Transcription
In this video, we will break down a very interesting tweet from the creator of Claude-Code. For those who are not yet aware, Claude-Code, which we all use, was created by Boris Cherny. I hope I pronounced his name and surname correctly. In general, on January 2, 2026, meaning very recently, he released an interesting tweet where he shared 13 of his techniques on how he uses Claude-Code in real work, because he genuinely uses it in practice. Once again, guys, in this video, we will break down 13 techniques from the creator of Claude-Code himself, on how he uses Claude-Code in real practice. I will try not to drag this video out; we will stop in more detail on some points. I will show them, so to speak, more practically, how to do them. Some points we will simply discuss and move on. In general, I hope it will be interesting. So, let's go.
So, this tweet was released on January 2, 2026, from Boris Cherny. I will leave a link to this tweet in my Telegram channel. There will be a separate post with this video. You can also go and familiarize yourself in more detail. Here he writes that he, Boris, is the creator of Claude-Code, and now he will share thirteen techniques that he uses in Claude-Code. He also writes here that this setup can be very vanilla, meaning very simple, basic, but in reality, there are interesting things here. In general, let's not delay.
The first thing is that he opens five instances of Claude, and they run in parallel in five terminals. How does this happen? Everything is quite simple. Look, go ahead, just open any of your terminals on your computer. I already have two terminals open, right? So, for example, you type Claude here. Let's press yes. Okay, I have some minor errors here. Now I need to do it like this, right? Here we opened Claude. Here we open Claude. Here, continue without these settings, right? We already have two, respectively, instances of Claude-Code running.
Next, for example, you want to make it three or more, you click here and simply open a new terminal window, respectively, and in the new window, you launch Claude again. Again, I have some minor errors here. Right. And, accordingly, you simply open the number of terminals you need and launch Claude. In this way, for example, you can work on one project where you need a backend and a frontend. So, in one terminal, you work on the backend, in another terminal on the frontend, right? Or you work, for example, on different tasks. This was the first technique from Boris Cherny. So, he opens one or five tabs, meaning five terminals, opens Claude in each, and works on some project. He also uses a notification system to know when he needs new input, in which terminal, right? So, this was the first technique. In principle, this technique is quite common. I know a significant number of people who use both five terminals and ten, right? So, they launch them, and their agents simply execute independently, respectively, write code. It's quite interesting, quite cool. You really need to learn to use this gradually, because it greatly speeds up your work and your workflow.
Let's also move on to the second technique immediately. It's very similar to the first because it also runs 5 to 10, respectively, Claude instances in parallel on Claude AI Code. That is, he runs it in Chrome. So, once again, these first, say, five sessions he runs locally in the terminal, and here he runs them in Chrome. In principle, it's all the same, just, respectively, different places where the instances of Claude run, respectively.
Okay, let's move on to the third point, to the third technique. Here he describes that he uses Opus 4.5 thinking for everything. Of course, he writes that this is the best coding model he has ever used. I completely agree with him here. A little later, I will show you how to enable this thinking mode, because it really makes a difference. Of course, he writes here that it is slower than GPT-4 and similar, but believe me, or rather, believe Boris in general, then me, first of all Boris again, then me, that Opus 4.5 in thinking is truly the best model for coding that this world has ever seen. Purely, again, my opinion and, respectively, Boris's opinion. Let's go to Visual Studio Code and I will show you how to use this model. When you launch in thinking mode, for example, Claude is on the right. Here, first of all, you need to go to the model command. Here you can select the model. I have Opus 4.5 set by default. But how to enable Thinking Mode? Look, if you have some old updates, version up to, say, 2.0, some beginning, I don't remember exactly now, I won't make it up, you will have something like a clock here. You click on it, and thinking mode is activated. But I recommend everyone to constantly update Claude, and in the new update, first of all, in Claude-Code, secondly, in the extension, here, if we click here, and there will be, respectively, a toggle for thinking. It is currently off. So, we are using the regular Opus 4.5 model. But if you want to use the Opus 4.5 Thinking model, which Boris uses on a permanent basis, then simply turn on this toggle and use it always. Of course, your model will think longer, your answers will be longer, but overall, it really changes a lot. Therefore, I never turn off this toggle and constantly work in thinking format.
