, so I've talked about prop. So, so, damn, well, of course, >> come on, Ivchik. It's fine. >> I hate, I hate everything, everything related to the app now. They are disgusting. I won't put emojis. Like this. Like this. Skill extension. Okay. So, let's move on. Where did I start? And then we have life hacks. Accordingly, keys, that's Escape, that's she + tab. Planning mode is also an important thing. I've talked about it. What's important here? The second, the third, the third important thing is statusline. What does status line do? How is it configured? It's configured with the command Status Line, you don't need to install anything else. And it allows you to run additional scripts every time. And in the end, I just have statusline running for me, give me my status line script. >> Dad. >> Ah. >> Dad, you didn't tell me that. >> With what water? Say hello to everyone, >> Dad. >> Yes. Say hello to everyone, to the camera, say hello to everyone. >> Hello everyone. >> Super. Thanks. Go ahead. And I'll continue wipecoding. Okay, can I? No, Dad still didn't tell me what happened to the water. >> What happened to it? >> I'm more likely. >> Everything is fine with the water. Everything is fine with the water. Look, I'm even drinking maté now. >> No. But >> run to Mom. >> To the pool, >> and the water in the pool broke, the bus broke, >> and therefore they couldn't take anyone to the pool. But tomorrow they promised that they would fix the bus and take everyone to the pool. Mom, the bus. >> So, yes. Statusline and this statusline of mine, which is simply, in my format, you can ask to configure everything that concerns this matter for yourself. Stunning, yes. So, in general, yes. In the end, it shows the selected model. In fact, you don't need to show extra information anymore. It shows the current session. It shows the session for the day and shows how many tokens I've consumed per month. How can this be viewed? Ah, you can view it with the command, npx CC usage. Here. And it will give you a table, which will show, how many tokens you have consumed, per month. I have a $100 subscription. I'll talk about it a bit later, but here, well, I'll run it again now so that it aligns the formatting. >> It still didn't align. >> Well, okay. Input output cost is normal. So what's here? What can be done here? There is a breakdown by day. The main trick is that clodcode stores everything it does, all calls, in your system, and therefore you can calculate it. That is, this is all within the subscription that I pay for, but here it is calculated from the price per million tokens, which can be found, for example, on Openrouter or in the clod API. Therefore, probably the most profitable subscription that exists now is the clodcode subscription for $100. I, in fact, completely cover my tasks, both for work that brings money, and for, and for some simple fun, doing various things, pranks, which also may not bring me money, but bring great pleasure. But, in fact, I hope that someday they will bring money too, because there are very interesting things. Ah, so, the price comes from here. The price comes from the calculation of these things. Ah, but I pay the same price, $100 per month, accordingly. Damn, you're getting on my nerves, that it does this every time. I will, accordingly, do this. Statusline. So, Max subscription. Ah, Max subscription. It costs $100. So, where is it? It's here. Hmm, no, >> in general, yes, it costs, $100. There are, there are several modes for you. You can use the subscription for, here for how much? $17 or is it per year? Probably per month. Per month, I think $20. Ah, yes, $20-20 dollars, if per month. But this will be enough for five Opus requests. This is a demo mode. And in general, I definitely do not recommend using it. Another thing, Max 5X. They say 5X, but it's not 5X, it's much more X, it lasts a very, very long time. When you have this subscription, you can check your limits, usage in the settings under the Usage tab. Here you will have, limits shown. Here it says that my limits will reset on Thursday at 7 AM, and you also have five-hour limits. 24 hours of this limit have been used by me, and in 22 minutes they will reset. There are also general limits on tokens. And I'm starting to run out of the $100 subscription. And the only solution I see is creating specific agents for specific tasks on the internet, because not all tasks absolutely require Opus. Just yesterday I tested the story with detailed planning and passing this plan to antigravity on Gemini Flash. It's cool, everything works quite fast, and in general it proved successful. Everything is through TD. And this is the future. Where I think I need to go when I master all this is writing red tests and clear, clear, clear planning, then writing red tests, and then I don't know what to do next, then scale and launch many agents. That is, if we say that you are currently engaged in transferring programming to agents, then the next step is to create an AI manager who manages AI agents, and then the next step is to manage these AI managers who manage agents. And then there is also a 20x subscription, but I have never used it. $100 is enough for me for now. Perhaps at some point I will switch to it. I have used, probably, everything that is on the market. I tried cursor, and antigravity, Google for ultra, for how much is it now? $130. Then what else is there? Open AI Codex for $200. But the biggest profit I got was only from this subscription. Nothing else came close. Well, the second coolest subscription is Google's subscription for $20. It's