Transcription
Good evening everyone and welcome to a new stream, which is entirely dedicated to the topic of skills, that is, the skills of your AI agents, because this is perhaps the most interesting innovation in AI, well, certainly in the last six months, or even a year. And I would say it's one of the main breakthroughs in the practical use of AI since MCP, yes, because MCP appeared more than a year ago. It was a big, big story, and it also became very important. And then the next big step happened with these very skills, the skills we will talk about today. And what's interesting, by the way, both MCP and Skills were invented by, that is, Claude. And this somewhat explains why skills were adopted so slowly and unhurriedly by all other companies. And in fact, even now, OpenAI and Google have not fully implemented skills, but we will talk about that today. In any case, this is a very important thing and opens up huge opportunities for your agents. >> [music] >> What will we talk about today? So, well, let's start by telling you a little about myself. I am Alexander Svet, a practitioner, enthusiast, and developer. And I am conducting this stream at the Profail school. And until February 20th, many of my AI courses are on sale. For example, on February 23rd, my new course on chatGPT starts very soon. And not just ChatGPT, but everything related to ChatGPT, or rather, how to get the most out of working with both ChatGPT and Codex. And we will have a very interesting adventure. Let me show you this course. It's starting right now. How to get the most out of ChatGPT? There are only three spots left, and we will dive into all the most non-obvious possibilities of ChatGPT. We will create custom GPTs, we will learn to do deeper search correctly, and of course, we will work with Codex. Codex is a coding agent that comes with a ChatGPT+ subscription. And we will actively use it, not only for coding, although we will make a Telegram bot and a website, but we will use it for working with files, because coding agents are actually needed not by programmers. So, if you want to learn to work with ChatGPT and everything related to it, that is, not only ChatGPT itself, but also all the services that come with it, then you are welcome. This course starts very soon. And also, some of my other courses are also on sale now, both on code-code and on practical development. So, in general, take a look. And today we will talk about skills. First, we will discuss what they are, why they are needed, how they work, what services support skills, and how skills differ from, for example, Claude MD files, MCP, or, for example, sub-agents, which also seem somewhat similar to this. We will definitely practice working with skills, meaning we will look at where to find skills and how to install them. I will give you an overview of the most interesting and important skills that, in my opinion, everyone should have. This is an incredibly important topic. We will discuss security when working with skills, because skills are one of the most interesting, but also dangerous things in AI right now. Because essentially, you are just installing instructions inside your computer, and if you haven't checked them, they can be very dangerous instructions. Moreover, it's not that they can be; a quarter of all skills currently on the internet contain malicious elements. So, skill security is a very, very important topic. And at the end, we will create, well, we will talk about how to do it, what skills I create, and how you can also do it. If you are interested in AI, join our chat, and we discuss all these things there. Well, let's begin. Let's start with what skills actually are. In fact, skills, if you've been to my previous streams or are taking my course, I think you've already heard about this concept. But I wanted to gather all the most important things you need to know in one stream, because it's such a new thing, skills appeared only a few months ago, about three months ago, 2-3 months ago. And since it's a very new thing, everything is very confusing, very strange, and not very convenient. It's very similar to how it was with MCP a year and a half ago. It was roughly the same. MCP appeared, and at first, nothing was clear. Where to find them, how to install them? Everything was very complicated. Now it's all elementary. But a similar thing is happening with skills, and you can see how their adaptation is going. But it's still not as good as we would like. So, a skill is essentially a prompt that is loaded gradually. It's not that you give your agent a large instruction all at once. No, it's loaded bit by bit, loading only those sections that the AI decides it needs to look at. So, in practice, how does this look? Look. For example, if we open some typical skill now, for example, the Cloudflare skill, to work with Cloudflare, what does any skill consist of? It consists of a folder where it is located. And from this folder, its name comes. So, the name doesn't come from the file. The file is always called Skill.md. We will talk about how it works now. But just so you understand, the name doesn't come from the file, but from this folder. And what's in it? Essentially, it's a large prompt. And its feature is that it's loaded gradually. So, the model always sees this small part, this description, what this skill is about and when to use it. And by the way, honestly, looking at this description now, it's not the best. Today we will talk about how to write a better description. And the model reads this small description, for example, understands that, ah, I'm working with Cloudflare platform databases now. And accordingly, if it's relevant to it, it starts reading this whole big document. And as you can see, this document also contains links to various documentation. So, the model can read it and think, "Hmm, where is it, for example, specifically about our databases, or something like that, because I'm currently working on a database project." And it will go and find information in these documents, for example, about R2, or other things that are relevant to the specific project. Not all skills look like this. Skills can be very different, but they all have roughly this structure. So, as I said, the skill name, this folder is called and gives the name to this skill. Skill.md is mandatory. This file is where the main instruction, the initial one, is. And then various optional folders. For example, there might be a reference folder with some additional documentation. Well, as we see here, we have a reference folder. And here is additional documentation on various technologies. In some cases, there might be, for example, some script examples. For example, if you look at the PDF Skill, there will be a scripts folder that contains ready-made Python scripts, optimal for working with PDFs. So, for example, to work with PDFs, the agent essentially just needs to run various Python scripts. And in this way, it will, for example, delete some pages from a PDF, or edit the content of a PDF file, or do something else with a PDF. And for example, here there is reference.md, which is another text file with some reference information, but all the scripts are in a separate folder. And in some cases, for example, assets, like templates, diagrams, resources. So, in principle, everyone invents to their own taste. Again, the only thing that is part of the standard is the fact of this structure, that we have a Skill.md file, which is located in some folder. That's basically it. Then people start inventing how they want it and how it actually works at the level of, for example, your coding agent or the program you are working with. If your program, for example, Codex or Anti-gravity, depending on what you are working with, Cursor, supports skills, then essentially it just knows that it has access to these different skills, for example, this list of skills. And it sees the description, this small description of each of these skills. So, for example, for PDF, this small description is this: "Use this skill when you need to do something with a PDF file." And then this is written. So, essentially, with each of your messages, these small descriptions of all available skills are added to its memory. For example, if you have 10 skills, it will be 10 descriptions. If 20, then 20. But in principle, even if you have 100 skills, it won't create a gigantic context for the model. It will be a couple of pages of text. Well, yes, it's quite a lot, but not that much. And it's very token-efficient. And accordingly, the model sees these descriptions, sees which skills are available to it, and starts to choose, "Ah, I was asked, for example, to edit a PDF, and accordingly, I will read this whole file, and if, for example, something relevant to me, for example, there is a description of some script to delete a page from a PDF, I will read these scripts and use only those scripts that are written here." So, in a sense, it is indeed a prompt, it's not something fundamentally new, it's the same text instruction that you give to the model. So, in fact, it's a prompt, if you look at it soberly. But because of this, because of this very efficient loading, meaning you don't constantly bombard your models with huge chunks of text or documentation, no. Because you only provide the most relevant information, and then the model itself decides: "How much more do I need to read? How much more do I need