And the fourth technique is about Claude MD. Here he says that they have, so to speak, one Claude MD per project in their GitHub repository, and they share this Claude MD between teams. So, roughly speaking, in short, Claude MD is a file in which all the information about what you are building about your project is stored. Now I will show you how to create it, right? So, for example, you finish a coding session, and Claude-Code itself updates Claude MD with what it has added, what needs to be done, the project description. In the next session, you launch it, right? In a new window, say, the next morning, figuratively speaking, Claude-Code first goes into the context, respectively, this Claude MD file, and using this file, it is already in the context of what you are doing. So, in reality, Claude MD is a super important file, and it needs to be initialized in every project because it stores the entire context of your project. Claude-Code, respectively, can quickly get into what you are doing. In general, they use Claude MD. This is a very good practice, it is necessary to use it in principle. And they have it, so to speak, single, and it is shared, respectively, through the GitHub repository. They are constantly adding something, deleting something. Right.
Let's go back to Visual Studio Code again, and I will show you how to initialize this Claude MD. Everything happens automatically. You just type the command `init`, right? And your Claude MD is initialized. So, you wrote some code, or rather, Claude-Code wrote some code for you. Then you, respectively, type `/init`, and your project is initialized, meaning Claude MD is initialized, which contains the description of your project. Here, purely for the video, I've put together this YouTube agent that comes up with ideas for videos, just for demonstration purposes, so to speak. And here you see, we have Claude MD, just a Markdown file with a project description, which includes Project Overview, Commands, Environment Setup, Architecture, Claude Style. Because in this project, I have, so to speak, one Python file, I don't have such a large Claude MD. But if your project is larger, you will have much more information in Claude MD, and of course, you will have many more, respectively, lines and more context about your project. Therefore, guys, Claude MD is needed in every project. And Boris, with his team, of course, has it. And this is an excellent technique.
And the fifth technique from Boris during code review. For example, Boris, or rather, a colleague from his team made a pull request. Boris reviews the code and sees that there is some inaccuracy, some error, right? For example, Boris immediately tags Claude right there and writes: "Add to Claude MD, never use enums here, always use literal unions, right? For example, as he says, here are examples using a special command that is installed via `/install github action`, Claude understands that it has been called, it reviews, respectively, this code. It opens Claude MD, adds to Claude MD, never use enums, always use literal unions. And commits this change, meaning during code review, it is quite convenient to use, but note that here too, there is work with Claude MD. And again, Boris emphasizes how important Claude MD is, so never neglect its creation, its maintenance, its use, because it's not enough to just create it at the beginning, right? Many people create it, but it's important to maintain it all the time and, so to speak, always add something to it, write something in it. You can create separate slash commands for this, so that after each correction, Claude automatically adjusts Claude MD. Right. But, in general, it needs to be maintained.
The sixth technique for working with Claude from the creator of Claude-Code himself is that sessions, right? So, most of Boris's sessions start with Plan Mode. First, he plans, right? Some task. Here he says: "I want to add progress notification rendering for skills." And he asks, "Can you make it better and better for subagent progress?" In general, look, first of all, Boris enables Plan Mode. I will show you how to enable it in Claude-Code now. And he first plans, plans, plans what he wants to do. He can spend a long time planning, sitting in Plan Mode, correcting something, removing something. After he likes the plan, he has read it, fully understood it, he likes it, he switches to auto-accept mode, and Claude-Code, as he says, usually, from the first time, provides him with quality code. And here he adds that a good plan is really very important, with which I maximally agree, because if you don't have a good plan, then Claude-Code likely won't be able to adequately implement some feature or some function, respectively, in your project.