also top because it has antigravity, it has nanobana and Veo, in general, everything is there, but in this format, yes. This is probably the best subscription, because I use it, and it, that is, clod-code with Opus by default everywhere and the version for chats from the phone, and clod-code in general for chats too. So, answering questions. Mm, this is the mode, yes, the mode. The mode that I launched here. Bypass Permissions. Yes, that's it. Dangerously skip Permissions. That's it. Tell me how to better Figma MCP to make a landing page through screenshots. It doesn't allow losing design steam. If it remains, maybe code can you advise? I probably can't tell you about that. I used Figma MCP 3 months ago, and in general, you do it by screens, but there are significant limitations on the context window, so I won't talk about it in this stream. And so I bought Max, enjoying the quality. And the voice was of some hero from Dot speaking. Yes, it was an Orc. It was an Orc from Warcraft speaking, but I've stopped doing that now. Why is it worse to use finsrv, where there is already the same clod? The thing is that all these stories, like cursor and clod, are engaged in reselling tokens. That is, they buy tokens, or they have a partnership agreement with Anthropic, and they resell tokens to you. No provider will give you a more profitable solution than Anthropic itself, because only Anthropic can subsidize your token costs, because in the end, everyone pays for electricity and there are no free tokens. If you use free tokens, then you are the product, meaning your data is used somehow, and so on. Therefore, the most profitable offer can only be from the manufacturer. Yes, if you want to get the best price in the store, you go to buy something, I don't know, expired, or you buy something wholesale, or you go to the manufacturer. Therefore, it works the same way here, the most profitable offer for clod. Namely, clod has various workarounds and so on, but it's easier for me to pay once and earn money and not worry about it. So, I've talked about the Max subscription, about where to launch, and about tools. Ah, this shell that I have running is a terminal emulator, and it's called. It works great, I've completely switched to it. Thanks to Andrey for the recommendation. Why is it good? Firstly, because it supports tabs, it looks like this. And I can rename the tabs. For example, I'll open a project here now. So, CD, documents, GitHub and chat. Let's fucking go. And I'll tell it: "Just run the project." And now I can rename this thing. And I don't know, maybe there's some demo or something else here. In general, I have some work-related things written here by project. So, here is one project, and I switch. Here I might have, yes, it launched it. I have a project on next, nothing superfluous, so I'll turn it off now. So, run the project. Background. Here, by the way, when you run some projects, you can use the down arrows, for example, to launch it, open it and kill the process. And we'll kill it now. My running process. So, K to kill. And so, let's do it again. Everything, the service is running. And now in this mode, how I use it, I divide it vertically. Here I also have a hot option command here. And the screen immediately divides. And here I can launch something there? Next, next. Let it be like this. Ah, it will work like this, right? It will work like this. Local host 3000. And it opens. This opens. >> The chat opened. Somewhere I had a history for it. Everything, it's now open for me, it's running here. I can minimize it somewhere else. But the main idea is that here I usually run the development server, locally in the left part the agent, and I can also run several more of these agents, but it's more convenient to run them vertically, that is, command Shift down, and I divide and run the server there, for example, and on the right I run another agent in some such, in such, in such, in such, in such a mode, my work is usually built. So, I'm asking now. And then for voice, I use Wprflow. This is a thing that allows you, I have it set up on the FN button on Mac, just to launch dictation on Mac. M. What is my, tell me, what is my link for the overlay for OBS? Okay, let's try to add it to BS now. Hmm. So, can I test it? I've now added Overls. And in theory, new messages should fall there, ah, new messages should fall there. So, OBS OBS browser OBS. Hmm, well, if not, then automatically. What text? As if, maybe this thing broke again. Well, okay. Ah, in general, in this mode, yes, I caught the chat ID. So, next. So, tools, accordingly, and I've talked about it. 2 is called. Then I use proflow constantly. What else do I use? Probably, well, Super Power Super Power to talk about. Mhm. It freezes on Windows, it constantly freezes. So, but at the same time, did you separately configure settings with permissions for commands, although it's needed if Danger Skip Permissions is enabled. Ah, listen, I don't quite understand. Probably, probably I configured it at some point. Now, in general, I don't configure it. Now everything, everything works and writes. Probably for some specific emps, but I almost don't use MCP. Another question from Laimer. The peculiarity of atyorm is that it suggests writing paths in folders, and you have to type the symbol yourself. No, this is not atyorm, this is generally in the terminal, it writes it for you with tab. But in general, you don't even have to do this if you've never used it, you don't need to do this at all. For example, I asked clodcode to write an extension that launches code