Transcription
Not long ago, the developers of Claude-code shared with everyone the principles they themselves use in their work, and talked about how they themselves work with Claude-code. The original of this post can be seen on the X network. Here are 13 points that the developers told us about and shared these points with us. So, you can see the original of this post via this QR code. Alternatively, a link to this post will also be in the video description. What do we see here? Someone named Boris, who introduces himself as the creator of Claude-code, tells us that he wants to share, in fact, with the whole world his tips and recommendations, or more precisely, a guide on what exactly he uses in his work. He and his team use in their work. And the person's name is Boris. I want to read "Cherny", but I don't know, yes, how his name is originally pronounced, so, apparently, Boris Cherny. Something like that. I, to be honest, haven't been interested in who exactly the creator of Claude-code is until this moment. And since this post is currently gaining quite a bit of popularity online, that's how I found out about it. Apparently, the person has some Russian-speaking roots, judging by his surname. In principle, if you study each of these points, they are quite small, about 5 to 10 lines each, then we won't see anything supernatural, some kind of rock and science here. That is, no secret secrets are revealed here. In principle, these are understandable points, recommendations and tips that Boris himself and his team use in their work. I recommend that you familiarize yourself with all of this yourself. Read each of these little points. Also, I posted a PDF on my Telegram channel, in which I simply gave a brief summary, yes, of each of these points in Russian. If you're too lazy to read in English, then jump into my Telegram, subscribe to it, you'll see all of this there. Here is the QR code and the post from January 11th on this topic is posted here. Within the framework of today's video, I will not dwell on all thirteen points. I became interested in three points that I want to discuss with you and to which I want to draw special attention. These are points 3, 4, and in his, in fact, notes. Let's go through each of these points. And moreover, I will show on my real project, which I am currently developing, how these points can be applied. Well, considering, in fact, the recommendations from Boris, and let's see how it actually works, and let's look at the effectiveness of all of this. So, let's start with point number three. I have the numbering of the points here according to, in fact, his post, so that you can easily grasp the connection. The first thing he recommends that we do, or rather, the third thing he recommends that we do is to use the OPUS 4.5 model in our work and use it in Finking Mode. What is this? If you are not familiar, let's take a look. We can open Claude-code. Here we open the terminal. Then we run the commands clD, in fact, clD-code in the terminal so that it opens in our project, yes? And then, by entering the command /model, we can see, in fact, the model that is used as the main one, which you have chosen, yes, as the main model. You can do the same if you open, for example, the extension. I recommend installing it. Some actions are not always convenient to do through the terminal, so here in the extensions, select clD, enter, here it is clD for VS Code and install it. Here you will have an install button. After installing, here in the right part, a Claude-code icon will appear, which also launches the Claude-code version as an extension for VS Code. And here, exactly the same, you can see the model and change, in fact, the model that Claude, your code uses by default. This can be done by entering the command / or exactly the same here on the right, click on / click on the switch model button and look at the models that are presented. In fact, I have OPUS 4.5 set by default. Next, the next point, Finking mode. Exactly the same, you click on / and activate the slider here. Finking, as the prompt tells us, is advanced thinking. So, in this way, you activate a function that allows the agent to think more deeply, so to speak, about tasks. This may take a little longer than without this mode. But again, as Boris tells us in his post, in the long run, it gives better, more accurate results in task completion and for agents. Also, what else, what else I want to draw your attention to, it's always better to use the latest versions of Claude-code, that is, always update it. The version can be checked if we enter the command Clot --version. Well, and in fact, here it gives you the version of your Claude-code that you are currently using. You can, this is probably the latest at the moment, here you can write "update the version to the latest" and, accordingly, or how to update the version to the latest, it will give you the command and run this command. Here, in fact, everything is as simple as possible. Next, the fourth point, which Boris talks about, is that they use the Claude-code.md file in their team on a permanent basis and constantly update this file. What is this file, Claude-code.md, and what is it for? When you implement any projects, look here, I have a project open in VS Code. And by the way, if we open the terminal and run Claude here, then here in the tips, you will see a recommendation that, dude, you haven't created the Claude-code.md file, create it. And the Claude-code.md file is an instruction, or rather, a description of your project. That is, if you are developing a project entirely using