Transcription
Hello everyone. If you have installed Claude and are disappointed with the results of its work, then you have most likely configured it incorrectly. In this video, I will show you how to better configure it for your tasks. Let's go. And the first thing we see, accordingly, is to open our project and launch Claude. That's it, Claude has launched. Accordingly, what do we need to do first here? Well, and what can we look at? Let's start with basic settings. or rather, not even settings, but commands that exist. A small overview of the main ones that you will need. The first is status. It will show you what is installed, what is not working. That is, well, basic information about your account, which you can view. We can also scroll right and left with the arrows, this is config and usage. Accordingly, config is the basic settings that you, as a rule, configure during installation. And the second is usage - this is how much you have spent, how much you have used, and how much you have left for the week, for the current session specifically. All this can be viewed again. We will also go through additional settings here. The only thing that should be enabled is thinking mode. You should have the setting true selected. Accordingly, to select it, you need to press Enter. The next thing you can also choose here is the language, but it's better to leave English by default, because it will communicate with you in English. Well, and in principle, that's probably all. Let's move on. The next command we have is s memory. This command will allow us to configure the project memory, Claude's memory about you, about your project. Accordingly, what it needs to take into account. The first is user memory. It can be configured superficially, for example, by writing "answer me in Russian" here, for example, or in any other language you like. We save. Okay. Next, well, accordingly, you can also configure not only the user, that is, the user is you. And you can configure, for example, memory for the project only. Accordingly, I write websites. I write a website. That's it. Well, we've configured it, for example, like this. This is just a basic superficial demonstration of how it should look. These files should be fully formatted to your needs. And, accordingly, then Claude will take this context into account in order to work specifically on your project. I will delete this because I don't just write websites. That's it. And we won't save. Let's move on. Accordingly, there is also an option to use Claude through such a chat. In terms of functionality, it is a slightly cut-down, cut-down version of the terminal. Accordingly, almost everything is the same here, but some settings, you see, there is a square. This, accordingly, means that most of the settings are configured through the terminal, that is, through such a field. The only important setting that you can find here and which is convenient to find here is General Config. And in General Config, you have such a setting. This is, in fact, bypass. That is, Claude will ignore any, uh, permissions given to it, or rather, not ignore them and not take them into account. That is, it will work completely autonomously. This is such a ymode, in fact, when you allow Claude to do everything it deems necessary. I want to say right away that this is a good setting if you want to do something quickly, for example, and, uh, for a limited number of prompts. But in all other situations, this setting will rather hinder you. Therefore, I do not advise enabling it, but as it were, know that this is a rather important setting that allows you, uh, to essentially allow Claude to do whatever it wants on your computer. That is, it will not need to ask: "Can I save this file?" It will not need to ask: "Can I delete this file?" It will simply do what it deems necessary. But if you have some old project, for example, and it's already working, then it can break everything there. Therefore, please, well, use this setting yourself. What else is there? In principle, nothing else that would be useful to you. This setting is not always pulled from here, so we don't pay attention to it. Everything that is here, by and large, does not affect your work in any way, since, for example, the terminal can be launched like this, the chat can be launched like this, so, by and large, you don't need this tab anymore. Let's move on, by and large, to the main information for this video. This is what you will need to use Claude to the maximum. The first and most important is agents. Agents are, in fact, already, uh, ready-made instructions for each specific task that you, for example, solve with your Claude. Accordingly, it is better to create agents custom for yourself, but keep in mind that these agents, uh, should be a limited number, that is, there is no point in inflating them. For example, let's take as a basis that I want to create a website for I want to create a website with flower delivery. Let it be so. What do we have here? We create a new agent. Here you can choose whether it is for a specific project or for you in general. That is, well, I advise doing it for a specific project, because, for example, today you need a website, tomorrow you will need some kind of calculator, and those agents there might not suit you. We select the project. And there are two options here. The first is to configure manually, but I don't advise it, because, well, it will be a bit complicated if you are a beginner. Accordingly, you can generate it using Claude. This will be much easier. I want to create an agent that will select a stack for my task. For example, it will be an online store. I need an agent that will help me decide on a stack for my website. Accordingly, we click, we send. Now it will generate it for us. I'll tell you right away that it's better to create several agents, to build a structure for your project. Imagine that you have, for example, a small team. Accordingly, the first person, the first agent will be the one who, for example, is the architect. So, we generated it, removed its edit and launch, for example. Well, of course, if you don't understand, it's better to