So, let's go back to Visual Studio Code. How to enable this Plan Mode and where is it located? Look, if you are working in the official Visual Studio Code Extension. Here at the bottom, there is "Ask for Edits" mode, "Edit Automatically," and "Plan Mode" - this is exactly the mode Boris was talking about. You can also enable it using Shift and two tabs. That's it, you switch to Plan Mode. Honestly, I also use Plan Mode on a permanent basis. Before doing any feature in a project, regardless of whether it's big, small, complex, easy, simple, it doesn't matter at all, I first switch to Plan Mode, describe what I want to do, and Claude-Code in Plan Mode thinks through a clear plan for implementing, respectively, writing any feature. When I like the plan, I, respectively, confirm the plan, but I switch not to auto-accept mode, but to manually accept mode, because I still look at the code, try to look at the code that Claude-Code writes, sometimes I don't like something, and I ask it to correct it. But, in general, look, Boris, again, emphasizes that it is a good practice, before writing the code itself, using Claude-Code, to first plan it clearly in Plan Mode. And, honestly, Plan Mode really works very well. I like it very much.
The seventh technique from Boris is the use of slash commands. He says that he uses these slash commands multiple times a day, meaning many, many times a day. And, respectively, this saves him from constantly typing the same commands, because he uses them often. In principle, commands are a really very useful tool. He gives an example that he has his own command to commit, push, and create a pull request for code on GitHub. This is a really very convenient command. I believe that, in principle, anyone who is actually involved in development should have such a slash command that, respectively, pushes your code to GitHub, right? This is really very convenient. You can also come up with a large number of other different commands. For example, add a command, I don't know, for example, `short`, and there you have some prompt, like, write a clear, understandable, short answer without fluff, right? So, you don't need to type, for example, 10 words, you just type one command, and, so to speak, this prompt, which is ten words long, is attached to your request. In general, slash commands are very convenient, of course, they deserve a separate video. Right.
But in short, to create a slash command, you can, of course, open the config, do everything yourself, but honestly, I don't strongly recommend doing it, because everything is constantly being updated there. And, honestly, it's much easier to just tell Claude: "Create me a new slash command for making a pull request on GitHub." Just write that, and, respectively, Claude itself will create it, it will make, respectively, an entry, correct, rather, the `claude_commands` file, create a new command there, and you will have, respectively, a new command. You don't need to open anything, search for anything. Right, as you can see, `create_pull_request`, `create_pull_request_for_current_branch`. And here is the instruction on what to do with this command. Again, commands can be completely different. This is one of the examples that Boris provides, what command he uses on a permanent basis. I know many guys who use a large number of commands, and they really simplify their work. Therefore, commands are also a quite convenient, useful tool. They should be used. Again, to create commands, just tell Claude-Code that I want to create such a command. I, for example, have custom commands like `push` to push code to GitHub. And for this command to appear after changing the config, you need to close Claude-Code, open it again, and here you will have the command, for example, `pr` for pull request, right? In this way, you can create various slash commands that will save you time and effort.
Okay, we've covered the seventh technique and slash commands, that Boris uses them. Let's move on to the eighth technique from the creator of Claude-Code - these are, of course, sub-agents, which have a lot of advantages. At least they unload the context and do not overfill the main context, respectively, of the main, so to speak, fork, if one can express it that way, of Claude-Code, right? So, you have some sub-agents, for example, a simple example. You have a sub-agent for writing backend, a sub-agent for writing frontend, a sub-agent that Boris gives as an example, which, by the way, he open-sourced, so to speak, called Code Simplifier, which simply simplifies code. Boris open-sourced their Code Simplifier Agent 16 hours ago, so to speak. He showed their prompt here. I won't dwell too much on this Code Simplifier Agent now. Nevertheless, it is interesting. If anyone wants to examine this Code Simplifier Agent in more detail, see how it works more or less in practice, then in RAN University, I have analyzed and reviewed this Code Simplifier sub-agent in more detail. Honestly, it's quite an interesting agent. In general, for those interested in examining this sub-agent in more detail, go to RAN University. There will be a more detailed review of this sub-agent. And we return. In the eighth technique, Boris, of course, says that sub-agents are used because it is really convenient. They can work independently of each other. So, one, for example, agent does, again, the frontend, the third simplifies code, again, a quite interesting agent, although simple, a sub-agent that Boris open-sourced.