in the terminal in the same folder. So, CD user downloads clodangers Permission does this. So, in general, well, to use clodcode in the terminal, you don't need to learn terminals. And in general, it's an unnecessary thing. Therefore, ask it to do something to make your life easier, and clodcode will handle it. I say again. No, no, there are no, as if, tasks that it wouldn't, wouldn't, wouldn't, wouldn't cope with. So, it freezes on Windows, constantly, by the way, fromspor. Well, if anyone has any recommendations on how to use it on Windows, then feel free to write. So, YouTube Studio Help, why don't you want to work, my friend? The chat works here, but here, in this link to OBS, it doesn't show. Okay, we'll figure it out. Accordingly, another tool that I use is, yes, it's Superpowers. What is Super Power? Super Pers is a set of skills that are installed via a plugin. You need to perform two such commands within clodcode to add it to your system. Plugin workpl, this command and this command. After that, you will have, firstly, you will have skills installed in your system. It's probably pointless to talk about skills now. I just talked about them in a separate lecture about skills that I gave, the link was in the chat. And there you can probably do something else, register, and you'll get the recording. for participants in Loom. So what's in this project? In fact, all that's here are skills. And this is the coolest thing. I have great faith in skills, just like with MCPNOP they did it and then everyone started using it. The same with skills. It's a very good thing. The main difference from MCP is that only this description is inserted into your context. But you can actually call them yourself. And I use these things constantly. This is a brainstorm skill that is launched here, for example, brainstorm. And here you say what you want to do. And I want to say that not all messages are shown in my overlay for OBS. >> And then it launches a cool flow where it asks several questions, launches a cool flow where it asks questions. But we'll figure it out.
I will clarify the contact. In what mode do you use overlay? And in tape mode? And what behavior do you observe? So, I observe behavior, that what I have, messages are simply not displayed, not displayed. Behavior, that my messages are simply not displayed. Yes, the broadcast is active right now. I see the message in the main, main window of this program. So, I will now show what I mean. No, I immediately launched the server and simply opened, simply opened the window and put it in OBS, but the message does not come. What is happening now? And the brainstorm skill works both for new features and for the purpose of trying to figure out what is happening? To try to extract context from you, from a person, from your head. Because now the main task in general in AI assisted development, in codewriting, in AI programming, whatever you call it, in any product, there you are doing, I don't know, you are programming, or you are doing something else. It is to extract from the head, since we don't have a neural network yet, you need to transfer the message from your head to this machine. Therefore, in this mode, in this mode, BSR does this very well. I will clarify the main window of the program. What page? And this page is like this. What can you say about AI? I can say that AI is amazing. I am launching code research to understand why SSE is not working on OBS. I understood that he launched it. Is there any point in learning JS, and then frameworks? I don't think there is any point in learning programming in general right now, because you need to move to a higher level of abstraction. If, perhaps, some basics should be studied, like what a variable is, what a function is, but in general, it makes sense to learn some different, more architectural things, that is, how the internet works, that is, more abstract things, just to understand what you are doing. So, you want to make a website, figure out how a website works, that a website is some kind of set of HTML, JS files. And how can this website, for example, be hosted on the internet? I need a domain, right? What is a domain? A domain is an address on the internet that is linked to an IP address in some DNS table. Okay. And what is this? Ah, in general, some things like how to deploy, what a server is. This is useful. Forget about JS, frameworks, nobody needs that anymore. I understand, of course, that many guys may disagree, and that is entirely your right. But that's my opinion, I had 15 years in frontend, and at first I hated everyone who wrote on jquery in my time, because I wrote document.querySelector. And when they wrote dollar signs, I was like: "Yeah, what do you mean, are you crazy?" But now the same thing is happening. And then I realized that for business, how do you make money? For the person for whom you will make a website or what you will do for them, it doesn't matter to them what you made it from. What matters is that it works. Therefore, all this will be decided. In general, just learn to do all this correctly and work effectively, effectively transfer knowledge from your head to this machine. Then everything will be awesome. So, I found the problem. The agent indicated that the mode switch is on, but you said that you didn't switch modes. When you say that, did you add OBS as a browser source in OBS and only watch there, or did you also open OBS in a regular browser? Listen, I opened it in both places. That is, I have Overlay added in OBS, and I also have this page open in the browser for