to read? How much more do I need to read?" And because of this, skills work surprisingly well. So, if you take and just stuff this much information into the model's context window, it doesn't work. No, of course, to some extent it will work, but in general, it's not very effective. And all this trick, all this magic of skills, lies precisely in the fact that skills are loaded bit by bit. That is, they start with a small description, then, if necessary, the model reads the whole file, and if necessary, it goes to reference folders, asset folders, script folders, and there can be others. So, everyone invents their own. And why is all this needed? We actually need this to explain to the model how to solve a specific task. For example, how to edit a PDF, how to upload websites to Cloudflare or to Vercel, how, for example, to create motion graphics using, for example, Remotion, a technology. So, a very specific task is taken. Skills are only for specific tasks. It's like domain expertise. Or I always like the example, remember when Neo in The Matrix, he suddenly learned Kung Fu, remember? He stands there and says, "But I don't know Kung Fu." And then, "Whoosh, now I know Kung Fu." And this is a very similar story. The model is also like, "I don't know how to edit a PDF. Now I know how to edit a PDF." So, it's the same story, when the model loads the necessary skill into its context. And it turned out to work surprisingly well. So, again, there's nothing revolutionary about this. It's not some fundamentally new thing. No, it's essentially the same text instruction. And when skills first appeared, many people laughed at it, because, well, they were like, "Seriously, you've just come and told us that you've simply packaged prompts into additional files, and this is supposed to work well." But it turned out to work so well. So, this is how LLMs perceive information, that this gradual approach, when the LLM itself decides how much more it needs to learn, works surprisingly well. It doesn't overload the LLM's context. The model itself chooses how to use the skill. And, well, let's see how it works in practice. For example, I've already created two chats in Codex before the lesson, just so we don't waste our time on it. One is like this, and the second is like this. And to the first one, I say, "Create a web game in the style of Mario with modern technologies and beautiful effects. Importantly, do not use any skills." And then, if you look, what's happening here, it started working. I wouldn't say it did much, it worked a little bit like this and did something. Well, let's see, what did it do? Where is it? It made this game. Let me move the window a bit. It made this game. Well, it's not bad, right? But if you look closely, everything is a bit strange. Well, why are the enemies hanging in the air with the coins? The hero looks a bit strange, honestly. And, well, it's not bad, right? It's not like the model failed, right? But now they've become transparent, and now they've disappeared, see? I jumped on one, and more disappeared. So, it needs to be refined, right? So, this is a typical situation when the model does something, and then you go around and say, "Listen, change this here, do this there," right? So, it works, but not very well. It's really not very good. It has glitches and bugs. Now let's see what happens if I give the exact same prompt, and I just say, "Create a web game." So, everything is the same, only one phrase changes. Importantly, use the "develop" skill. This is a skill built into this new Codex application, which I really like. And it's called "Develop webgame," meaning it's for developing such games. And here's its description: "Create well-functioning web games." And everything is written here about what needs to be done. It's told how to test them. And not just, you know, do something haphazardly, but test it so that it all works normally. And this is a skill that you can use to make simple games like this. And if we now look at how Codex worked with this skill, again, only one phrase changes, right? So, in the first case, I said not to use any skills, in the second case, I say, "Importantly, use this skill." And then, if you look, it tests much more, does much more, right? And don't expect a miracle now. But first, look, you see, there are some bugs here, right? Everything is flickering somehow. Well, it's strange. And let's look here. And here, suddenly, we have a normally looking game, right? So, it's also, well, it's not perfect, right? It also needs refinement, but notice that it generally works normally, right? They don't hang in the air, everything looks more or less pleasant. When I kill them, they start becoming semi-transparent and blinking, right? So, this is what a skill gives, right? It just taught Codex to make such a simple game. This doesn't mean it teaches it to do it perfectly, or anything like that. No, no, it just gives the basic foundations. If I had written a different prompt, maybe a more extended one, maybe with more details, shown references, it would have done even better, of course. But here, I'm interested in showing that it made a working, normal game right from the first try. Not like this one, where, you see, something is flickering here, it's unclear, right? And well, well, this is actually a non-working game. The enemies are hanging in the air, which is wrong. And yes, all this can be refined. And a skill, in fact, helps to avoid this eternal refinement, right? When something is done, and then you start describing, "Okay, please make sure the enemies don't hang in the air," and so on, right? But here, everything is done precisely according to our prompt, as it should be. And there's even an end of the level. Boom, everything is great. So, this is how a skill works, right? Yes, it doesn't perform miracles. You can't expect miracles from skills. You still need to prompt the model well. You still need to show good references, right? It won't do all the work for you, but it will explain to the model how to do it in principle. And therefore, it will immediately do it much better. If you want, you can also make a working version out of this. You'll just spend more time on it. Therefore, it's better to find the right skill for your task, use it, and everything will be great. And work. Now let's see what services support skills. And here it's a bit of a complicated story, because on the one hand, almost everyone supports them, for example, Claude, Claude Code, Codex, Kill, Rot Code, Anti-gravity, right? So, almost all coding agents support this, but there are a few important outliers, let's call them that. First of all, ChatGPT itself. ChatGPT itself does not support skills, specifically skills that you install yourself. ChatGPT has built-in skills, and it knows how to work with them. The same applies to editing PDF, Word files, Excel. And by the way, if you ask it directly, "Hey, what skills do you have?" it will even tell you what skills it actually has. But you can't add them to ChatGPT yet. I think they will change this later. I think they will definitely improve it. Maybe by the time you are watching this, ChatGPT will already support adding them. Well, if you are watching this recording sometime, you know, one day in the AI world is like 7 days in the real world. So, maybe if you watch it in three or four days, there will already be support. But for now, you can't add skills to ChatGPT. Codex has full and very cool support. It supports both adding skills and has built-in skills. So, Codex is great. GPT does not. With Gemini, it's a disaster, because Gemini does not support skills, only Gemini C and Anti-gravity support them. Their coding agents also support them, but Gemini does not. Perplexity, as far as I know, also does not support them. Honestly, I don't use it much, but the last time I checked, it didn't support them. And my response also said it doesn't support them. If this is not the case, please let me know. But I think, in any case, support will appear soon, because skills are becoming a very massive thing. Very, very, very massive. And everyone understands that they work really well, that they are very token-efficient, so I think it will all happen in the end. And now I want to close these game processes so they don't overload my computer. So, yes, so we've correctly understood the services. Coding agents almost all support them. ChatGPT, well, it won't support your own. It has some built-in ones. Gemini does not support them, Perplexity does not support them yet. So, essentially, it's more of a tool for development. Although I disagree with this. I use skills a lot, not for development. And I think Claude understands the full power of skills much better in this regard. It's not for nothing that they invented them, because with Claude, you can use skills in their regular application. Now let's discuss how they, that is, how skills differ from Claude MD or, for example, from MCP or sub-agents. Well, let's go through each of these points and see what's what. Let's start with Claude.md, and I say Claude.md, but it can actually be Agent.md, any other agent rules, Gemini.md, anything, right? So, Claude is just a common name for a file that describes how your agent should work. And all such files are essentially just files that contain various project rules or project or general global rules. And they are always loaded at the start of your agent. So, you