Claude-code, then the first action recommended is to create a Claude-code.md file, which will describe your project's architecture, the technologies you use, the main points of your project. And, in fact, this file is recommended to be constantly updated. That is, you have made some corrections, some changes to the application. For example, during the day you did something, worked on your project. And then it is recommended to update this file so that the agent understands, yes, what changes you have made, and incorporates these changes into this Claude-code.md file. And this file, importantly, is constantly loaded into the context of your application. That is, when you give a command to your agent, it always has information from this file in its context. It always sees it, and for it, in its mind, yes, there is always up-to-date information about your project. Accordingly, it relies on this information and executes your commands based on this information. I developed this project in a semi-manual mode, so I didn't create the MD file. That is, I gave commands to the agent, so I didn't create this file initially. Let's create it together. Now, to create it, you enter the command init. The command is launched, and the agent will automatically analyze our project, our files, our folders, and will ultimately produce this file. The same applies to updating this file. That is, you have completed, yes, some part of the work, run this init command, and your agent will update the current file and make corrections to it. Creating the file, naturally, takes some time, so here, be patient. By the way, while the development is underway, or rather, the creation of this file is underway. In one of his points, the first and second, I think, Boris wrote that he runs from one to five terminals in parallel in his work. And in the second point, he writes that he also runs 5-10 browser tabs in parallel using the Chrome extension. This Chrome Extension is an extension for Claude for Chrome, for managing tabs that you have open in Chrome. I have already made videos on the channel about how to use the CLD extension, the Claude extension for Chrome, in my work. You can watch these two videos. Here is the new CLD extension, I copy the site. Watch how this works. Well, and in fact, returning, yes, to the fact that you can perform multiple tasks in parallel, you can additionally open a new terminal here and give the agent some tasks in parallel, open several of them, yes, several terminals, and work in parallel in each of them. For example, you can perform a task separately for the frontend, separately for the backend, and so on. I don't use this much in my work, it's not very convenient for me. And I don't have tasks, yes, that are super important for me to complete in parallel. Therefore, since, yes, the project is still connected into a single chain, backend, frontend, it's important for me to do everything in order. So I don't use this, but if you need this feature, you can actively use it. For now, while Claude is thinking and working, a small digression. Now I'm addressing those guys who left comments under my previous videos saying that the Chrome extension is a paid version. Why aren't you talking about it in the video? Guys, all AI tools are paid tools. And this doesn't surprise me at all, to be honest. Do you really think that companies spend millions of dollars to train their models? And do you really think they will give these models to people for free use? These are commercial organizations first and foremost. And the goal of existence of any commercial organization is profit. That's how the law of capitalism is written. Therefore, guys, all those who write this nonsense to me about paid or free, the default thought in my head is that all tools are paid tools. Are there any free ones? You can, yes, give me some information now. You know, there are videos from guys with these cheap clickbait headlines, like how to get 12 months free from ChatGPT. Guys, if you are looking for free tools, go there, watch those videos and look for free cheese in a mousetrap. In my worldview, good tools must be paid for. And, in fact, that's how the world works. Here's my answer to possible questions about the paid or free nature of some tools. Claude-code is a paid tool, yes, if I summarize all of the above in two words. Okay. Meanwhile, look, Claude-code has reported on the work done. It wrote that it created the Claude-code.md file in the root directory, yes. The file includes commands, architecture, code organization, app patterns, and important context. In fact, if we open this Claude-code.md file, we see, yes, that it was created on the left in the structure, then here we will have a description, in fact, of how the application that was created works. I repeat, this Claude-code.md file is always loaded into the agent's context, and it uses it when executing your commands. Okay, let's move on to the next points that we will discuss today. I've already talked about OPUS 4.5 and Finking Mode, that it should be enabled by default. I've talked about the clD-code.md file, how to create or update it. I showed you the init command. Information from it is always loaded into the agent's context, so always keep it in a fresh format if you are not working precisely, yes, giving commands to Claude-code by selecting a part of the code and giving it a command to change it. But if you give it some large tasks. And now let's talk about the sixth point. Plan Mode and then Auto Accept. What is this? So, Boris in his post recommends always using the planning mode. What are these