leave everything as it is. Continue. Model, look, choose based on the task, since the architect is a rather important role in this case, he can be given a smarter model. And, accordingly, we choose some color, let it be green. Next, we choose where to save all this. Let it be the project folder. That's it, we press Enter. Our agent is created. Here it appeared here in the folder. And here is its instruction. Let the next agent be code review, which will check our code. Let's describe it. It should check our code for security and quality of work, so that there are no authorization tokens that should not be there, and so on. Let it think for itself. That's it, we launch it. It was created. We also leave it as is. Continue. Let the model be saet, because such a task is more routine than aimed at some super-skills. Let it be saet. Well, let its color be red. We enable it to have memory. Okay. And let's say, for example, let there also be a debugger. Create a debugger agent that will find and fix errors in the code. We click continue. Uh, in principle, bugs are also such a thing. Let it be a smarter model, but in fact, it could have been given a regular model. Everything is okay, we saved. Here we have created three agents, I will attach them to the Telegram group. Accordingly, you can go, subscribe, download them. Accordingly, we have created agents. What other settings do we have that we could, in principle, consider? Let me show you. There used to be such a setting as MCP. Accordingly, these are additional services that could be connected to our Claude code, which Claude could use at its discretion. Now there is, in fact, a section with plugins, which significantly simplifies this process. And, in fact, the MCP section is no longer needed for us. But I will explain. Basically, all the most necessary, most popular tools that were connected through MCP have become available through plugins, and installing them has become much easier. That's one. MCP is now only needed for services that are not popular or, for example, are simply not available, for example, in the basic plugin marketplace. Here. That is, if you have some custom, uh, product, which has MCP, let's say it's Hrefs for sessions or something like that, this tool might simply not be in the plugin marketplace. And in that case, yes, we can connect it as, uh, as before through MCP. Here. But at the current stage, we have plugins, and installing them has become much easier. Let me show you what it is. We open it. And there will be three tabs. These are already installed. These are what I have installed. We will show this a little later. Marketplaces are, in fact, a collection. That is, it is a collection. So, well, if we take games as an example, these are some packs, for example, which contain a lot of plugins inside. That is, you install a marketplace, and then from this marketplace, you selectively pull out the plugins you want to use. I have installed several additional ones. I will attach them below the video and explain what these marketplaces are. Basically, these are design plugins that help, accordingly, create a more interesting design than Claude did without all these plugins. I have this Superpowers. This is an additional tool that allows Claude to think a little better. Here. But, accordingly, everyone needs to look, some will like it, some won't. Let's look at the plugins themselves, what's here. I have installed context 7 - this is a service that, in fact, has documentation for many frameworks, sockets, anything you want. Accordingly, it helps Claude to better search for documentation, which allows it to work better. And there are two plugins for design. This is frontend design official from Claude and an additional one that was recommended to me in the chat. It is also, accordingly, involved in generating good design. Here. And Superpowers is what allows us to additionally enable such a brainstorming mode within Claude. What else is there? Code review is, in fact, an additional agent, which can be installed as a plugin. If, for example, you are not satisfied with the quality of the agent you already have, then you can install additional plugins that are already tailored to the instructions. The instructions themselves, what's inside, can be viewed on GitHub. I will provide the link. The next is GitHub. It allows you to connect to Git itself and, accordingly, work with Git. What is this? How and why it is needed, I will show in the next video. There is Play - this is a plugin for browser automation. It might also suit someone. Here. But I have the ones I showed. Here. Let's go back. Those that we have already installed. And let's show how to use them. We exit with escape to the chat and start s. And let's, for example, well, a flower shop, let it be a flower shop, let's create some design for our project. And that's it, we launch. And now we will wait. Here it will ask for access. Accordingly, if you don't want to confirm them every time, choose the second option. Uh-huh. So, it has been generated. Let's look. We open it. Here, in principle, such a design is ready. I would say that it is very good. The only thing is that it cannot generate pictures, so it substitutes emojis somewhere, well, and puts placeholders somewhere. Here. Well, without pictures, yes, I agree, it's not as beautiful as with pictures, but nevertheless, it's quite good. Let's look at the second card. Uh, well, the picture turned out to be large, but in general, it's not scary. You can ask to correct the bouquet composition. Listen, it's quite nice, in my opinion. Reviews bouquet, reviews, bouquet skins. Well, in general, not a bad design, I would say, so, mm, okay. I, in principle, like it. Let's try to ask it to generate a second one. This is what we generated frontend design. This will be our next one. This is a design for two pages of the site. Let's see what the prompt was, because we had the opportunity to fully evaluate, so to speak, the work. We send the same prompt. So what does it represent? Uh-huh, these banners are shaking. Well, okay. It chose different colors. Well, in principle, it's very similar in blocks. Uh, probably, the