Next, the ninth point, the use of post-tool-use hooks for formatting code that Claude-Code writes. Hooks. What are hooks? In short, these are, so to speak, some commands, some sets, maybe rules, that, for example, here again, Claude-Code made some edits, wrote some code, and a hook is automatically called to check, respectively, code formatting, check indentation, something else. Or you can create your own hook. For example, Claude-Code again modified a file, wrote some code, and you can create a hook again for checking formatting, some indentation, something like that. And you can also create a hook for syntax checking, so that, for example, Claude-Code, respectively, eliminates syntax errors behind it. This is, again, one of the uses of hooks. Hooks have very different usage scenarios, but Boris gives an example that he uses a post-tool-use hook for formatting the code that Claude-Code generates. He says that it usually generates good, well-formatted Claude-Code, but the hook, respectively, handles 10%, for example, of poorly formatted code, right? With a hook, you ensure that everything in the code is adequate. Again, to create a hook, you just type "Create me a post-tool-use hook for formatting code and checking syntax" for example, right? You type this, launch it, Claude thinks, now it will simply create a hook in your local config, a hook that checks code formatting. Okay. Here. Again, you can use the hook that, respectively, Boris posted. I am simply showing you an example of how to create these hooks, right? So, you just type in the chat, and Claude-Code will create a hook for you. Now it will be launched after, for example, Claude-Code has edited some file, added some code, and a hook is launched, and the formatting is checked, for example, and the syntax of the code that Claude-Code wrote. This is quite a useful thing. I didn't use this before and, respectively, learned about it from Boris.
In the tenth technique, Boris says: "Do not use dangerous skip permissions," meaning do not give Claude-Code complete, so to speak, freedom, with which I completely agree. Instead, he does pre-allow commands, respectively, that Claude-Code can run, right? So, he has, respectively, 21 bash commands that Claude-Code can run. In principle, there is nothing particularly interesting here. Simply, instead of giving Claude-Code complete freedom, he gives it, respectively, pre-approved commands that Claude-Code can use without asking for permission to use them.
In the eleventh technique, he says that many Claude-Code tools run and use it. So, here he gives the example of MCP Slack, meaning Claude-Code itself manages Slack for him. So, again, here is MCP, Slack, MCP. Again, nothing particularly interesting. He just says that he uses Slack's MCP. Again, nothing supernatural. I think that MCP is now used by a large number of people. Again, Boris just says that he uses Slack MCP.
In the twelfth technique from Boris, as you can see, he runs Claude-Code for 2 hours and 47 minutes a day. At the same time, he uses 2.4 million tokens in thinking mode. This is actually very, very serious. And in such cases, Claude-Code or the prompt itself is used by Claude to verify that the task is progressing, right? So, you can, for example, in a recent Claude-Code update, create some task, "Create me, I don't know, a Twitter agent, right?" So, you write this and then say, "Create me a Telegram agent," figuratively speaking, you do it like this, and Claude-Code, and, respectively, Boris can do this, and, respectively, Boris checks with this method whether the background task he assigned to Claude is working. Or he has a special stop hook for this. In general, if you, respectively, load Claude-Code so that it works around the clock, you can prompt it as I showed you now, or use a special stop hook that does this deterministically.