start a new session with your agent, and Claude.md is immediately loaded into the context, and it makes sense to describe, for example, what development standards you have, what are your key project rules, how you conduct development. So, essentially, it's a way to tell your AI agent, "Here are my rules and my preferences, this is how I want to work." Then we have, well, let's get back to skills now. Then we have MCP. MCP is a protocol for connecting to external systems, that is, for example, to databases, to other services, to a calendar, to, I don't know, Notion, to any external systems. And MCP, which is very important to say, they are constantly connected and consume a lot of context tokens. The use of MCP is also changing significantly. In some, for example, OpenClaude, a very popular agent, a personal assistant, there is no MCP at all. All MCP is done through skills. So, there is a skill for using MCP. And when needed, the model reads this skill and then uses MCP. It's a rather clever system. So, MCP, due to their extreme token inefficiency, are being used less and less. Because if you connect some MCP, it's always in your context, and some MCP take up a lot of context. Especially poorly designed MCP. For example, GitHub's MCP is terrible. It's really just very bad. People just didn't think about how to make it properly. They made a terrible MCP. And some other large corporations also have not-so-good MCPs, and they take up, you know, 20-30,000 context tokens. This is just monstrous. And there's no point in keeping them on all the time. But it's important to understand that MCP still gives you access to some data, and you can't, well, directly replace, for example, replace MCP with a skill. You can do it if you want, and they do it, but it's not exactly the same thing, because a skill can describe how to make calls to some database. But again, the problem is that, um, for example, if your context accumulates, then this skill can gradually disappear from our context or take up very little space. But MCP doesn't. MCP is always here, always active, always occupying a lot of context, but in this regard, it's more accessible to models. And models can actually use MCP well now. So, it's not that MCP is completely disappearing, but people have started using MCP much more thoughtfully, meaning they started thinking, "Hmm, maybe I don't need this 25,000 token MCP that's always on."
occupies my context, because I will turn it off." That is, MCP is a good thing, an important thing, and essentially, it simply provides access to external data and tools. That is, this is its essence. And sub-agents are a slightly different story. That is, they are specialized assistants, and they are needed to delegate certain tasks. And it's not actually you who will be delegating. It's the main agent that delegates. That is, for example, you tell Claude, or Codex, or anyone. You say, "Analyze my codebase" or something like that. And your main agent understands that if it starts reading all the code now, its context will be terribly filled. And therefore, it says, "Listen, you, sub-agent number one, go read what this project is all about. You, sub-agent two, go see, um, what databases are connected here. You, sub-agent 3, go figure out what's going on with the front-end, right, how everything looks externally, right?" That is, it gave each of them such commands, and each sub-agent went, figured it out, filled its context, and simply reported, for example, "The front-end works like this." Another one returned and said, "And the databases work like this." And it turns out that they filled their contexts a lot, but the main, right, our main agent, received only the key information from them, and everything is fine with it, it understood everything. That is, again, you cannot replace sub-agents with skills. No, you can't, because it's necessary for individual agents to go and fill their context and then report only the necessary things. That is, they are like parallel workers. And so now, I think you understand that a skill is essentially a folder with expertise, and the LLM accesses it itself, or you can tell it to do so. And it accesses it as needed, right? That is, if it understands that this needs to be used, it will use it. And skills work best for complex and repetitive processes. That is, for example, well, I, for example, have a skill through which I manage my Telegram channel. That is, I have such a Telegram channel. I'll show it now. You can subscribe to it. It's called AI Long Reads. And, um, let's open its preview. Here. And here are translations of various cool English-language articles that I read every day. Several translations of the most interesting articles are published. And all these translations are done using a skill, because it's a complex thing, right? That is, accordingly, it's not just about translating, right? For example, you need to have a list of references, for example, how to correctly translate different terms, what style to maintain, right? How to check if the translation is correct. That is, in general, it's a very complex skill, a complex procedure, actually. That is, I always tell everyone that my agent is doing this, and it looks like it's all so simple. No, it's not simple at all. Creating a skill for such an agent is a complex story. And it's precisely about such a more complex story that we will talk about, we will talk about it in my course, which is also starting soon. It's a deep dive into LangChain. That is, here we will, it starts on March 23rd, and we will, first of all, explore all the possibilities of LangChain, right? You will learn how sub-agents work, how it all works, hooks, and other interesting things. We will talk about advanced LangChain features like the ReAct loop and much more. We will create our own LangChain SDK. And one of the homework assignments will be to create your own well-thought-out skill, precisely because, well, I think it's an important skill to learn how to make skills yourself. We will talk about this today, but we will talk about it in an overview format today, and here we will have a whole homework assignment about it. That is, essentially, if we summarize all of this, right, LangChain defines the rules. MCP provides external data, Skills teach what to do with them, right, if it's a complex process, and sub-agents can perform various actions in parallel, including those related to MCP, and skills, and anything else, actually. That is also a typical story. For example, how I do, say, translations, right, for my channel, I accumulate a list of articles that I need to translate, and then I just send them to my main agent and say, "Create a sub-agent so that each of them goes and, using the skill, translates one of these articles." And it turns out why it works this way? Because if I start telling my main agent, right, to just translate these four articles, it will start translating them, and its context will grow and grow and grow. And in the end, it can lead to the fact that on the fourth article, it will insert something from the first article because its context has grown significantly. No, that won't work. For everything to be high quality, each one needs to be spawned separately, that is, a sub-agent appears. Each sub-agent reads the translation skill, performs its small task, and that's it, it doesn't clutter its context with anything else. That is, they all work together, they all work in a unified tandem. And Sergey asks, "Is LangChain like rules?" Yes, it's rules, essentially. For example, in LangChain or Chainlit, it's called Rules. Essentially, yes, it's just rules that describe how you see working with this agent. Sergey also asks, "Can you launch sub-agents yourself, or does only the coding agent decide to launch them?" No, you can, you can, you can create different ones, we will also create different types of sub-agents. And you can, well, you just have to tell the coding agent, right, to launch these sub-agents now. It will launch them. It's just that usually it decides itself that this needs to be done, because it understands that, ah, this is a complex task. It will consume a lot of context. Let me launch sub-agents now. So, this is about how they all differ. Let's not go further yet. I'll answer Dmitry's question. Don't we save the main results of the work and conclusions in Agents ND? And that's also, uh, another matter. Just look, Agents ND, MD, Cloud MD, and others, they are of different levels. That is, for example, in your specific project, you might have in Cloud MD or Agents MD or in any rules written down your preferences, right, what we do, and at the same time, the current state of the project. It's just important to describe it compactly enough, so you don't need to stretch it out like "War and Peace," but you can do it. At the same time, there is, for example, um, LangChain MD or Agent MD, which will be at a global level, not at your project level, but global. And the global one will describe your key principles of interaction with the agent. For example, "never, ever load an unverified project," right? That is, "don't deploy an unverified project." Well, some things like that, and you can write them in these global rules. Now, where to get skills? Essentially, there are, well, a few key places. First, in, well, I always do it, I just search with LLM, right? That is, I just say to the LLM, "Go search and find me skills for my