Claude Code Zero to Hero in 60 Minutes

0:01 Hello everyone, my name is Abhishek and welcome back to my channel. In today's video, let's understand clot code from zero to her. Basically, I'll cover right from installing clot code CLA on your terminal to understanding different features of clot code to using concepts like MCP skills with cloud code,

0:22 building enterprise applications with clot code, and even how to use open models with clot code. I'll cover all of that in today's video. And just like any other zero to her tutorial on this channel, even for this I have created a GitHub repository. So after watching this video, you can follow the notes

0:41 using the GitHub repository. Let me quickly go through the repo because it has a very useful information. So this is a repository. It's called Cloud Code Crash Course. I'll share the link in the description and also in the pin comment. Now if you go through this repository you can find different MD files. The

1:02 video also covers the concepts in the same order starting with what is clot code then installation and configuration of clot code C cla how to use your first session how to use the built-in /ash commands within clot code. How to switch between the permission modes and settings and then how does clot code

1:24 handle the agent memory. Then we will understand how to manage context efficiently using clot code which is a very very important concept. This will also help you understand how to save money while using coding agents like clot code. Then we'll understand how to create a custom slash command. Then we

1:44 will understand how to use hooks. What exactly are hooks? How can they help you secure your entire environment? Then we will cover MCP servers, sub aents within cloud code, skills and plugins, headless mode. You don't have to get into the cloud code. How can you use one time command or one time execution using clot

2:06 code and finally some best practices exactly what we are going to cover at any point of the video. Let's say you could not follow the command that I'm using or maybe after watching the video you want to revise a particular concept. You can just head to this GitHub repository. You can go to the relevant

2:24 topic. Let's say you want to revise the memory management or agent memory. You can just go to this file. You will find everything that you need. Perfect. So this is how you use the GitHub repository. Now that you understood what we are going to cover in today's video, let's get started. What exactly is clot

2:44 code? Now in very simple terms, clot code is an AI agent that is developed by the company Anthropic. Now if you are a beginner, you will have this question. What is the difference between an AI agent and AI assistant? Very simple and this is a good question. On day-to-day basis, you use tools like Chad GPT. Chat

3:09 GPT is an AI assistant. Basically you can go to chat jpity you can ask it a question and it will answer the question for you. Just imagine within your organization you are given with a task to create a python application and deploy the python application maybe a simple todo python application. You can

3:32 go to charg you can provide it with the same question. What it does it will give you a list of steps. It will tell you this is how you create a Python application. This is how you package the Python application and this is how you deploy the Python application. Now it is your responsibility to take these steps

3:56 and execute these steps on a machine. So with charge GPT you can achieve your task but there is a human intervention. That is how AI assistants work. Whereas when you look at an AI agent predominantly how it differs an agent can execute the end to end action on your behalf. So you can go to the agent

4:20 you can explain the same thing I want to create and deploy an application. It will create the application. It will deploy the application and it will get back to you saying this is the URL for the application or this is the application that is deployed for you. That is the difference between an AI

4:40 assistant and AI agent. In simple words, assistant can answer your questions. It can provide you with the list of steps for the action to be completed. Whereas an agent can execute the end to end action on your behalf. Now clot code is an agent. It is not an assistant. Cool. Now are there multiple agents?

5:05 Definitely yes. Along with clot code there are agents like codeex, there is copilot from GitHub and there are any number of agents. Then Abishek why did claude code become very popular? There are many reasons we are going to cover in today's video when we'll cover about the different features of cloud code you

5:27 will understand why this became very popular but one of the main reasons when cloud code was initially launched it was launched as a CLI tool for example let's say I'm working on a code base uh let's say I'm working on this repository there are different files I want to add a new

5:48 feature once I install Cloudwood. I can just head to the terminal and if I type cloud, you can see an AI agent is running here. Now I can just go to this agent and provide it with the steps or provide it with the task. It will complete the task and once it is done I can see the executed task here. probably

6:11 I can ask it to add one more empty MD file and it will do it for me or I can say deploy this application and it will do it for me. Perfect. So it started as a CLA tool and developers and DevOps really like this experience. Instead of having the agent running somewhere or instead of having agent on their

6:34 browser, they have the agent running within the terminal. But then Anthropic decided to make the adoption more viral. They want more people to use it, not just developers and devops engineers. So eventually they also added a VS code extension of cloud code. Today you can go to Visual Studio Code and within the

6:58 extensions if you search for Cloud, you will also find a Visual Studio Code extension. This is the one. And there is also a desktop application for cloud code. There is also a mobile application for clot code. So you can use in any of these ways. However, I would recommend to use the CLI option of cloud code

7:22 because within your organization this is how you would be using most of the times because of the well integrated experience. Perfect. So summary what is clot code? Claude code is an AI agent that is developed by the parent company Anthropic. Predominantly it is used as a CLA tool but it also supports VS code

7:46 extension, desktop application and mobile application. What are some of the features of cloud code before we start using it or before we start learning the different concepts of cloud code? one it is a terminal native one as I explained then you can use cloud code for creating applications

8:09 for writing test cases and any of your day-to-day activities. The good thing about cloud code it supports a very good context smart context that is you don't have to explain cloud code a lot of things once you start claude here it understands your application it understands the different files in it you can just start