card design here is even more pleasant, as it were. Let's open the second one to compare as well. Mm, well, it's maximally similar, actually. Uh, I even, to be honest, don't see much difference. Well, the only thing is that through the official Anthropic plugin, the design is slightly better than here. Well, in principle, it's good everywhere. Well, the only thing is that there is such a bug. It can also be fixed. Uh, well, in principle, okay, right? Let's move on. That's it, our designs have been generated. And now let's move on. Let's now check, first of all, how much text our current project occupies. And let's see, yes, what we have. Look, the current context of our project is 50%. Well, here, accordingly, there is full detail 62 and 124,000 out of 200,000 tokens. What can we do with this? First, let's enter the Init command for now. This, accordingly, is that we create a Claude MD, which will record the detailed history, the detailed stack of our site. Let's create it. Uh-huh. We see that it has created such an instruction for our current project, but we only have, as it were, two design versions. This is not enough. Let's now turn to the architect. and select what stack for our online store will actually be needed. Use the architect agent to select a stack for my website. I want to create it entirely using and on web coding. Let the hosting be beget. Beget. Well, in principle, that's probably all the input. Let's ask him and ask me questions that you will need to select the stack. A important point is that such an instruction, to ask additional questions, greatly helps to find common ground with the neural network, if one can put it that way, because the neural network does not always understand what you are telling it, that is, well, and we, accordingly, do not always form our thoughts correctly. Therefore, if something is unclear, for example, you missed some detail, it can really ask this question and, accordingly, give you a more accurate instruction than if it did it without this, just at its own discretion. Let's press Enter. So, here's the list of questions it sent us. Let's write. They are very convenient, of course, to look at all this in the terminal. The first hosting is simple, if needed, we can switch to VGS. Well, it's better to share. This is the second question about the type of store. A full online store. Payment let it be yukassa. Fourth, what do we have there? Uh, catalog. In the catalog, we will have, well, 500,000 products, for example. This is quite a lot for an online store. Admin panel, yes, an admin panel is needed. Customer accounts. Well, let's have them too. Seventh. Integrations, yes, they are needed. Eighth, you can answer, if you don't understand what you are being asked, you can answer, accordingly, just yes or no, or I don't understand at all. It will clarify this question, or, accordingly, simply skip it. SEO marketing, yes, I need an SEO friendly stack design. Well, I like the first one more. Let's scale. Expected traffic. Well, a maximum of 10,000 people per month. Eleventh is your level 100% coding, because I don't understand anything at all. Takes deadlines. Skip this question, because, well, and 13, yes, SMS and notifications will be needed. That's it, let's launch. So here is the proposed stack for web development, PHP framework. Okay, let it be Laravel. And in the next videos, we will try to create a full-fledged website on Laravel. We will see how easy it is to assemble an online store yourself using Claude. And, accordingly, well, we will deal in more detail with deployment and so on. For now, we return to the topic of our video, which is configuration. Accordingly, we have configured these agents that help us simply work with our project. I advise keeping them no more than five to a maximum of ten, but this is, as it were, for your needs, so look for yourself. Well, let's analyze the context. I have already shown the main context setting. What do we have now after one more? Uh-huh, 71%. Okay, that's quite a lot. We have two settings. The first is clear. It completely clears the chat history. Accordingly, everything that is not in your files will simply be deleted. Such radical cleaning does not always help us. And there is, accordingly, such an option, it is compact. Let's do it. Accordingly, what will be done? it will analyze your entire conversation, make a summary of it, and, accordingly, it will remain with this summary, and everything else will be erased, and it will continue to work based on the context of the summary it has left. Accordingly, when Claude starts to answer you strangely, to lag, or to forget what you were talking about, it means that its context window has become clogged, and it is advisable to clear it. Also, if it overflows, then Claude has a setting that it clears it itself, it does all this compacting itself. Uh-huh. Here it has compressed our file. And let's now check the context. The context is now 18% occupied. Additionally, I would like to show this. This is Shift TP. Below you have a setting. This is switching between modes. The planning mode is needed when you want to start some large-scale task or, for example, you have accumulated many errors and you want to process them all at once. Accordingly, it is better to use plan mode first, so that it goes through everything, checks everything, makes a detailed list of what you will do with it further. And this will remain as, uh, a ready-made plan for the next agent, for example, some debugger, to go through and fix everything. The next mode is, well, automatic correction acceptance, it is called like this. And the next mode is exactly the one that will confirm each correction. These can be switched. Well, that's probably all. We have reviewed the main settings. If anything is unclear, you can ask questions in the comments under the video, subscribe to Telegram, I also answer questions there. And if anyone has any additions, something I haven't said, you can also add it under the video. And thank you all for your attention. In the next video, I will try to explain how to work with GitHub and why it is needed at all. That's all, bye everyone.