topic," right? That is, it goes, does research, and finds everything. But if you want some verified places, they also exist. First, my favorite place is playbooks.com. What's great here is, for example, if you search for a skill, right, official skills are marked. And most importantly, most of the skills that end up here are verified, right? For example, you see the inscription "safe," meaning it's safe. Plus, there's a "health score," indicating how well it's made, right, how well it's described, because, well, some skills, honestly, are completely talentless, and it's not a fact that they will work well. Someone made them talentlessly, resulting in a talentless skill, and a complete mess. Well, you can also filter, for example, "official only," right? To use only official skills. This is Playbooks, a great service. Then, I think you're familiar with this service, Context 7. And Context 7 has a separate section here. This is, if you just go to Context 7, there's a new tab here: Skills. And this is, essentially, their own catalog of skills, right? And you can also go here, read its description, see, for example, what their trust score is, that is, how much they think this skill can be trusted. I don't know for sure if they verify it. They don't explicitly write about it, right? Although here, they write that it's a verified skill, an official skill. They write this because Playbooks really verifies every skill that has this "safe" badge. And Context 7 is a good resource. I don't want to say anything bad about them. It's just that I don't really understand how safe it is, because, yes, this trust score, it tells us something, right? 89 is probably good. And the checkmark, that's great. But it would be great to see some inscription like "verified." It's really pleasant, it gives peace of mind. And Vercel has its own skills. And they write that they supposedly verify skills, but the wording there is also like, "we do everything possible not to miss bad skills. If we missed one, let us know." Well, it's not very reassuring. So, you can use it or not, but it's also a quite popular resource. There are many different skills here too. So, personally, honestly, I mostly use either the search with LLM to find the skill I need, or I go to Playbooks. So, I mostly use these two combinations. Now, let's talk about the most unpleasant topic: how to install skills. And here, honestly, it's a mess, because it's such a new technology, and there are no good unified rules. They are slowly forming, slowly appearing, but it's still very clumsy in this regard with installation. Let's look at the options. First, there's manual installation. What does manual installation mean? It means that, for example, you go to Playbooks, want to install something, for example, a Code Reviewer, a good skill, it has a good rating, it's an official Google skill, and you click here to go to its GitHub. What do you need to download? Essentially, you just need to get access to this CodeReview folder. But the GitHub's inhuman peculiarity, you know, is that you need to, well, you can use commands like clone, but let's not get into that now. You can just download an archive. Hop. You download it and extract it. And here you need to find where, probably, where our skills were. I've already forgotten myself where they were. Let me see. Ah, where were they? In the Gemini section. So, it says you need to find Gemini Skills. Here, Code Review. Here. If we find this folder, right, in this downloaded archive, we copy it. And then we need to paste it into the system folder of your coding agent. First, they highly recommend enabling hidden folders. And then, depending on what, like Codex, dot, on Mac, for example, dot Codex. And where are the skills? Here. And just copy it here. Hop, done. And it will appear immediately. And interestingly, with skills, you don't need to, uh, you don't need to restart the coding agent, it will load everything itself. And for example, what else do we have? Who else? Oh, there's Kill Code, there's Kaira, right? Kaira. Although Kaira is a separate story, I think. Let's not touch Kaira. It has its own skills. Ah, well, first, Kill Code, right? Just copy it here, or there's Kill Code too. So, essentially, every coding agent now has a skills folder where you need to copy it. If this folder doesn't exist, and it might not exist at first, just create it. Just create a Skills folder in this system folder, copy it there, and everything will work. This is manual installation. There's a simpler option - installation via terminal. That is, if you noticed, when I was looking here, you see, there's a command here, right? What does that mean? Well, first, you can just copy all of this to your agent, and it will do everything, right? But if you want to do it yourself, you just copy it, launch the terminal. I'll launch the terminal now. It can be the VS Code terminal. I use it on my computer. I like it. It's just a nice terminal. Okay. And you just run this command that I copied in it. And it will ask: "Install this Playbook?" Yes, it needs an update, probably. I had it installed. And it found a lot of agents for me, 36 of them. Well, a bit too many, but okay. In short, it found them. And you can install it for all agents, or only for specific ones. For example, I want to install it only for Codex. Where is my Codex? Hmm. Hey, where are you? They didn't find my Codex, did they? Hmm, maybe I'm missing something. No, it can't be that they didn't find Codex. They always, always should find it. Very strange. Ah, well, yes, it, uh, it found it correctly, because it will install them in agent Skills. This is also a universal new folder, right? So, it actually found everything normally. It's just that, you see, Codex Droid, they are moving to a new format where all skills are in agent Skills. That's why, yes, that's why it found it. Okay. And you, well, in any case, I don't know, let's choose, say, uh, what do we have here, cursor, right? I'll choose it, for example, and do it. Well, I think I've selected everyone. Well, okay, it doesn't matter. And then I choose either a specific project or globally. Skills often make sense to use globally, because, um, well, because they don't consume as many tokens, and if it's a universal skill, why not? Okay. And then you just install it. There's also a SimLink thing on Mac. This is when they will all be in one folder and then replicated from there. Well, if it's convenient for you, you can use it on Mac. In short, you just choose where to install it, and it will be installed like that. And the same, for example, works exactly the same way with Context 7, right? So, the same thing, if you go to Context 7, the same thing. Again, you can just tell your agent to do all of this, and that's it. Just copy it here and say "yes." And it will now, accordingly, um, install all of this. It installed it in Codex. It's just a slightly different installation. It immediately installed it in LangChain, cursor, Codex, and the universal agent skills. Everything, it has accordingly, uh, installed it. And that's it. So, this is a pretty simple method. Either enter commands manually, or just tell your agent to execute this command, right? So, with the agent, all of this can also be stopped. Fortunately, times are changing. These wild terminal times are becoming a thing of the past. And in modern programs, everything is much nicer. For example, the new Codex. Look how beautiful it is here. Here we see installed skills, recommended skills. Click the plus sign, and it will be installed. So, it's becoming human-friendly, honestly. And that's great. But the only thing is, it's a bit complicated to add new skills. You can also ask it to do something like "add," well, any third-party ones, "add this skill to your folder." But you can also just copy it manually, but everything that's already prepared is done very conveniently. You see, just click the plus sign, and that's it. Everything is great. And with Claude, everything is also great in this regard. And with Claude, I mean, not LangChain, but the application itself, everything is great. Then you just go to settings. I'll show you now. You go to settings, to the "Keep abilities" section, right? And here you will find all your skills that you use. Now, something is wrong with mine, I'll make it smaller. It's too expanded. All your skills that you use, you can enable, disable, delete. All the skills that come by default. And you can click the plus sign and upload a ZIP file, or create a new one, or, for example, create from instructions, right? So, well, there are different ways here, everything is also quite human-friendly and convenient. So, you just upload a file with the skill here, and everything is great. So, the new Codex and Claude have human-friendly installation. All others are not very human-friendly. If you work, for example, with LangChain or Droid, there's a thing called plugins. A plugin is when you can, for example, install a skill like this through a plugin, and the convenience will be that it will be updated automatically. That is, it will essentially be a connection to the GitHub repository where the skill is located. And if something has changed, it will be picked up immediately. This is very convenient. And Sergey asks about this, that if you download it as an archive, you can't update the skill from the coding agent. Correct? Yes. So, with manual installation, there