checking. And we are now trying to debug this. Pay attention to this parameter here. This is the context parameter. It is very important because when my parameter exceeds the value, probably 60, I always definitively clear the command context window with Clear. This is also an important thing, by the way, which I didn't mention. This is about clear commands, and there are quite a lot of them. You can configure different things in the config. And I would probably recommend diving into it only after you have mastered some basic things. In general, from what I have done, it is probably only disabling auto-compact so that there is more context. Got it. You also don't see messages in the browser in tape mode. I see messages, but on the website I see messages, they come, but with a significant delay, that is, I see old messages. And in OBS, I don't see anything at all. So, if all this is important, why do these evil HRs make a fuss? I don't know, I don't know why. It's bad to clear context. You need to start from scratch and brainstorm again after clearing. Well, the main idea is that you need brainstorm once when you have identified what your problem is. Brainstorm will write one large file after that. And with this file, you will move to a new session. Okay. Now there are messages in the browser. They come to OBS. How big is the delay? Several minutes. And, listen, well, actually, there are like a couple of messages from the very beginning, with a delay of about 20 minutes. Nothing else interesting is happening there. I asked for it always. Okay, cool. In the idea, yes, after clearing, you don't need to brainstorm. After clearing, you need to do something like a plan or execute a plan. We will get to that in more detail, hopefully. If we debug my error, which occurs. Let's see now. After this, by the way, yes, you need to exit planning mode. You need for yourself to be in the mode. Listen, I need it to work always, because that's its main application. I use it for OBS, so that messages from the chat are displayed on the stream. Are designers still needed now, or have they been replaced? Of course, they are needed. Everyone is needed, and developers are needed, and designers are needed, because nothing has changed. New tools have simply appeared, and a designer who... What is happening now is not that. First, please write the design, as it is written in your instructions. I had an error that the plan mode was set here. I launched the brainstorm mode in plan mode. And after that. And I always try to monitor that this works in exactly this sequence. Because if you let go at some moments, say: "Well, okay," then later, this will be everywhere, you will give concessions everywhere, and then you will still have to redo it. Therefore, the story when you write the design, then write the plan, then the implementation, it works for me in 95% of cases. The story when it immediately starts to bang and write corrections, and then correct these errors, and then say: "Yes, you were absolutely right, and I was wrong, sorry, I'll redo everything, start over." This is an endless cycle, which ultimately takes a lot of time. So, this is sex to describe. Please write the design, the design file, as it is written in your brainstorm skill. So. Please write. And by the way, about designers. I think designers have gotten superpowers. Because, because you no longer need a person who writes code. You can write code yourself. This is the biggest, the biggest joy. So, I still haven't grasped why you don't like it when the 100% result is worse. What needs to be looked at clearly? Yes, the result is worse, because the higher it is, the more context is filled, it's very similar to how a person works. At the end of the workday, you come, and your performance at the end of the workday is worse than at the beginning, because you have a fresh head at the beginning, and at the end you have a lot of information already there. The same with AI. The less information it has, the better it works. Design in OBS. Plan recorded. Well, okay. The next thing I do is write plans. This is a small task now, but it works. This approach works very well for large tasks, where, for example, OBS can work for hours. Here I use something like ultrac. Ultrac only works in code-code. This is a keyword that enables what is called the token volume for reasoning. And this is simply a feature of the code tool through the API. But I usually at this stage, I enable the reasoning mode to the maximum for writing the plan. In planning mode, it will give specific instructions on what and how, where to edit. This plan, once written, you can then take to different agents, even to rock-code. It will do it more or less. This is also a skill. This is also a skill from this Super Powers plugin. Another very important thing I didn't mention is raycast. Raycast is a super cool thing that allows you, in my case, if you saw, quick window management, that is, left, right, well, yes, left, right, some things like that I use. Then everything related to quick hotkeys on buttons, for example, I have ultra, that is, what I write often, or RR launches for me. So, what happened? Yes. RR launches a prompt that I often use. It's just to conduct a comprehensive research using XMCP and insert the current date here. Inserting the current date is very important. Then Superpowers allows you to launch, two options. You can launch either in parallel mode using WorkerThreads. At this stage, it's definitely not worth doing. But Agent Driven launches different sub-agents, can make them, moreover, in parallel. And