8:33 working on it then it also can be integrated with git let's say you want to push your changes to git it supports it and the most important thing the pure speed of cloud code is what makes clot code or the adoption of cloud code very very huge. Perfect. Now that you

8:58 understood what is clot code, let's move to the next topic. How to install clot code. Now for this we will head to the official documentation. You can just search for clot code docs or you can find that link in the description as well. Once you are here, head to the quick start section. There are two

9:19 prerequisites. One, a very basic prerequisite. You need a terminal. Most of the laptops today have a terminal by default. Or if you have a specific terminal requirement, you can also go for that. For example, in my case, I'm using ghosty as the terminal. Once you have the terminal, the other

9:39 prerequisite, you need a active clot subscription. But Abhishek if I go to the pricing page I can see try claude for free. Now you need to understand this is only for the claude assistant. This is not for the claude agent. Just like how you can use chat GPT for free. You can also use the

10:02 claude assistant for free. You can create an account. Once you have the account, you can use the claude assistant to ask the basic questions just like how you use chatb on your day-to-day basis. But if you want to use claude as an agent, maybe through your CLI or through the desktop application, then you either

10:23 need a pro subscription or the max subscription. The primary difference between pro and max subscription is the number of tokens that you can get. But Abishek, what are tokens? Don't worry as we move to the next topics in the video you will understand tokens in detail but at this moment just understand the usage

10:46 of clot code agent you can get more with the max subscription you can create more number of applications you can write more number of test cases or you can build more projects if you go for the max subscription and max subscription can go up to $200 as You can see here this is $100 but if you want 20x usage

11:09 then you have to pay $200 and additional taxes as well. Perfect. Now you might have a question Abhishek is there any way I can get clot for free? Yes, there are some options. You can use clot code with light LLM or you can use clot code with open router and you can possibly use their free models but you will not

11:33 get the same coding experience or you will not get the same agentic experience while using claw code with anthropic proprietary models. The models like fable the models like uh opus sonut the free models will not give you the same experience. Perfect. Towards the end of this video, I'll also show you how to

11:56 use clot code along with the free models. Now going back to the quick start, assuming you have the both the prerequisites. I have the terminal, I have the cloud code subscription as well, then you have to just execute this command. The same command works on Mac OS, Linux as well as WSL. If you're

12:16 using Windows, then better go for WSL. You can just install using a single command Windows subsystem for Linux. Copy this command. Execute this command on the terminal. This single curl command will download claude and it will install claude for you. I already have this so I'm not executing it. Perfect.

12:37 Once you are done with this, just type cloud and you should see something like this. If you don't see something like this and if you run into any error, there are some troubleshooting steps here. Just follow the troubleshooting installation. Now that this is clear, let's move to the next step. How to use clot code to

12:59 interact with your project. Now, this is actually very simple. In fact, I explained this at the beginning of this video as well. Anyways, now let's deep dive. The first thing that you do when you join a project, you download the source code of the project or maybe you clone the source code of the project.

13:18 Once you do that, you switch to the directory. Now to interact with this project using claude, all that you have to do in the directory just type claude. In my case, I'm using a repository called Argo CD because it has so many source code files. Now I started claude. I can just ask

13:39 claude explain the goal of this project. If you observe carefully I haven't provided Claude with any context because this is a simple activity because this is a simple task. Claude can go through the source code files. It can perform the analysis and it can explain you what is this project. In no time it explained

14:03 Argo CD declarative GitHub CD tool for Kubernetes core goal keep the live state cluster matched to the desired state in the G repo. Further it explained few more things as well which is good. Now not only basic explanation of the project you can also use clot code to make some changes. For example, I will

14:26 say in the readme file add this is the best project. Okay, this is a vague thing I know but just to show you using cloud code you can instantly make changes to your source code files as well. If you are working on a feature just provide some information of the feature. It will go through the files.

14:49 It will understand where to make the changes and this way it will get back to you saying these are the lines that I want to change. So in this case it says I'm adding this new line. This is the best project and it is asking me for the approval. Do you want to go ahead like do you want

15:09 me to go ahead and make the changes? I'll just say yes. Now you might have a question. Abhishek, at the beginning of this video, you told clot code is an agent and it can perform actions without your approval. Then why is it asking for approval? See, basically coding agents have advanced a lot. Now they can perform

15:32 dangerous actions as well. For example, if you provide a wrong command or if you provide a wrong instruction, it can delete all the files from your laptop. To avoid that, what anthropic did in the default mode like if you're using claude, you can see here this is called manual mode. This is the default mode.