will be no auto-update. This is a bit inconvenient. But if you are in LangChain, definitely use plugins. Again, in my course on LangChain, we will also talk about this. Now, I think we should take a short break, because we have about 20-30 minutes left. I think it makes sense to take a 10-minute break, and then we'll return and talk about the most interesting skills, and discuss everything else, and create our own skill. So, let's have a 10-minute break. If you have any questions about skills, write them down, and after that, we'll practice with the best skills that I use. 10 minutes. Be sure to come back. Good evening, everyone again. And we continue with practical work with skills. And let's answer a few questions that have come up. First, Bogdan asks: "Without describing how to work with PDF, doesn't the model itself know how to work with PDF?" It generally knows. That is, in general, it understands how to work with it, and it will ultimately fulfill your task. Remember how I showed, right, the example with the game? Well, it made a game, right? Well, a bit crooked, a bit buggy, enemies flying in the air, but the game exists, right? And the same will be with PDF. That is, it will make a PDF, maybe just a bit crookedly. You'll have to tell it, "Listen, it turned out a bit crooked, let's redo it," and it will redo it. In the end, it will probably all work out. You'll just spend much more time. But this way, you just ask once, and it will do everything well immediately, because there are verified skills, oh, not skills, but scripts, which it doesn't need to guess how to create, and typical problems to avoid are described there, right? So, in general, all this unpleasant work has already been done for you. And Alexander asks: "I found a skill for Remotion. How do I download it from GitHub? Or do I need to give the agent a link, and it will install everything itself?" You can do it in different ways, just look at what agent you have. Well, look, here we have, right, this. Let me show you both ways, right? The first way, let's really just ask the agent. This can work with many agents. The main thing is that the agent understands where its global skills folder is, right? So, well, if anything, you can tell it where it is. Well, let's just launch this Codex now and just say: "Install this skill globally." Globally, so that it understands that I want to add this not just to this project, but to everything. And I've set full access so that it has full access, because without full access it won't be able to do it, otherwise it will be locked in its folder. But you see, Codex has a cool skill called "skill installer." So, it has a skill installer. So, maybe others will have more trouble with this. Well, let's see. In short, let it do its thing and let's see. Besides this, how else can you do it, right? Besides just telling the agent directly, you can just download it, right? Just an archive. Let's download it now. It's probably quite large, because the whole technology is together, and therefore it will take some time, probably to download all of it. Let's see. For GitHub, this is a very large size. Hundreds, over a hundred megabytes is rare on GitHub. It's even interesting why it weighs so much. Are there pictures or something? It's strange. It weighs a lot. Let's see how our Codex is doing. While it's also, probably, it's also probably downloading, so it's taking a considerable amount of time. Let's see. Remotion, if you don't know, is a very cool thing. By the way, it's on my list. Here it is. Remotion is creating videos through code, using React, regular code. And through this, you create motion design, animated visualizations, and promo videos. So, try it. It's actually cool. And on YouTube, you can find my stream about AI news in February. And there I show, we create really cool animations with Remotion. Yes, it's somehow huge. So, well, while it's loading, I'll show you how it looks, because if you haven't seen it, it looks like this. Where is it? So, I had something. Hmm, here.
He. Here. All of this that you see now, it is made by code, right? That is, this, well, this specifically Hixfield made, but you too, in principle, can make Motion graphics of the same quality. by code. That is, this, this, all of this that you see now, it is code. It is not generated by Sora or Veo, it is specifically code. And this is very cool, because you can regenerate it every time, and it will be exactly the same, with the small changes you need. This is precisely, among other things. Well, not among other things, this is essentially what everyone uses for this now, because the skill is very cool, but you also need to know how to prompt it correctly. That is, the skill itself will not work miracles. Here. Meanwhile, Codex has found the correct path. I will install it there now. So, while it is installing there, we. Oh, and this one has downloaded. Let's just find where, where we have it here. Probably somewhere around here. Where are the skills? Ah, maybe not in the cloud. Strange. Let's see. Here are more skills, yes, probably here. But you need to find it, and then you can simply copy this folder manually, and that's it. That is, I understand that it is a bit clumsy. That is, I am not arguing, this is like, the reality is like this, well, new technology, and everything new with AI, it is always very, well, not exactly unpleasant, well, inconvenient in terms of usability. But Codor, by the way, also managed. Here, it's ready, the skill is installed. Ah, globally, right? So, let's see where its folder is, where Codex's folder is. It should have installed there for us now. Oh, wait, ah, somewhere here it should have Codex Skills Remotion, right? Here, it's all installed. That is, you see, Codex can be asked like this, and it will install everything. Let's look further, what else do we have. So, the first skill that I really like is brainstorming, right? That is, it turns raw ideas into thought-out designs or concepts. It asks questions and suggests approaches. And this is a very cool story. If you, if you don't fully know what exactly you want to create or want to think through all the nuances, then brainstorming is simply an amazing skill. I practically do everything with it now. That is, for example, well, let's say, imagine that, ah, you want to have some kind of stream. Let's think about what we can do. Well, about skills, by the way, about skills. I want to have a stream, like a workshop about agent skills, right? That is, skills for cloud code or Codex. And I want to think through the concept of how to do it. Use the Skill Brainstorm. And let's think together how best to do it. But before that, first use Tavel and MCP to research what skills are, what is their current situation. Gather information first, and then let's brainstorm all of this with this information, right? Excellent idea. Let's start with research. And it will first go and research what it is, because you cannot, if you are talking to AI about AI, you cannot do it without research. It must go before that. And look, see how many searches it is doing. It is searching for information, so, it is already doing four searches and, well, to find up-to-date information, so that there is something substantive to talk about. Because otherwise, it will start, well, saying all sorts of nonsense to you. Excellent, a powerful base has been gathered. Now let's read the skill brainstorming and apply it to our task. And then the cool part will begin, where you will simply start communicating with it, just like you communicate with a person. And you will compose, well, the picture you need from all of this that I found out from the research. Here is the key picture of Agent Skills as of February 2026. What are Skills? The ecosystem has grown explosively. And this is true, how is the folder with Skills plus optionality structured? Yes. What is the main audience of the stream? Photographers, developers, non-technical people. Well, let's say non-technical people, but generally mixed. Let's have a mixed audience. I think we have photographers, and non-technical people, and developers all here. Understood. A mixed audience is an explosion and an opportunity. We will need to find a balance between the wow effect for beginners and practical depth for those who are already in the know. So, lecture plus demo. I'm showing. That is, it asks, right, what will be the format of the stream. Lecture plus demo. Excellent. Excellent format. Well, honestly, any format I chose, it wouldn't say it's just excellent. Well, okay, nice. The next question, it will ask us now. Meanwhile, I'll ask such a question. Levitskaya asks, if possible, in a nutshell, how to work further with motion in Codex? Well, look, the question is what you want to do, right? That is, for example, well, imagine that I have a folder, right? For example, here is a folder with all sorts of YouTube data, YouTube statistics. Well, these are like open data. And, for example, I will now simply open this folder in Codex. And it can be not only Codex, you can do it anywhere. That is, it is in any coding agent. And, um, for example, let's say this. Use the Skill Remotion to visualize this data in the form of beautiful animated infographics. That is, analyze this data, think about how it can be beautifully presented in the form of infographics and then animate it using and create such a video. Okay. Let's just, it's a bit wrong now, remotion. Like this. Like this. And now it will, accordingly, go and read this remotion skill that we installed. Analyze our file. Here. Understood the task? Here it is reading this skill. In general, let it do all of this here. That is, it will simply go and read this large 76 MB file. Well, it won't even read it, it will read it using code. That is, it won't load the entire file. It will extract key metrics for storytelling. And it will do this using code, here, using Python commands. So, let it do it, and we will see what it comes up with later. Okay, and it asks us: "What stream duration are you planning? 