this is a cool thing that allows you to save tokens very well. And usually at this stage, when I'm solving some big task, my context window is already filled. It's already about 70%. And in this mode, you shouldn't continue working. You always need to clear. And I always ask to do the following. I want to launch this prompt in a new session now. Please give me a prompt to launch a Subagent Driven session. To launch this plan. So, professionals and top-level specialists. All who improved and didn't skip, work much better. All programs with designers and technologies are accelerated, can do more. Yes, yes, all true. So, I'm sitting on a project now and it's exhausting to do it in 4 hours. Adding features is incredibly tiring. You really feel that your skills are not enough for natural logic. Well, in general, yes, but such things help a lot because they enforce discipline for you. And another thing about, about things like MCPs, for example, XII for research, they help you. But in general, about these things, I talk about systematic work with these things and some help, on my course. The link is in the description, the very first link, click it. There are still five spots left. 5 months left and it starts next week. Exactly a week from now, by the way. So, send the source files to make a normal banner for you for the channel, if you want. No, thanks, I don't need a banner or anything. I consider all this irrelevant now. If you describe the process step by step, how can it be described? Tell me about the process. Well, probably starting with brainstorming. After that, the design file is written. And after that, I launch in writeup. The plan is written. After that, I ask to write a prompt, which I will execute. Okay. This plan removes. Okay. And after that, I launch context clearing. I launch the prompt and add sub-agents. The most important thing is that this plan already exists and it is the context state. There are various MCPs that allow you to avoid this, yes, Memory Bank, for example, one of the most famous. You can simply ask, you can simply ask for a plan of seven tasks for acceleration, for fixing, for nothing. Yes, only I will write now, I will use sub-agents. And what is the main feature of sub-agents? First of all, each of the sub-agents. A sub-agent is a full-fledged codewriting session, but it has a clean context, it knows nothing. Therefore, in this case, my master session will prompt each of these sub-agents. They can work in parallel, they speed up the work, and you can also configure them. That is, which ones they use. Minus terminal and minus sessions are these things, glitches. But in this case, I just switch and this is my work mode, I just write the plan, leave the plan, and then switch to another tab. And here I realize that my chat is working, everything, it's working. It's spinning here, showing something, what it's doing. Everything, if something is needed, I'm already in another tab and doing some other task in parallel. And in this mode, I work. The advantage of this plan is that you can, for example, open antigravity and tell them: "Here, I have such a plan." By the way, yes, I haven't gotten used to it yet, but through raycast I also assigned to Thank you, Andrey. Thank you, Andrey. Through raycast I also assigned everything related to, yes, antigravity. Antigravity can also launch the same skill if I open it and feed it the same plan, it will pick up all this and use it. In general, while it's doing it. Each of the sub-agents has its own context. What else? >> What else? What else is interesting to say about them? About sub-agents, yes, probably. A maximum of 10. A maximum of 10 works well. And they also eat tokens well, because you work 10 times faster, because 10 agents can be launched. Not all tasks support parallel work. So, I understand correctly that, in fact, a set of skills, yes, it's a set of skills. Plus, they are also packaged in a plugin to make them easy to install. And there are commands to launch. These, in short, yes, presets. You don't create a new tab for each new session. Usually by projects. I have tabs by projects. Like, there's one working project that I work on for a week, it's in a separate tab. Well, and some things like that. So, have you tried warp? I have warp. I have a year-long subscription left, but I've practically stopped using it. Sometimes I open it to test new, chat GPT, but in general, I don't use it much. Okay, done, plan executed. Mmm, SSE works always. So, and he, I understood that he launched it again. Once more. 3.1. Uh-huh. So, empty. And I'll change it here. So, okay. For now, the BS overlay will still receive messages depending on the visibility state. Updated. So, yes, let's see what else could be. I'll send it now. By the way, this is also one of the life hacks that I use extremely often - copying images to the clipboard. This was one of the reasons why my Mac was slowing down so much. And so, and it helps him, for example, to do this. Ctrl V. This is the current state of my two overlay windows and the transmission screen. Use the skill for debugging. Yes, it probably receives new messages. This is a very good skill he has for debugging, called skill. You can also launch it now. The project is old. Many things are missing in this project, for example, such as Now I will launch another session, I will ask him to do it. Use the skill codewriter to update the documentation for this project. Damn, it writes like this. Documentation is extremely important. Without documentation, the agent will most likely be very, very slow. Therefore, it