15:53 In this mode, it will ask you for the approval. Abishek, can I change the mode? Can I move to something called auto mode? Yes, if you just press shift tab, you can see the modes are changing from manual mode. Now we are in the plan mode and now we are in the auto mode. If you perform any activity in the auto

16:15 mode, it will not ask you for the approval. It might give you some suggestions but it will not ask you for the approval. So this is how you interact with your project using clot code. make some changes, interact basic questions or even add some test cases. Now before we move to the next step, I

16:35 want you to understand some basic shortcuts because this is predominantly CLA experience. We will be using the CLI. So there are two to three shortcuts you should understand before going to the next step. The first thing let's say you are asking claude code show me the test case coverage. Okay. Now claude

16:59 will start performing the action but you want it to stop. Maybe you provided a wrong instruction or maybe you want don't want it to do it. Just press escape. Right? Lot of times you want to stop cla, you want to interrupt claude, you just have to press escape. You can see interrupted

17:20 or maybe you want to go to the previous instruction. Okay, now you interrupted it. Now you want to claude to go back to the previous action it performed probably adding this is the best project or even before that. How would you do that? Just press escape twice. Right? If I press escape twice,

17:42 it will show me the list of things I asked Claude to do in this session. Just like the history command, if you're using Linux, you know the history command. Almost the same thing. But the thing here is that here using the up arrow, I can move the state to a particular thing. Right now, this is

18:02 where I am. If I want, I can go to this instruction. Show me the test case coverage. Or if I want, I can go to the previous one as well. For example, I'll go with explain the goal of this project. And you can see when I press on enter restore code and conversation. So, it is taking the complete thing to that

18:24 particular state. I'll just say yes, restore code and conversation. And there you go. Now it went back to the state explain the goal of this project. So it is similar to the history command but using the up and down arrow you can take your clock code to that particular context. This is again very useful

18:48 especially when you're working on a huge activity. Okay. Now how do you exit from claude? Just use C and press Ctrl C twice. Right. If I press Ctrl C twice, I exited out of the cloud code experience. Or even I can just say /exit as well. Abishek I exited clot code. Now I want

19:16 to go back to the pre same session of clot code like I want to go back to the previous session of clot code. How can I do it? There is an option. You can just say claude - resume and it will show all the list of the sessions. So these are the previous sessions that I worked with claude. This is a previous session

19:38 explain the project goal. I'll just click on it and it will start from the same thing. So these are some basic day-to-day commands that you would use. Press escape once if you want clot to interrupt. Press escape twice if you want to look at the history. and go back to a particular point in the history.

19:57 Press Ctrl C twice if you want to exit cloud code. Use cloud hyphen resume if you want to go to a particular session that you already used. Now that this is clear, let's move to the next step. Built-in /comands. We just looked at shortcuts. Shortcuts are definitely useful, but they can only help you with simple

20:19 things like switching between the modes or exiting clot code. What if you want to do more with clot code? Let's say you want to see the model that you are currently using or maybe you want to switch to a different model. How do you do that? This is exactly where built-in slash commands can help you. They are

20:40 called slash commands because you have to start your message with a slash. When you start your message with a slash, cloud code lists out the built-in slash commands that you can use. There are close to 30 to 40 built-in slash commands. You can go through each one of them or you can go through our MD file

20:59 as well. Anyways, I'll cover the most important ones. As I told you, you want to look at the model that you are currently using. Just type /model press enter you will see the model that you are currently using. I'm using sonnet because I'm working on a simple routine task like whatever I'm

21:20 explaining it's routine task. So I don't want to have to go with fable or opus that's why I'm using sonnet. Let's say I want to switch to haiku or I want to switch to fable file. I can just use the up arrow or down arrow and press enter. If I use the down arrow, I'll go to haiku model. I'll click on enter. You

21:40 can see model is set to haiku 4.5. Now the next instructions will start using hiko 4.5 until I switch to a different model. Similarly, you want to track the usage. Again, very useful because especially when you're using Claude Pro Access, that is the $20 thing, you want to track your daily, weekly, and monthly usage

22:06 because 2 days down the line, you might work on an important activity. If you consume all your capacity currently like today, you consume all your tokens, then your weekly limit is exhausted. 2 days after when you are working on a critical activity you will not be left with any usage you know cloud code tells you you

22:28 have hit your weekly usage you definitely don't want that to happen so always run the / usage command and once you look at it just press escape you will go back to the terminal another important slash command in fact very very important slash init remember this command because what it does

22:52 imagine you're working on a new repository. Okay. So you joined an organization you are given with a repository and this repository nobody has worked it with it using clot. You are the first one. So you will just to go to claude and you will just type /init. This will prepare a file called claw.md

23:16 everything in capital. This cloud.mdi provides the guidance to the cloud agent. Okay, this provides the guidance to the cloud agent about the repository. For example, there are some important things in the repository that claude has to remember. Maybe how to build the package, how to deploy the application,

23:38 what is the folder structure, what are some important limitations of the project. The claw.md file will have all of this. Whenever cla performs an activity if required it will look at the cloud.md file. This way it can reduce the token utilization for you and this way it can remember more information

23:59 about your project. Abishek what if the claw.md file did not capture something important. No problem. You can go to this file and you can edit the file. Claude agent looks at this file and it understands what you are trying to explain. That's why claw.md file provides the guidance to the project.

24:21 Now this has become very very popular today. If you look at even the popular repositories for example if you go to argo CD right this is an open-source project. Even within this project you can find the claw.md file. So you can understand the utilization of this. I'll just type claw.md and you can find the

24:43 file. Right? So how do you create this claw.md file? Just go back to claude and type /ash in it. There is already claw.md file but this will create a new one just to show you and I'll quickly show you what is the context in this file like what is this file updated with. I'll go

25:08 to the new tab because Argo CD already has the claw.md file. So it will say the file already exists. I'll take a new tab cat.md and quickly show you. So this is the content of the file. This file provides guidance to clot code when working with code in the repository. What does it remember? Common commands.