In the thirteenth technique, Boris talks about, probably, the most important technique, the most important thing that he considers. This is to give Claude-Code, respectively, a path for verification of what it has done. So, figuratively speaking, it corrected something for you on the frontend. You need to, respectively, give it the opportunity to check what it has done. Or, um, I don't know, Claude-Code made you a new handler for a Telegram bot. You need to give the opportunity for, respectively, Claude-Code to verify what it has done. And this, as he writes, increases the quality of the final result by two to three times. I completely agree with him here. In any case, when you write any code with Claude-Code, you need to, respectively, somehow verify it later and with that, so to speak, create a feedback loop, something like that. And, consequently, the quality of your code will increase.
Here, for example, he uses Chrome. How does it roughly look? Look. First of all, you need to download the Claude Chrome extension. How is this done? You go to Manage Extensions, go to Chrome, type Claude here, download the strictly official extension. Here's how it looks. 2.9 ratings, because, honestly, this is all in beta. I will also leave the link in my Telegram channel on how to set all this up. So, figuratively speaking, you download the extension from Chrome, then you go to your Claude-Code, type `Claude update`, type `Claude Chrome`, set the flag, and then you can, respectively, run Chrome. Let me show you an example. I hope it works now, though I'm not sure, because it's still in beta and doesn't work very well. I only managed to do it once before the video, and honestly, I didn't particularly like this mode. So far, it hasn't proven itself to me, but in general, you understand the essence, that after each change, you need to give Claude-Code the opportunity to verify what it has done. So, it can correct itself. And this is super correct logic. I completely agree with Boris here.
So, as you can see, this command is not working for me at the moment. Okay, let me try to launch the flag again. As you can see, it works intermittently. Right, I couldn't show it to you before the video. I managed to launch this Chrome. It always, honestly, disconnects, respectively, this connection. So, you have this Claude-Code connected to the Chrome extension in Chrome, respectively. And when you type something here, for example, Google Chrome turns on and shows that Claude-Code is doing something there. In principle, it's a convenient thing, theoretically, but now it's all still in beta, in such a wild beta version. Right, as you can see, beta. And, honestly, it doesn't work very well here. Everything works intermittently. I won't even set it all up again, honestly. If anyone is interested, you can set it up. I will leave the link in the Telegram channel. It works for me intermittently. And, honestly, in practice, I don't use it yet. And, in principle, if, for example, you are creating some Telegram bots or something else, this verification path is not particularly suitable for you. Therefore, in the thirteenth technique, the main essence is simply to give Claude-Code the opportunity to verify what it is doing at all.
So, guys, we have broken down 13 techniques for using Claude-Code from the creator of Claude-Code himself, Boris. Thank you very much to him for sharing his tips and techniques on how he works in Claude-Code. Honestly, in my opinion, I like everything here. I used many of these techniques myself, respectively, before Boris's tweet. But what caught my eye, what I don't particularly use. First, slash commands, honestly, I didn't use them much. Now I've started using them much more often. Next, I tested the Code Simplifier sub-agent. Again, for those interested in getting acquainted with this sub-agent more closely, you can do so in Ry University. There, I get acquainted with this sub-agent more closely. Also interesting for me was the hook for code formatting, but I expanded it a bit, and I have, respectively, code formatting for spaces, for indentation, and so on. I also check the syntax. This hook works coolly. I honestly don't use this tenth technique. So, for me, so to speak, an opening, one might say, was, first of all, the thirteenth technique. I didn't think about it much. So, now I've approached it more thoroughly and I really try to ensure that Claude-Code can verify what it has done, right? Also, again, hooks and sub-agents and slash commands.
In general, again, guys, I hope you found it interesting to watch the 13 techniques from the creator of Claude-Code, how he uses Claude-Code, how he recommends using it, or rather, not recommends, but shares his, so to speak, workflow. Honestly, for me, it was very interesting. I hope it was for you too. Be sure to write in the comments which techniques you use when working with Claude-Code. Because, again, there is no ideal workflow. Everyone has their own workflow, with its pros and cons. So, share your workflow in the comments under this video. Be sure to give it a like if you enjoyed this video. I thank you for watching. See you soon.