modes? If we open our extension within VS Code again, then here there is a pencil, and if we click on it, different modes of Claude-code operation change. The same can be achieved by holding Shift and clicking Tab. The same. In fact, Plan Mode is a mode where you give a task to your agent, and it first plans the execution of this task, makes a list, yes, of actions point by point. And after that, you can correct, change these points. That is, it first creates a list of actions that it will perform, without performing these actions. And only then, after you fully approve all of this, it proceeds to execute these points. And, in fact, Boris recommends dedicating a lot of time and attention to planning. Planning, planning, planning. And after you are satisfied with what the agent, the points, yes, that the agent recommends to do, after that, put the agent in auto mode, and it will automatically perform all these actions. Let's try to use this method in my project and see how our agent handles it. So, what I will do is show on the project that I have open here. Let me show you what kind of project this is. I have a Telegram app that contains various tutorials on different topics that interest me, that I am studying. These are code, low-code, and AI coding. How does this Telegram app work? When you open it for the first time, the first action you need to take is to watch the onboarding video. That is, without watching this video, I do not let users into the main part of the application. This video shows a three-minute video, it talks about the application so that people don't ask questions later, yes. And it gives a code word, so that people don't skip the video, but watch it completely and then enter the code word into the field. This application is functional. And to open it, you go to my Telegram, yes, you see the QR code on the screen. I've also mentioned this. In the pinned message of my Telegram channel, there is a message. Here there is a "Open" button. By clicking this button, you go to the bot, to this one, and here, in fact, you open the application. What will we do now? What do I suggest we do now? So, here is my onboarding video, yes, we can play it, we can watch it. >> So, here it shows certain information, I talk about it. And after that, here we enter, in fact, the code word from the video. And then we get to the main screen of the application. In fact, if we close the application and open it again, then there will be no onboarding video, of course. What I want to do is, within the current video, make it so that this onboarding video is not this, yes, initial barrier. And it doesn't show up for users who open the application for the first time. But for those, a yellow panel will simply appear here at the top, yes, of a yellow color. This will be an alert that will pop up for all users who do not watch this video. So, look, we will immediately let them into the main part of the application, bypassing the onboarding video, but so that people watch it anyway, because it's important for me that people watch this video, as it talks about this application, so that people don't get confused later and don't ask questions 500 times. Everything is explained there. But, naturally, people often skip this video out of laziness and don't watch it. And so, to somehow motivate them to watch it, we will display a yellow alert here, yes, a warning. And for those who watch it, we will give a bonus and the opportunity to watch, the next, yes, videos in different sections. So now some sections, namely, low-code and coding, have some free lessons and some are available by subscription. Accordingly, I will make it so that, well, I will ask, yes, the agent so that for those who did not watch the onboarding video, even the free lessons will be unavailable. Let's implement this. I hope you understood what I mean now and what functionality I have incorporated. Look, you see? For example, if we open one of the tutorials, there are two free lessons here, and the rest are available only by subscription. So, in fact, if someone tries to watch by subscription, then a red alert pops up here. Let's implement this functionality precisely using this plan mode. So, first of all, I will, of course, launch the application locally just to make changes. So, I've launched the code, I'm launching the front. And here it gives me a link, a local one, which I can use to insert this link into the bot. The bot is created on the Jetbot platform, so I go there. Jetbot is a platform for developing chatbots. No-code platform. Then I will open any of my test bots here. I will write "test" and insert the link to open my app locally into the button here to make changes. I open the bot. I open the application. And here, locally, my app is running, in fact. Let's make sure that everything is running correctly, we'll put "Easy Def 1 2 3" headers here, for example, and we'll see, yes, that the changes have been made. So, we are exactly where we need to be. Then, let's return, so that when I watched the onboarding video, I have a separate field in the database that changed from false to true. So, let's return the value in the database to the previous option so that when I open the application again, the onboarding video is shown. So, let's return to the initial state. So, I changed this flag in the database. We open our application again and see that the video is opening again. By the way, access is denied because, because my videos only open on a specific domain, but this is not critical now, as I am opening the app in a different bot. We don't need to watch the video now, in