25:35 how to lend the project like how to verify the lending of the project, how to build the project locally, how to build the CLI, how to lend the user interface, right? What is the folder structure of the project? So let's say you're asking clot to create a feature. It will immediately look at the cloud.

25:53 MD, it will understand everything about the project. whatever it could not understand directly. Maybe you have passed some instructions, it will look at cloud.md and it will start working on the project. Even if you don't provide cloud.md, it will still work. But when you provide cloud md, it can work more

26:13 efficiently. Another slash command. I'll just escape. Another slash command slash login. Maybe you have two cloud accounts, right? One account you are using along with open router, one account you are using with the clawed frontier models. You can use the slash login and you can switch between the anthropic accounts.

26:37 Right? So press / login. It will open or it will ask you for the token. It will go for your browser authentication and it will login with the relevant user. Along with these things like these are four important commands what I just covered /init /mod/ usage / login along with that there are some basic things

27:01 for example you want to see all the built-in slashcomands and you want to understand what exactly each command does just press /help. Now, this will show you the built-in commands and their use cases. Or maybe you want to clear all this context. Okay? Like just like when you're working

27:24 on Linux machine, you want to clear the complete text on your screen. What you usually do? You press clear. Now here you will just type /cle and this will erase all the context within this session. See, I'll just do /cle. Everything is gone. But you have to be careful with it. Now

27:46 you're not only clearing the text on your screen but you're also clearing the context that agent has about whatever you are performing. Maybe you are working on a task then you have to be very careful about slash clear. I would recommend this thing whenever you are switching between the tasks. You are

28:04 working on three task. First task is complete then do/clear. Do not frequently do SL clear. You are completed with task two, then again do SL clear, then again do SL clear once your third task is done. You might say, but Abishek, what is the point of doing SL clear? I'm fine. I'll

28:27 just keep using the same cloud session for multiple tasks that I'm working on. There is a problem. You know, every model has something called context window. For example, if you are going with haiku model, if I am correct, it has a context window of 256k tokens. After that, it will start hallucinating

28:50 or the performance of it will go down. If you're going with opus or fable, then yeah, the context window is 1 million tokens. So, it can remember a lot of context. what I'm trying to say so everything that you see on the terminal within the session you see huge amount of text that is basically context

29:15 whatever you have in this session agent will try to remember so if you keep using the same session there will be lot of text on the screen this is what agent remembers that is called context if it exceeds 256 tokens then the performance of the haiku or the sonet model will come down if it exceeds

29:36 1 million which is actually huge but still if it exceeds the performance of opus or fable can also come down but Abishek is there anything that I can do like I don't want to clear everything you can also do something called / compact what / compact does it will not clear all the context but it

30:03 reduce the context. Compact you can understand, right? So it will just compact the context. Instead of having like if you have uh 250 tokens of context on your session, then this might come down to 125 token context, right? So keep using /clear and / compact wherever it is useful. You can look at

30:30 other slash commands on the GitHub repository or you can just use /help and look at all the slash commands. Just spend some 30 minutes to 1 hour. They are definitely useful. Now that this is clear, let's move to the next topic permission modes and how to configure custom permissions. By default, there

30:51 are four modes in cloud. The default mode, manual mode. If I press here and let's say I use shift tab, you will notice this will change from manual mode to edits mode. And when I press shift tab again, this will change to plan mode and finally auto mode. Let's understand what each of these mode means in detail.

31:16 The first one, the manual mode that itself explains. Claude or the agent does not take any action without your permission like any write action by default. It can only read the files. So it will read the files. Let's say you ask it to create a pull request or maybe you ask it to update a particular

31:39 module. It will not do it without asking your permissions. It will list out the changes and it will then ask you will you approve these changes. only then it will move to committing those changes exactly what we have seen for example we are in manual mode now I'll just say update readme file with a new line in

32:03 the beginning saying what is this project I just said add a new line to the readme file you will notice it read the file it did not ask for a permission That means read is granted. Now it is trying to go through the file. It will try to add the change. If you observe for read it did not ask our

32:30 permission. But when it is writing now it says do you want to make this edit to the readme.md file. That means only if I press yes this will go on. This is the manual mode. The most safest mode. You don't have the risk of deleting any files like your agent cannot accidentally delete your files or agent

32:51 cannot accidentally do anything. Then there is this mode. If you switch to accept edits only. Now what does this do? Files edits are approved automatically. Command still ask for your permission. What does this mean? It can make the change like the change that you have seen here. this change is made. But in

33:16 edit mode, let's say you ask it to commit this change. For committing the change, it has to run the g commands. Let's see what happens. This should ideally ask for a permission. So it is trying to run get add readme.md get commit readme.md. You can see do you want to proceed? That

33:42 means in edit mode it can update the files. It can read, it can edit. But if it has to execute any permission, it will take your consent. So even this mode is safe for you. You don't have the risk of accidentally deleting the files on your machine. Perfect. I'll just say no for now. And I will change this to