60-90 minutes, a full workshop. 60 minutes, 60-90 minutes is ideal. Enough for theory, and for a couple of serious demos, and for Q&A. Another important question. What tool are you focusing on in the demo part? Uh, it's interesting that it doesn't even offer others to choose from, only cloud code. Well, okay, let's go with cloud code then. Okay, if you like it so much, let's do cloud code. And it's nice that it writes like that, that it knows about me. You know it well, and it's your primary working environment. Last key question before I suggest approaches. Now it will ask us the last question. What should the viewer take away from the stream? The main takeaway. What do you want to take away from the stream? Let's think. Uh, probably to show people what and how to start, to learn to write. I think it's all together. I think we all really want everything together and ambitiously. Okay, I see, it continues to write that I see three options for how we can proceed. Approach one: from zero to your own skill. Recommended. Showcase of power. Masterclass for builders. Well, that means builders in English. Who is building? Let's go with this approach. It looks reasonable. M, we will send it to him now, yes, let's go with this first approach, and in the meantime, let's see what it's doing here. Here, you see, it went. What has it done here? Yes, it has collected statistics. I am moving to the code, adding structures for remotion composition and animated infographic sections. That is, once again, right? That is, everything we are working with is this dataset. Now it will open it, a large one of 76 MB. Now, um, it will probably take a long time to open. Okay, let's not even open it. I think it will load my computer. Okay. Well, in general, a large, large 76 MB dataset. Ah. And then it tells us all sorts of things, right, that, for example, ah, well, let's do it like this. Well, it's not important now. Let's just do this approximately for it. Let it do something here. Here. And it just describes each section for us, right? That is, each section of our workshop, right? That is, we discussed it with it, and it offers us such cool ideas. Bogdan, oh, Vadim writes: "Russians are waiting for three times more tokens for savings. Can you ask it to think in English?" Actually, it already thinks in English. If you look, if you read, it usually thinks in English. Ah, in general, in short, everything is excellent with it. Now I want it to finish, and then we will immediately continue with another skill. Let this one do its thing here. Here. And what other skills are there? Well, naturally, there are all the classic skills, like for working with PDFs, working with Excel files, working with Word files. Another interesting one is FFM Pack. This is a technology for editing video, well, cutting video, making splices. There is also a good skill for it. Skill Creator is a creator of new skills. This is a cool skill for creating skills. Well, we will be creating a new skill now. We will use it soon. And if you work with Vercel and in general with, well, like, make websites on React, then you definitely need these two skills. First, React best practices, right? These are the best, like, patterns, modern patterns of working with React, structures, states, optimization, testing. That is, if you are making a React project, this is the official React Best Practices Vercel made this skill, and it is very good. And Vercel deploy is a skill to help your agent deploy something better, well, that is, upload your sites to Vercel itself. This is the key thing. And also frontend Slides, I like it. It creates HTML presentations. We will do that now. Now, when, um, now we finish answering questions, and we will also make a presentation from what we are gathering now. Let's see what our thing is doing. How? Checking the visuals. Incorporated the main composition. Now I'll do a short walkthrough. In general, it's doing all sorts of things. Used the skill, gathered. Okay, and where is our file? Let's see. And where are you? Vo, there's something. Let's see. That is, once again. Yes, everything we gave it, everything we gave it, we just told it: "Here is our file of 76 MB, right? And here is a skill that teaches you to do visual, ah, motion, like, design through code." Let's see. Well, you know, honestly, very good. That is, imagine, this video it just made itself. Just now, that is, I gave it 76 MB of data, and it just assembled this video for me. It's all done by code. And moreover, based on, well, like, correct, up-to-date data. This, by the way, is really, I even, you know, Codex is great. I didn't even expect it. Cool. That's it. So, a forty-three-second video that completely covers with such nice animations and everything, all of it, all of it. We just took and visualized 7.6 MB of data. Just like that. That's the power of skills. And so, let's look here. Let's, let's, for now, okay, tell it to finish, and let's tell it this. Let's take all these sections that we've come up with, and format them into slides using the frontend slide skill. And now it will read this skill. Okay, here it is reading. See? By the way, it's in English, see? That is, everything it does, here it does in English. That is, it thinks in English, everything. So don't worry, you don't even need to ask. It will do everything in English itself. So. And now it's just describing in this skill how it will make such slides. Okay. And now it will do it. Yes, but of course, it's a cool thing. Just like this, to take and generate infographics out of thin air. That's really cool. Ah, that's it. And now it, let's make, um, ha, let's make a developer console, as if it's a console, you know, like a developer's. Yes, that's what concerns these, well, the skills that I use the most, but this doesn't mean that these are the only ones I use, because I choose skills for each project. If I'm doing some project, I look for skills for it and then use them in the project. It's just that these, well, are really often used, in fact. And what you definitely, definitely should try. Okay, it's creating now. Let's wait for it to finish creating, and in the meantime, I'll answer questions. Ah, Maria asks: "In your experience, what is more effective: modifying someone else's skills for your needs or writing from scratch?" For example, when a skill almost fits for some calculations, but you need to change something for your needs, add your specific instructions. Yes, I think it definitely makes sense to modify someone else's. That is, if in general it works, well, just not quite as it should, right? Just change it and that's it. I do that sometimes. Vadim asks: "Is it only specified in one place? Essentially, it's one folder. Just for each, do you mean when you install it, right? For each, for each agent, well, and for each program, they have their own skill folder, right? That is, for example, for Codex specifically or, well, for Codex, right? Here, skills, that's it. And here all its skills lie. And so on. And so for each one, that is, for Claude too. Here, if we look, where it is, here, clotd also skills, right? That is, everything, everything is the same. Okay, now it will generate it for us, and in the meantime, I'll answer more questions. Ah, Dmitry asks: "Can't you use deep research or connect web search directly to the LLM?" You can use deep research. It's just that for such questions, sometimes it can be a bit, you know, too simple a question for such a serious tool. That is, it won't always work well. And although, also look at what, what tools you will use. You know what my favorite tool for deep research is now? It's, ah, I'll show you. Firew. Ai. Okay. Ah, Sergey writes: "I don't have a skills subfolder in Integromat. If not, you just need to create it." That is, if it's not there, just create it. Here. Ah, look at agent. This agent really searches well. Set it to search five times a day for free, and it will find great things for you. I really like how it works. So Fire is really great, they did a great job. Ah, Bogdan asks: "Tavely or EXA? The latter wrote in a newsletter the other day that it is faster and more reliable than Tavely." Yes, they are really better than Tavely. I agree. Mmm, Tavely gives 1,000 free credits every month, and Exa doesn't. That is, for Exa, if you use it a lot, and I sometimes use it a lot, you will have to pay for it. And it's like, it's cool, but not so cool that I would want to pay for it. Look, that is, it used this skill. And here is everything we discussed with it, right? This is what should be on this presentation, right? Our, well, on the workshop, it just formatted it into these slides. Look, by the way, what a cool effect. Hop. Cool. Ah, that's it. And here, accordingly, you see, it just assembled a presentation for me. That is, we just talked to it, discussed everything. It, well, naturally, it processed it and everything. And now it has assembled this presentation for me. Cool. So, this is how you can use this frontend slide skill to, for example, quickly create something cool like this. Ah, now let's talk about security, because this is, in fact, an incredibly important topic. The thing is, a large study was recently published. Here is its number, well, in the archive of studies. You can read it. It was published in January 2026. They analyzed 42,500 skills that are currently on the internet. And, well, they collected 42,000, analyzed 31,000, and found that among them, 26% have vulnerabilities. 