principle. I've already shown you how it all works. Now let's do all the preparatory things. Let's move on to planning. So, I have already prepared a prompt in which, in fact, I have described in simple human words everything that I want to implement now. That is, what changes to make, in general, everything that I have talked about, I have written here. You can pause and read all of this. Let's first create a plan. Create a plan for changing the functionality. And then I have written the points themselves here. So, I added here, I had spaces, I needed to specify various pages, yes, that I use in the application. And now I just copy this request, open the extension. So, here we enable Plan Mode and just paste my request here and send it for my agent to think about. It's thinking, you see. So, finking is happening here, yes? Thinking. It's starting to think here. It also describes here what exactly, how exactly it is thinking. So, you can observe this. And we wait and see what it will offer us and what plan it will create. By the way, an interesting change. I haven't seen this before. Look, I asked in my prompt to display an alert of a certain color. I wrote: "Display yellow." And it, accordingly, asks me: "Orange is currently used with this code. Which one do you need, dude? The existing one or the actual yellow?" And it gives me the code. It also asked me about sections, which I also mentioned in the prompt. In general, the visual display of this window with clarifying questions looks a bit different. And here, you see, you can switch between tabs. This wasn't there before. This is apparently something they released in an update. It's quite convenient, where to block access. And when we have answered all the questions it has gathered, we submit the answers. As a result, it took a little over 10 minutes for it to think. It wrote to me that the plan is ready, the main changes: frontend - six steps, backend - one step. And here it is described in detail. True, part of it, you see, is in English, part is in Russian. But, in principle, if you want everything to be described in Russian, then you can specify this in the Claude-code.md file, which we created at the beginning, that it communicates with me in Russian, for example. Then it will use it. And here there is a huge, huge number of actions that it wants to perform to implement the task that we have set for it. And it offers us two options, that is, auto accept, yes, as for applying this plan, or manual mode of operation. And if we return to the post, yes, from Boris to the sixth post, then here, in fact, he recommends first using Plan Mode and then Auto Accept in order to apply those tasks, or rather, points, that Claude-code has described, in automatic mode. Now we can actually chat and slightly adjust this plan, add, change. Within this video, yes, since it's educational, I won't get too bogged down here. I'll try as an experiment to try to apply the actions it has written in auto mode, and let's see what we get in the end. The plan is approved, it proceeds to implementation. Here it has created the to-do list, yes, with all the points, and now it will execute all these points one by one. Let's see what we get in the end. To be honest, I am a small supporter of these auto applications of updates. I still like to watch and control what the system is doing. But, you see, I can't argue with Boris, who created the message we are relying on today. I clearly have much less experience, yes, in this. Therefore, I am a little distrustful of the agent for such large tasks for now, because, well, who knows what changes it will make. And at my current stage of knowledge, my knowledge in development, I am, yes, not a super specialist in development, I am a person who studies this topic, knows the basics, and I advocate for the fact that you still need to know the basics and then build tools on top of it, use them. Therefore, I still use a slightly different approach at the moment, when I solve specific tasks, yes. That is, in small segments that I can check, look at, ask clarifying questions, what did you do here? In order to learn independently in the process and gain skills. But, naturally, as AI develops, as I develop as a specialist in development, I may switch to this approach that the Claude-code team itself promotes. Since they use it, since they recommend it, yes, in such open, large access, then it is clearly a working topic, yes, comparing my micro-project to their huge projects, yes, to their Claude-code. This is, naturally, like comparing apples and oranges. It's impossible to compare. Therefore, clearly on their enormous volumes, on which all this works, it clearly works great, it clearly deserves attention. I will test this approach a little later, well, completely, from the moment of application creation to the moment of making changes, and I will definitely share this information with you too. Now, I am using the current approach more for experimental purposes to see how this feature works, how it will cope with its task. Well, guys, about 15 minutes have passed, maybe a little more, and the agent reports. Tasks completed. Here it describes that it has completed them, and then writes, "Restart for testing," and so on. Let's see what it has done for us in the end. So, let me remind you that my current user is now marked as new, who has not gone through the onboarding video. Something is not launching for me here. But let's restart it anyway. Let's restart the front. And now let's open it. Oops. So, it immediately redirects us to a page not with the onboarding video, as we