34:06 auto mode. Let's see what happens. Now I'll just say commit this. You will notice in auto mode because agent is in auto. It will not ask for your permissions. This is little risky and this is little dangerous. You have to be careful with it. Commit it. Especially if you're just starting with

34:27 cloud code. I would highly recommend you to use manual mode and then go to accept edits mode only once you are comfortable. You are good at you became good at prompt engineering. You're good at using clot code then go to the auto mode. But Abishek then what is this plan mode? Plan mode is basically when you

34:48 want to perform any thinking activity. Let's say you are designing a terapform architecture or maybe you are preparing AWS VPC architecture. Before writing the terraform code, go to the plan mode and start asking agent to prepare the architecture, provide you the detailed architecture. Once your plan is done,

35:10 you can go to auto mode or edit mode according to your convenience. So claude can only read and analyze in the plan mode. It produces the plan and it is up to you if you provide the approval then it will ask you should I change to the auto mode and proceed with the execution of the plan just like terapform plan. If

35:30 you are a devops engineer you already know what is teraphform plan almost the same thing. Perfect. So these are about the default permission modes manual mode auto mode plan mode and accept only mode. Now you might have a question Abhishek what if I want to customize. Okay these are the default modes. What

35:52 if I want to allow certain things and I want to deny certain things. I want to use the auto mode. So this mode here I want to use the auto mode but I want to make sure RM minus RF is disabled. that is I don't want claude to delete the complete file system on my local machine even

36:15 when I'm in the auto mode accidentally even if I ask cloud to delete it should not delete it or while performing any activity it should not delete files on my file system how can I do it yes you can customize the permissions just use the slash command one more slash command and again this is also a built-in slash

36:36 command just do slash permissions. Press enter. My bad. I should have done / permissions. / slash permissions. Press enter. You will see a tab here. Right. Permissions recently denied. Allow. Ask. Deny workspace. Go to deny. Click on add

37:02 a new rule. And here you can paste the permission rule. For example, I'll paste this JSON. My bad. I have the JSON within the GitHub file. I'll just paste it here. Which GitHub file I'm talking about? Don't worry. Go to our GitHub repository. I'm currently in permission modes. Go to this one and you will see

37:23 the JSON that I'm copying. As I told you at any point of this tutorial, this GitHub repository can help you. If you observe what is the deny rule? Deny rm - rf. Deny reading the environment secrets. Deny the web fetch. That is do not fetch anything from the internet. Think from the source code. Read the

37:47 source code. Analyze the source code. This way you can deny set of actions and you can also allow set of actions. Perfect. So this is how you define custom permissions. Now just press enter. It will ask do you want to configure this at your project settings level? Do you want to configure this at

38:08 your personal project level or do you want to do it at the user settings level? Now I have to explain you what is the difference between these three things. In fact that is asked in the cloud code certification as well. I have recently given cloud code certification like even if you're giving this topic

38:27 will definitely help you. So let me explain in detail. So when you are going with custom permissions, there are three levels at which you can define the custom permissions. One, local level. Now what does local level mean? Saved in docloud settings.local.json. So anything that you configure at the

38:50 local level, this is your personal setting. Then you have an option where you can configure at the project settings level. What does this mean? Let's say you're working in a team and you want to share some clot port permission settings within your team. You want every developer within your team to use the

39:12 permissions. So you can select this option the project settings. It is created include settings.json and you can share it with your team as well. Then you have the user level settings. What does this mean? User level settings is applicable for every cloud project that you are doing. For example, right

39:36 now I'm working on Argo CD project. Then maybe I'll switch to Argo workflows. Then maybe I'll switch to one of my personal projects. If I configure this at the user level, all the projects that I'm using with clot will get this permissions. The same thing is explained here. Project settings saved in doclaude

39:57 settings.local.json. This will be your personal thing. You can commit it to git. You cannot commit it to git. It's up to you. Project settings dotcloud settings do.json. You can share it with your team. You can share it within your enterprise. User setting. This is applicable for your

40:17 entire projects that you work with clot. Perfect. I've explained the same thing in the GitHub repository as well, right? You can find the exact same thing. Your personal defaults for all the projects. This is cloud settings.json user level dotcloud settings.json you can commit it

40:38 to git share with your team. Settings.local.json. This is just for you ignored by git. I know I'm repeating this but there is a reason why I'm repeating this. This is very very useful especially when you're working in the organization. The default four modes are not good enough. You want

40:58 to customize the permissions. Imagine you are working within your organization. You use the auto mode that is you are telling agent to go ahead. I will not you know you don't have to ask me whether to approve this command, deny this command. It is dangerous because at some point agent might read some

41:17 document, agent might think on its own and it might delete important files from your file system. On the other way, you cannot keep using the manual mode also because every time cla will ask you to approve the changes. Imagine you are working on a different task, the claude agent terminal will stop for your

41:40 approval. So you cannot even use that like you cannot do multitasking if you go with manual mode. That's why the permission/comand is very very useful. In fact you can pause the video here go through the document until and unless you are clear about the permission modes and the settings. Now that this is clear let's

42:01 move to the next topic. Let's look at two topics together. Cloud MD and memory and managing context. The reason is simple. We already looked at most of it during the slash commands. I mean the /built-in commands. Let me quickly walk through them. The first thing the claw.md file as I told you what exactly