is very important, what we have codewriter for files. I have GitHub with useful repositories. Repositories with useful skills. Now, yes, here is the skill for setting up GitHub repositories. Therefore, some things can be looked at here, like those skills, for example, cler, it's there. And where should it be put in the chat? In the broadcast chat. >> Let's go. Added debug logs. Now pre-render Overlay. Open the console. Show me what it outputs with the OBS prefix. Done. Coded updated. Line 61. Okay, super. Let's see what's next. BS transmissions. And then we execute what this agent asks me to do. There are no headers here. So, about exo mcp mcp mcp, I am not using them now because they consume context quite well. This is the most important thing you have when working with an agent, and there is a context command that shows you the usage of your context right now. Mmm, here MCPs only use Ex, extremely economical. I used to use KCК 7 Sequential Thinking, but in the end I just settled on XMSP because it's awesome. For example, I need to get documentation for libraries. Libraries. We use them in this project. And I launch research through Ex. It will quickly gather the necessary files, read everything that is needed here in PO and save it, I don't know, in the Docs folder. And after that, you can update codewriter again. And in this case, you will have code as your good map of the area for the AI agent. The AI agent will work more economically because it won't waste time analyzing the entire project over and over again. So, SSE works, initial chat history loading. Question about loading, now I show that if, yes, the problem was. If not, show a screenshot. So, >> look, why is only displayed on the screen? All the latest ones should be. This is research on best practices 2025. And then, then let's go. In this project, by the way, there are large ones. I probably wouldn't use next here, I would probably use, rather, white, or maybe even, most likely, white, maybe something in Python. Ah, I've stopped liking Next because of its quirks. But in general, in general, yes, something like that. So, what's there? Chat, no questions. An hour of broadcast, and now another 10 minutes of debugging and we'll finish. If you have any questions, please write in the chat, I'll gladly answer them and run off. Copy the console to him. By the way, such a thing can be solved very easily with an MCP for browser code, but I don't want to add it here yet. Let's finish the context. Now everything up to 50 usually works fine. Some quick edits can be done like this. But again, ideally, you should install an MCP with Chrome for debugging. And plus, the use of MCP in codewriting is that you can configure them for the project, turn them on and off via MCP Jon, which is already in the project. This is a good story. Another interesting thing that I want to try at the beginning of this year is, first of all, development directly on the server immediately. I think with the growth of agent capabilities, this could be extremely interesting, and possibly successful. I think there are good cases from guys who wrote and shared in the codewriting chat. And what else is interesting to try this year? Well, first of all, scaling agents. Well, here I'm launching one codewriting session for a simple task. It's interesting to see how all this will develop further and, and use multiple sessions for acceleration. And in general, I think acceleration is, of course, a powerful thing. So, reload. Now it should show the last five messages. Let's check. Let's check. I've collected it for OBS, I'll figure it out later. Updating. So, it shows the last five messages. That's true. And here in the browser, I probably need to, I don't even know, refresh it. Ah, yes, it's fine. And the last thing I ask you to do in this context is, please give me for OBS styles that will remove the blue backgrounds, because now I'm showing blue backgrounds on OBS. Checking. Mmm, nice. So, they are just slightly cropped, but we need to continue to develop these styles and somehow configure them further so that everything is good and neat. Something like this will be the plan and here it will probably be normal. Everything. Well, in short, you can work. You can try to test and write a message in the chat and see if it will be displayed here now. Mom, >> what neural networks for writing code are you using now? Clt opc. Will $100 be enough on opus, for example, to write delivery applications? It all depends on the person. It all depends heavily. In general, I can say that it's enough for a lot. That is, a $100 subscription, I think, will be enough. Like, you need to do something extraordinary for it not to be enough. Or just grind every day for, I don't know, 10 hours. Just don't leave the terminal. It also depends on the fact that, for example, if you make a plan, first you double-check the plan, you spend time, then you have, in the end, tokens, tokens can be. Have you heard anything about SGART from Renat Abdullin? I think that's his name. No, I haven't heard. >> So, committed and pushed. So, in general, in general, that's it. My child is starting to go crazy. I'll go get him. Thank you for joining this stream. I hope the stream was useful, and you learned a little more about codewriting. And perhaps, it became a little clearer, and perhaps less scary, to work with it in the terminal. Subscribe to the channel, subscribe to the Telegram channel, join the codewriters group, come practice together. Something like that. See you next time. Where is the end stream button? Here it is.