asked, but to the main page. We see that there is a message at the top, yes, with this alert that we asked it to create, and with the text. So, the user will have information flashing at the top, saying, "Dude, watch the video." But at the same time, we have opened access to other videos for him here. Here he can watch videos. Okay. Let's look further. The next task was, we see that, ah, this alert is shown on every page, as we asked. Then we had the story that we didn't give access to the lessons from the Low-code section. Here. But it didn't do this. That is, you see, the open lessons that were initially open for me remain open. That is, people can access them, although I asked to fix this. So, this one task was not done by them. This needs to be corrected. In the AI coding section, the same. A person can go to a lesson and start watching it. Here too, yes, access is denied, because I'm from another. In general, my videos can only be shown on a specific domain. This is all okay. This is not protection from the agent that was made, but protection that I made. So, one task was not fully completed, it needs to be corrected. Further, if we go to watch, we get to the onboarding video page. Everything is as it should be. If we enter the code word "огонь" here, the alert disappears. And it disappears on all pages. Awesome. Look, in general, it coped with the main part of the tasks. Accordingly, we just need to finish the task, ask it to finish the task. That is, specifically, yes, describe it in the same way. Exactly the same way as we did within this task that I gave it completely, yes, exactly the same way, we will remove the points that it has already completed, and also put it into plan mode. It will think, again, come up with a solution, give us points, and we will run it through these points. And everything should, in theory, work. So, it actually coped with the main part. It coped quite well. I'm not looking at the structure, yes, we're not taking into account how it wrote this code correctly from the perspective of architecture, technology used, and so on. This is not the task of the current video. The task of the current video is to show that with AI tools, you can do various things. Naturally, you need to control everything, yes, as I said, above, so that, after some time, the AI doesn't, well, doesn't redo everything so much that it doesn't turn into some incomprehensible mass that can no longer be fixed. But I am satisfied, in fact, I am satisfied with how it coped with most of the tasks. Therefore, to finish this small task that remains, it will not be difficult. What else I want to say, yes, in conclusion, what needs to be done, and this, by the way, is also recommended by Boris in one of his points, yes, of that post, which I am referring to today. You need to check, naturally. That is, after we have made all the changes, we can, we cannot, but we need to launch an agent, which will need to be a sub-agent, yes, which will need to be created, which will check all the changes that you have made, for their correctness, for their operability, and draw some conclusions, yes, based on the check. Or it will say, "Dude, let's make further changes, because this is a weak spot here, here, I would redo it, and so on." And in this way, you insure its main agent against itself and the changes it makes. And either you create a sub-agent yourself, or you can, by the way, also use an agent called Claude-code Simplifier. This is also in the post that was published on January 9th, not so long ago, and which, the Claude team, released into open source. That is, this is their internal agent that they use to, well, judging by the name, code Simplifier, a code simplifier, that is, the agent that they run to further analyze. It analyzes the code that their main agent has made, for the purpose of making some corrections to it. But this is also a topic for a separate video. Perhaps I will test it later and record a video on this topic. Well, let's summarize the video. If the Claude developers themselves share some information with us and share some methods that they themselves use in their work, then clearly they are not taking it from thin air, just to write something, but clearly it is a working topic, tested a million times and used in work. Therefore, it is definitely worth paying attention to and definitely taking something for yourself. Maybe not all points, but some part. I have told you about three little points today, and we have actually seen how this works. We confirm, yes, that the topic is working, so follow it, use AI tools in your work so that they help us in our work and simplify our lives. Of course, jump into my Telegram channel. You see the link and QR code on the screen, it will also be in the description of this video. There I write about code, low-code, and coding. Also, jump into the application that I showed you today, yes, which we modified. It works, it's publicly available. There are also lessons on no-code, low-code, and coding. Some of them are free, some are available by paid subscription. The subscription now costs only two thousand per month. These are super low prices that you are unlikely to find anywhere for the amount of information I provide there. Well, and hit the like button for the video. Subscribe to my YouTube channel. I will also continue to write interesting and useful things for you that I find and consider worth discussing. Write comments if something is clear, unclear, if you want to discuss something, want to brainstorm something. I also track them. Well, and with this, we will finish. Thank you all.