42:22 does this do? When you type /init claude agent creates a file called claw.md. What is the responsibility of this file? As we looked in the argo cd project. This file provides guidance to clot code when working with the code within this repository. What kind of guidance? What the project is and how is the project

42:47 structured? You know, I'm just telling you the common things that are in the claw.md file. Let me write it for you. Most of the times this is what you will find in the cloud. MD file. The first thing the project structure. what is this project and what is the structure of the project so that claw

43:07 does not have to go through every time then what are the commands to build the project test the project and any inline commands the project supports so that if you're working on a feature claude can automatically build the project test the project and probably share you a local URL

43:27 what exactly is the coding style of the project this is the most important ing because every organization has a different coding style. By default, claude can generate any coding style for you. If you make sure the claw.md file is updated with your coding style, claude makes sure every time it

43:48 generates the code for your organization, it keeps the same style. But Abhishek, what if it is not present in the cloud file? As I told you, if slash in it, I mean when it creates a cloud. MD file, if that MD file does not have a coding style, don't worry. Go to the cloud.MD and just update it. And then anything

44:12 else that Claude keeps getting wrong. For example, in the cloudMD file, you have worked with claude maybe for three different task and every time it is doing something wrong. You can go to claude.mmd and you can tell claude hey you are doing this task wrong every time make sure next time when you work on a

44:35 similar task be careful with it even those kind of things you can provide in the claw.md file the one that is generated by /init command whenever you work on a new project go ahead and type /init you will see an md file like this this is about memory now let's learn about managing the context. Okay, I've

44:58 been using this term called context for a while. I have been using token for a while. What exactly is context? What exactly is token? Simple context is nothing but everything that Claude is working on currently. Everything that it can see including the messages including the files that it is reading commands

45:20 that it has access to the content that is within your repository becomes the context. If you're not providing any external context this is the default context that it can get. If it is working on a GitHub repo everything that is within the GitHub repository and every message that you typed becomes the

45:40 context for the cloud agent. Now what are tokens? Tokens are nothing but let's say you provided claude with a prompt and it replied maybe it provided a detailed answer. Everything model calculates in terms of tokens. You can consider 3/4 of a word as a token. Now the costing of the large language model

46:06 depends on the tokens. So you have to be careful how much output is your model providing. If you're just using your model for very trivial activities for example you have fable model and you use fable model for trivial things like you are asking fable model to explain the project you don't have to do it because

46:29 fable 1 million context or 1 million tokens the output tokens is basically $50 US. So using fable if you generate output of 1 million tokens you can calculate roughly one token is 3/4 word. So if fable provides an output of 1 million tokens you will be charged with $50.

46:55 On the contrary if you are going with opus 4.8 which is also a powerful model but this will only cost you $15. Probably if you go with sonet it might only cost you $5. So every time you're working on a task, you have to choose a model. If you're working on writing documentation, just

47:16 go for sonet model. If you're working on a security task, yeah, you might have to go with the fable model. If you're going with an enhancement, you can go with opus 4.8. That is why I keep saying tokens are very important. I hope this is clear now. What is the difference between context and token? Once again

47:36 context is whatever your claude agent is seeing and whatever you have provided the information to your cloud agent. It can be through files, it can be through text, it can be through architecture diagram as well. Every information that you provide before it works on a task is basically the context or even

47:57 information that it gets while working on a task is basically the context that it has. Fine. Now let's understand the concept of hooks because cloud is very powerful. If you decide to use it within your organization, you should definitely use the hooks. I'll give you a very simple example. Imagine you are working

48:17 on a task with Claude and accidentally one of the users or even it can be you. You write a prompt and within the prompt you commit the API token or accidentally you share your API token with claude. It can be directly within a prompt or maybe one of the files within your repository

48:40 has an API token and there is a chance can read it. Of course, you don't want this to happen. How can you prevent this? You can prevent such thing using hook. Not only this thing, any kind of sensitive thing, you can prevent using the hook. Now, you can define hook at various levels. For example, claude

49:01 supports six to seven very popular hooks. The first one, pre-tool hook. Now, this is a type of hook. What you deal with is the pre-tool use. This executes before a tool runs or before clot tries to run a command and if required the pre-tool use will block the execution of clot or will block the

49:26 action of the clot. Similarly, you can run something called post tool use. As the name suggests after a tool succeeds let's say claude is running an action and the action is successful then basically you can use this post tool hook this is good for formatterers or this is good for llinters

49:49 maybe using claude you are writing an application application is written by claude then you can tell claude to run this post tool use if you configure it claude will definitely run it and it will verify the linting according to your organizational standard. User prompt submit. This is another kind

50:10 of hook which I'm talking about. This will verify when user sends a message before the model sees it. It verifies if there is any kind of sensitive information that is being submitted. Very very useful one. Right? If if you're working in an organization, you should use this user prompt submit type

50:32 hook so that if user writes a prompt before the prompt is sent to the model, the user prompt submit hook will verify if there is any sensitive information. Similarly, there are other hooks like stop. This hook will execute when claude finishes responding. It can be successful, it cannot be successful.