20. Think about it, that's more than a quarter. That is, every fourth skill on the internet has some kind of vulnerability. And the most common is data exfiltration, that is, data leakage from the context outwards, disguised as logging or backup. That is, you see, you might not even realize that something is wrong, right? Your data will just leak when this skill is used, because, well, the agent will just do it like it's for backup, it won't even understand that it's doing something bad, your agent. It thinks it's backing up, because the skill wrote it to send the data there for backup, and it will send it. Or, for example, 11% almost 12% of skills try to get more permissions than they need, right? That is, for example, a skill should do some, um, some simple thing, right? But instead, it asks for permissions, I don't know, to do something on your hard drive, for example. In 0.7%, by the way, a very small percentage of hidden, well, hidden instructions, prompt injections. And I think such a small percentage is because, honestly, modern systems are already very well trained to recognize prompt injections. But to recognize, precisely, this data exfiltration or to recognize this, that it's trying to get more permissions than it needs, that's already harder. That's why I think they dominate for this reason. And plus, there are also, for example, various attempts to infiltrate your, this is called supply chain, that is, for example, installing fake, non-existent packages for you, installing various non-existent dependencies, well, that is, which they don't actually exist, but instead you will install something harmful for yourself. So these are the key things, and this is a real serious, huge danger. That is, if you just go and install skills en masse just like that, it's really dangerous. It can end very badly, because they also do it cunningly. They disguise themselves as various safe skills, as if they are official skills, as if they are some kind of, like, just some skill, like, for PDFs, right? In general, this is a serious big problem. What to do about it? There are two very simple basic security rules. First, use official skills, right? That is, again, if you use this Playbooks, right? Then you can immediately just, well, take and set yourself to official, official only, right? And this, in principle, will already protect you in many ways, because, well, official ones, they, as a rule, are still, well, normal. And also, if you use something unofficial, be sure to check it before use. You can also check here in Playbooks. It has such a Tools and here is such a tool, Skill Security Scanner. And, for example, ah, you just need to copy it like this. You need to copy the path to the skill. Copy it like this, paste it here, and SCAN. And it will now go, scan this repository, scan these skills, and it says: "All safe." Like, sometimes it might say something else. This is, well, like, not a 100% probability, right? That is, like, you saw, right? That many skills here are marked, for example, as safe, right? Mmm, like, but you still need to understand that this is a relatively simple check. Well, that is, it's good, it's not like you can't trust it, no, it's generally good, but if an attacker really, really, really bothers, they can bypass this check a little. And to prevent this, there are advanced security checking tools for skills. That is, if you work somewhere in a corporation, in a large company, you need to use serious tools. Specifically, the best tool is from Cy, it has a skill scanner that uses several things. First, it, well, checks everything, everything, everything like this, well, like, the entire content of the skill. Plus, it also runs it through a service called Virus Total. A very good service. And, mmm, you can get its IP key for free and use it too. And plus, it also runs them through this AI Defense, but here, I think, it's paid, I think. Well, that is, you need to be their corporate client to get this key. That is, if, in general, you work somewhere in a corporation, you are simply obliged, really obliged, to deal with the security of your skills and, well, including through such serious tools as this security scanner. Ordinary people probably don't need this. That is, ordinary people are more likely to just, well, install some preferably official skills and check them before use. That's enough. Ah, and now let's look at what tasks I create my own skills for, right? That is, in what cases do you need to create your own skills and how to do it, actually. For example, ah, I have my own skill that conducts research specifically for my tasks. That is, so that you understand what it looks like. Here. M, now I'll start a new chat, one second. Here, for example, let's say, use AI Research Skill and find me information about, ah, the new Codex application that was released in January-February 2026. That is, what is special about this research, right? It goes through my skill. And this skill searches for information exactly as I need it. That is, not how it could search or how it searches automatically, but exactly as I need it. And what does it do? It performs a series of requests through Tavely. And this series of requests is also based not on simple requests, but it searches, for example, specifically what people are writing, for example, on Reddit about it. I don't need any junk blog posts or anything. I need to know what real people are writing. I need to find out what people are writing on Reddit about, for example, this new Codex application, right? That is, and I don't need, for example, any official press release that will tell me about it. I want to know how it's happening in practice. Here. And it, accordingly, searches for exactly that. Here, if you look, it, accordingly, ah, well, it analyzes official sources too, but it focuses most on what I need. And then it will show everything in a structured report, again, structured according to my needs, right? That is, not just some report, but it has a structure that I want it to give me. Let it do it for now, and then I'll say. Well, plus there is this translation skill, which, for example, translates articles for me. And I have a separate skill that creates good prompts for browsing agents, right? That is, for agents that do things in the browser. It creates good, correct prompts for them, where to click, what to do. This is also a rather interesting thing. And there is a skill that reviews Codex, that is, code uses this skill to run in a Ralph Loop, right? That is, in this continuous review of Codex. I can even send you a link now, because this skill, another skill of mine is publicly available, so if you need it, you can download it. I'll send it to the chat now. I'll send the link now. Here. This is my public page. There are only two skills here so far. One is this skill - it's Codex, and the second is Clot Agent SDK. This is the second skill I created. It is needed to correctly create these agents on the Claude engine. This is what we will be doing in the cloud code course, and you will use this skill for that. And plus, I have a separate skill for generating QR codes. That is, all these QR codes that you see, right?
For example, there, right here, yes, it's just my, well, it generates them for me based on my skill. But I actually have many more of these skills. This is just what you, let's say, will understand without delving into the context, because the absolute majority of my skills are, well, very narrowly specialized. Only I will understand what such a thing is needed for, right? That is, for my specific task. And so, look, what a report it generates for me. So, first of all, yes, it tells me which keywords it found, all sorts of key moments in the history that happened, key, you know, features, and then there's a comparison, and then it actually shows the sources. So, you see, here there are many, many, you see, many posts and, well, official announcements too. And so, for example, there are posts from programmers, right, so, and it forms all this into a good overall picture. I just go through it, I look through it, and I understand, yes, and also I said that the source was always with me. So, if I need it, I just go and read that source further, and that's it. So, I do these kinds of research for myself, right? So, this is my skill, which, in this format, does the research I need for me. And now, let's actually talk about how to create your own skills. And here you just need to understand, yes, why you are creating it, right, because, for example, first of all, there are a lot of existing skills. Look for them first. If you still, well, understand that you need to create your own skill, then, first of all, use a skill for creating skills. It always sounds very strange when