50:56 When you define a hook of type stop, it will execute. Similarly, there are session start notification where you can ask claude to send an email notification or a slack notification when a particular task is done. These are the different type of hooks that you can configure. Most important ones that you might use

51:16 this is the one user prompt submit hook, notification hook where let's say you are heading to lunch, right? and you're asking claw to work on an action, you can configure the notification hook. It will send you a message on Slack that the particular task is done. Okay, Aish, this is

51:35 looking good. But now, how do I configure a hook? Okay, there are different types of hooks that I can configure. But how do I exactly do it? You can just go to your terminal like go back to the terminal. You can do this at user level, project level, local level. The same thing that I explained. Let's

51:54 say you want to do this at the project level within the Argo CD project and you want to share it within your team. Here create a folder like dotcloud. Create this. Head to do.clude the one that we just created. And create a file called settings.json. I think there is a typo but yeah create

52:16 a file called settings.json. And within this file you can paste any kind of hook that you want to configure. Let's say I'm configuring the post tool use. This is the syntax for it. This is in the GitHub repo as well. Hook of type post tool use. What I'm saying if claude has performed edit or the right action

52:42 using the command execute this particular command. npx prettier write Guess what is this book actually doing? This is a post tool use. What we learned post tool use is executed once claude completes its execution. The same thing once claude completes the write or the edit action just run npx pretty. So

53:09 if Claude has written 10 different files, this hook is saying run the linting and improve the format or the linting of the written files. Similarly, you can say just verify if there is any linting issues within this or maybe you can say see if there are any white spaces or you can say there should not

53:33 be a particular word within the generated thing. Maybe it is writing documentation. You can say if it is a documentation change I don't want claw to write a particular word. If there is a particular word the hook should fail. This is how you can use it. I showed you example of post tool use but similarly

53:52 within the repository you can also find hooks for blocking. You can find hooks for different use cases even notification. For example this is how you configure notification. I'll remove this hook and let me quickly paste it. Hook of type notification hooks. Whenever a command is executed

54:18 and successful, run this script and display the notification within the terminal. I'm saying display this particular notification on the terminal. But I can just change this to send a notification to Slack or send a notification through the email. Perfect. This is how hooks are used. Now let's move to a very

54:43 important topic. MCP like how do we use MCP server within a clot? Lot of you might have this question but Abishek what is MCP? I made a detailed video on MCP but I'll quickly explain it here. model context protocol. MCP stands for it. Basically, large language models do not have context of everything. Large

55:07 language models are trained on certain things. But if you tell a large language model, let's say you are using Fable and if you ask Fable, go to my Jira and fetch the information from Jira or go to my Kubernetes cluster and fetch the information from the Kubernetes cluster. Go to GitHub, fetch the information from

55:30 GitHub. It cannot do it because large language models are trained on certain things. it cannot access your internal tools or maybe external tools. MCP is a way where you allow the large language models to access the internal or the external tools. Now Abhishek who configures these MCP

55:55 servers you don't have to do it the popular tools like Jiraa Kubernetes GitHub they already defined their MCP servers and within cloud you have the MCP client so you just have to install this MCP server within your clot right what are the steps that you have to perform only one step identify the

56:19 required MCP server and install the MCP server but Abishek why MCP why not agents can just use APIs now there is difference between API and MCP there are different tools each tool let's say if it starts implementing its own type of MCP or API my bad then it becomes difficult for the agents it has to

56:44 understand different API types it has to understand different standards that's why what people decided is a standard called model context protocol. So model context protocol each of these tools implements the same standards. So they define their MCP servers. You just have to install those MCP servers and claude

57:06 can start interacting with these tools. What are some popular MCP servers that we use maybe as DevOps engineers? GitHub MCP server, Jira MCP server, Kubernetes MCP server or your database related MCP server so that cloud can fetch information from it and act upon the action. For example, you say perform

57:32 troubleshooting on Kubernetes using MCP server. It will connect to Kubernetes. It will get the required data. it will perform the analysis and then it will give you the response back. This is the reason why a pod in your Kubernetes cluster is failing or if you ask explain about a Jira ticket again it will use

57:52 the Jira MCP fetch the information perform analysis and gives you the information back. Okay, now you might have a question how to configure these MCP servers and how to list the MCP servers that are already configured. Simple. Go back to your terminal. You don't even have to open the cloud port

58:13 CLI. Just type the command claude MCP list. Right now there are no MCP servers. So it would say no MCP servers are configured. To configure an MCP server just use the command claude add MCP. That's it. Now if you provide more information like exactly it says what

58:38 kind of MCP server you know you haven't provided the information I can just say for example I want to configure Postgress MCP server I can just say this claude MCP add- transport the Postgress and I am providing the location of the Postgress server that's it this will install the

59:00 Postgress MCP server Right? As simple as that. One simple command. Now if I do claude MCP list, you will notice checking MCP server health. So it figured out now there is an MCP server and it says Postgress is the MCP server and it is also connected on my local. It used the Postgress and

59:22 it connected to it. This way you can go to GitHub documentation. You can go to other documentation and you can connect the relevant MCP servers. What will you achieve with it? Now if I go to claude and if I ask claude what is the database that I'm using? Okay. Or what is the health of database? What are the tables

59:47 within the database? Write a select query for the database. it will write the select query and it will add the select query on this particular database. That is the advantage of MCP. You are allowing your large language model to connect to an external or an internal tool. Now let's

60:07 take a look at another important concept skills. Skills are introduced after MCP but recently it became very very popular. The concept of skills is very very simple. What does a skill do? Using skill, you can instruct Claude to perform an activity in a certain way. Right? So, what does a skill do? Using a

60:33 skill, you can instruct the Claude coding agent to perform an activity in a certain way. For example, you want to install a particular tool. By default, if you tell claude, it can definitely install the tool for you by going through its documentation. But within the skill, you can say, I want to

60:54 install a tool in a particular way. Best example is Kubernetes. Ask claude to create a Kubernetes cluster. It will definitely do that for you. But there are n ways of creating a Kubernetes cluster. There are different versions of Kubernetes, different architectures of Kubernetes, different

61:14 plugins within your Kubernetes. If you create a Kubernetes skill or maybe you call it Kubernetes installation skill, whenever someone ask Claude create Kubernetes cluster, it will refer the skill and looks at the instructions in the skill and according to the instructions it will set up the

61:36 Kubernetes cluster within your organization. Nice. As simple as that. But Abishek, what is in a skill? Nothing. A skill is basically a MD file. Yes, that's the surprising thing. Whenever you hear skill, you might feel okay this is something huge. Not really. A skill is basically a simple MD file.

62:01 So you can just create something called kubernetes md file and you have to share it with your claude in a particular directory like within doclaude you can create skills folder and within the skill you can put your kubernetes.mmd file call it kubernetes installation md that's it now when you use cloud next

62:26 time and when you access cloud let's say you provide a prompt install Kubernetes cluster. Now claude understands okay user is asking for Kubernetes installation. I have a skill relevant to that. Let me refer to the skill. It refers to the skill according to the instructions. It sets up the Kubernetes

62:47 cluster within your organization. That is the concept of skill. Right? So on your terminal just head to the dotcloud directory. If you are setting it up at the project level go to this create a directory called skills and within this you can set up something called kubernetes

63:12 installation skill or just say kubernetes installation md or maybe you can create another folder. I'm just telling you different ways of doing it. You can even say mkdr kubernetes install. Head to kubernetes install. And here just call skills dot md. And within

63:39 the skills.mmd just say I'll just give one line instruction always install kubernetes 1.36. Then you can provide list of instructions. You can say two worker notes like emphasize on it or enhance on it. Just for the video I'm writing in a simple way two worker nodes and use a

64:06 particular CNI like calico CNI and save the file. Now if you go to cloud within the parent repository if I just type cloud and if I say install kubernetus now what it does it will look at the skill it will identify okay there is a skill so let me go through it it will identify there is a

64:31 relevant skill and according to the skill it will start the instructions because I haven't provided more information you know there is no clarity it is trying to understand okay which Kubernetes cluster local or public Kubernetes cluster but in real time because you will provide a detailed

64:49 prompt and you will write a detailed skills.mmd this will become simple and this has almost become a standard now most of the tools for example if you go to this tool here if I go to insports documentation which is a database you can see in the installation of insch they have provided

65:13 a skills file. Now this is the CLA setup. They have the skills here. If I download this skill I can go to claude say download inforge or setup inforge it will refer to the ins skill. This way some people are moving from MCP to skills as well. I'll make a detailed video comparing MCP and skill. Right now

65:35 you don't have to get confused but just understand skills is a way of instructing your clock code. Now finally there is one concept that you have to understand the concept of plugins. What exactly is a plug-in? We learned about MCP. We learned about a hook. We learned about commands. We learned about

65:57 skills. Let's say you want to install a bunch of them. For example, there is a tool the best example you can take inforge itself. Let's say you want to install the MCP server of ins you want to install agent of ins skill command hook related to ins. What this company can do instead of providing individual

66:21 MCP server, individual commands, individual hooks, they can just create a plug-in and they can ask users this is the plug-in just install this cloud code plug-in and they can provide you a curl command. Once they have provided and once you install within cloud code, if you just type /plugin, you will see the

66:44 plugins. Right? Right now, these are the plugins that I have, but you can install external plugins as well. Plugins simplifies the installation of skills, MCP, hooks, agents. It bundles them all together and you can install all of them at once. That is the concept of plug-in. There are some built-in plugins. Apart

67:08 from that, once you install, you can see the list of them in the installed section. For example, I've installed caveman plugin. I've installed ins plug-in. This way you can do it. Now, apart from all this, within this GitHub repository, I have provided lot of information related to clot code. For

67:28 anyone that is getting started with clot code, they can just follow the MD files in the same order. I've provided more information about MCPS, skills, sub aents, best practices. just go through the documents and at any point of the video if you could not follow a particular command even you can find

67:47 that in the GitHub repository. I'll add one more section to this GitHub repository on how you can use cloud code with open router or maybe I will take that as a different video. If you want to understand that just let me know in the comment section how to use cloud code with open router so that you can

68:07 make use of free models within open router. Cool. I hope you found this video insightful. Let me know in the comment section if you have any questions. Thank you so much for watching today's video. See you all in the next one. Take care.