I ask, well, like, the agent, to use a skill for creating skills to create a skill. I have the feeling that I'm like a tongue twister: "Use a skill for creating skills to create a skill." And so, this is the first thing you need to do. And the second is to gather, and perhaps the most important thing is to gather the right context, right? So, you just need to think, what should really be in this skill, right? So, not just all sorts of information in a row, but what is truly needed. Well, for example, let's say, as I did, let me just show you with an example how I did, um, now, maybe I even have this chat, this chat history still. Yes. Well, I'm just communicating with it in English here, because I just do all these things in English. But, um, I tell it to use a skill for creating skills to create a new skill and to create an agent that will create agents based on, well, this cloud agents SDK. And then what do I tell it, right, what did I give it, right, I gave it a Markdown file of a very cool article, by a cool programmer, about working with the Cloud Agency SDK, right, so it's a good, really good article. Then I gave it a Markdown file, the official, like, Anthropic documentation. Then I gave it a Markdown file, not of the documentation, but of, like, the best practices, which were also recommended by Anthropic in another article. And then I made him a good quality transcript of one of the developers from Anthropic, who gave a two-hour workshop on YouTube about Cloud Ages Dec. I gave him the transcript. I, so, took it, I gave him four very high-quality tools, so not tools, but context. So, I didn't just, you know, tell him, go and gather everything you find on the internet about this topic and make a skill out of it. No, I selected good context. I understood that I needed high quality, I needed the best articles, the best advice. And I told him all this: "Now combine all this into a skill." And then it made the correct structure of the skill, how it works, what references, API references, patterns, right, and so on. And at the same time, well, it's just combined. So, I use a skill for creating skills, I look for the right context, and plus the right instructions. But the right instructions usually, in principle, this skill finds them quite well. And let's see. Where is it? So, it looks like this, right? Here. So, it describes when to use it, what, when you are building agents using the cloud agents SDK. So, the most important thing is that this, this should explain when, when it should trigger. For example, it writes exactly like this: "Triggers on requests. Build an agent, create agents with the SDK. Cloud agent SDK, right? So, it triggers on this. And then, accordingly, if it triggers, then it starts reading all this detailed information about how it all works, and various references to various, various, various topics. Here. So, for example, how I created such a skill. So, once again, be sure to use a skill for creating skills. So, don't do it manually. Select the right context and add the right instructions to it, so that it, well, sounds right. Well, this first initial instruction should be correct. Yes, we had a small technical glitch. So, here's what this file looks like, which, well, the skill that Claude prepared for me based on my good, correct context, right? So, I used a skill for creating skills, gave it the right context, and, accordingly, it prepared, um, so, this initial, right, instruction, when, when to use this agent, what to use, when you are doing cloud agent SDK, right, and there are trigger phrases, for example, cloud agent SDK or create agent with SDK, so this should trigger it, like, to launch the use of this skill. And then everything is described here, all these best practices from the context that I gave. And, accordingly, by API, for example, best practices, and by patterns, for example, code review patterns, patterns, using tools, so, best, best practices from all, from all the information that I gave him, right, and all together, right, this collects a good skill. So, you need to use a skill for creating skills, give it the right, thought-out context, not just throw it in, so that it just drowns in it. No, give the best articles that you want to be included there. And plus, so that it uses the right instructions, right, so that it, well, this, especially the initial, the first instruction, so that it is done correctly. Now, I will answer questions. Mi asks: "Did you connect search instructions separately via API?" Well, I just have TEL MCP connected. So, I just have it written in my skill: "Use T MCP," and my TNCP is connected. So, um, let me answer another question. Bogdan asks. So, where to get the right instructions? That's a question for you. Well, in terms of, you should think about where to get them. Well, so, for example, as I found them, right, these instructions, I found half of them somewhere, just, well, I, for example, I sit and read various new materials that come out, and bam, I see, oh, a very cool article on this topic, CL Agons Decay, has been released. I saved it. Then again, I'm scrolling through the Twitter feed. Bam, a cool workshop from this Anthropic developer has been released. I saved it for myself. And at the moment when it was time to assemble this skill, I was like: "I have saved ones, I took them out." Plus, I also told him: "Go and search again, maybe find something else interesting." And that's how it happens. So, um, Bogdan clarifies, so the AI won't find it itself, it can find it, it's just that it will likely find more than needed. It's unlikely to make a good selection. So, selection is something you should feel you actually want to include in this skill. Because if you don't feel it, then you probably don't understand what you want to include in this skill and how it should work. Then a good skill won't be created. So, a skill will be good when you have an inner feeling, an inner understanding of how it should work, what this skill should convey. If you have this inner understanding, then everything will work out for you, like. Sergey asks: "Can you create skills for chat gpt research? Put articles there so that it doesn't search and gets the right context?" Um, well, chat GPT, you see, just, first of all, GPT doesn't support skills yet. I think, at the moment when it starts supporting skills, it might be possible. Also, by the way, chat GPT Research, it's now, I think, available, if I'm not mistaken, via API. So, maybe you can make API requests. Or is it Google's, maybe I'm confusing, or maybe both. Well, anyway, one of them definitely. Or Google's, or, um, chgpt has deep research via API. And this is something you can do with a skill without any problems. Make an API call, make these and these searches. So, let's summarize what we learned today, because skills, although it seems like a small topic, are actually very important. So, we found out, yes, what is the difference between skills, cloud MDMCP, sub-agents, because these are fundamentally different things. And we understood that the most important thing is that a skill is a dynamically loaded prompt, a smartly loaded prompt, which saves us tokens, and, accordingly, and allows us to simply, um, get the most out of our AI tools. Where to get them? I most often get them from Playbooks, I just do AI research to find what I need on the topic. But at the same time, be sure, just be sure, it's very important to take care of security, because every fourth skill has some kind of problem. Well, and you also learned how to create your own skills. Sergey asks: "What skill do you use for for slides PPTX from Antropic?" Uh, no, for slides, I used, let me tell you exactly what it's called. Frontend. Now, let me find it exactly. Frontend Slides, I think it's called Skill. Now, yes, I'll find it on Playbox now. I think it's there too. Uh, frontend Slides. Here, frontend slides. And I used it. I like it. It works quite well, but, um, pptx from Anthropic is also good, but it's more for, like, PowerPoint, and this one simply creates HTML files with beautiful, um, with beautiful slides. Here. So, I hope after our stream today, it became clearer to you what skills are, how to work with them in general, and what to do with them. If you are interested in delving deeper into all these stories, then let's see what might interest you from my courses. First of all, very soon, "How to get the most out of chт GPT" is starting. There we will study deep research and many, many possibilities of chat GPT that you might not know. On YouTube, there was an overview lesson of my course. You can watch it here too. I'll send the link now. Watch it, I show many things there, including how to work with code, and how to work with files, and how to work with data. A very interesting thing. And also, on March 23rd, a deep dive into clД-code begins. This is, well, if you work with cloud code, I don't think anyone will talk about it in Russian like this. And on March 10th, practical C development for everyone begins. These are my upcoming courses. So, if you're interested, be sure to come, read what people write about them. You can go here, here are reviews, just read them. These are completely real reviews from people. Here. So, thank you very much for being here for 2 hours. I didn't expect it to stretch for 2 hours, but I hope it was interesting. And most importantly, practice, start using skills in your work. Goodbye.