Claude Code: Get the official certification to get hired!

My goal is to help you succeed. By the end of this video, you'll be able to pass the official Claude Code 101 exam and add this Claude Code skill to your resume to help you find a job. The reason is very simple: I've never had so many companies and entrepreneurs at the start of 2026 tell me, "We're going to close down." Just yesterday, I had a manager in the construction industry tell me, "We missed the AI ​​revolution; we're going to shut down. In six months, we'll be laying everyone off." So, why am I telling you to work on Claude Code today? Because it's the number one tool in companies right now. If there's one skill you absolutely must have in the job market, it's having the Claude Code 101 certification. And that's exactly what we're going to do in this video. And most importantly, at the end, I'll explain the fundamental mistakes you mustn't make, mistakes I myself made, and mistakes you absolutely must avoid with Claude. In a world transformed by artificial intelligence, there's one skill that stands out from all the others today, and that's mastering Claude's AI. I guarantee that if you master it, you 'll be able to take full advantage of it. The time is now. In a world transformed by artificial intelligence, there's one skill that stands out from all the others today, and more specifically, mastering the Claude Code tool. I guarantee that if you master it, you won't just survive the AI ​​revolution; you'll be able to take full advantage of it. And I'm in a good position to tell you this: in recent months, I've seen hundreds of people with absolutely no technical background, in dozens of sectors, successfully transform their daily lives thanks to this tool. And in this comprehensive training, I'm going to teach you everything you need to know to master Claude Code, even if you've never written a single line of code in your life. For those of you who use Claude Code and work in machine learning, data analysis, or are developers, please share your thoughts on this type of prompting in the comments. Do you think these methods are effective? If so, why? If not, why not? Hi Jarvis, as you know, we previously created the website for the platform's future marketplace together. Now, I'd like us to create the MVP of this platform together. I'd like you to be in the planning mode I selected earlier and help me create the application's MVP from start to finish. Just to give you some background, I'd like us to do the design in Claude. So, throughout our interaction, I'd like you to act as both a project manager and a full-stack developer. So you're going to give Claude the instructions and prompts so he can generate the entire interface for us. It's also important that you know we'll probably be using Supabase for our database. Now, help me get everything set up and plan the best features for our MVP with me. Okay, so tomorrow, if you were a boss and you saw an employee using AI this way, would you trust them? Let me know what you think in the comments. I'm curious to hear your opinion and feedback because my goal is to help you succeed. This is an official masterclass based on the Claude Code 101 course. And at the end of this video, you'll be able to take the exam immediately and add this skill to your resume. By the end of this video, you too will be able to say, "I passed the Claude Code 101 exam," and you'll see that you'll be among the 1% who truly master AI thanks to this Claude Code 101 skill. And above all, I'm speaking to those who are engineers, developers, and experienced professionals. Claude's level. So, what exactly is Claude? That's the first point we'll clarify: understanding what an agentic system is compared to a classic chatbot. And we'll understand how to configure Claude's functions, workflows, and how to customize its operation. So, what is Claude? First and foremost, it's a system we'll define as an AI agent. What is an AI agent? It's something designed to make decisions. And to make decisions, it needs to be able to collect information, interrupt the user by asking questions, and take action— in other words, it needs to be able to define scores and results that it must verify according to criteria. It's the combination of all these elements that will allow Claude to implement the agentic loop based on your prompt. In a standard exchange with an AI, what happens is your prompt is sent, and what happens next? Probabilities are established, and you get a response. But this response is nothing more than the probability of the words you specified. It's not real in-depth processing. The difference with the agentic system is that the model, based on your request, will extract certain contextual elements. This could be RAG (Research Anonymous Group), it could be internet searches, it could be files, it could be images. So you need to define what you want to use. Then, it will make decisions: What do I do with this data? How am I going to organize it? What do I need to work with the data? And then, it needs to verify the results. So, to be perfectly honest with you, the fact that it takes actions is, let's say, fairly well developed in models. They know when to connect to the internet or run code, but when it comes to verifying results, it doesn't know what's good or bad. An AI isn't a system that knows in advance whether information is true or false. It only has training data as a point of comparison. And if the training data generally indicates that information points in a certain direction, well, it knows that this type of information points in a certain direction. But if you give it your company's data and ask, "Can my revenue be improved, increased?", the AI ​​can't give you that kind of information because it has no basis for comparison. So, in that case, you're the one who will have to code the actions and calculation verification methods so that the model can perform the task. So, the design of an agentic system is about being able to build each of the blocks to perform the work. And it's not simply: "Hi, my name is Fabrice, I'm an AI engineer, now I make YouTube videos. Give me the magic prompt for the YouTube video." The problem is that this way of interacting is absurd, which means that, now you understand, an agentic loop is only functional if a model has acquired this skill and this learning. But as soon as we apply it to the business world, the model doesn't know what a relevant context is in relation to a company. And context, mind you, is n't the name of your company, or the SIRET number, or because you sell hairdryers or online training courses. That's not what a relevant context is. A relevant context depends on the type of result you want to obtain. So, it's this way of coding the instructions that will completely change the game in terms of the work. Why do we use the terminal rather than the desktop and the web app? Well, quite simply because it's the version that will allow you absolutely perfect control, to have the The latest features were deployed first, giving us the most advanced agentic systems and control. This is the only interface, truly the only one, that will allow us to address all the problems we've encountered so far. Firstly, because in terms of reasoning systems, it's the only one that will allow you to be ultra-surgical in your work, because we'll be able to precisely code the section of missing instructions when the model can't find a solution to a problem. So, we'll be able to activate precise sequence editing. Secondly, because we have an "advisor" mode, which is a supervised mode that will allow you to use your model with another model on top of it. There was a change within 24 hours: Haiku was available with Opus 3.5. Now, they've included it in Sonnet 3.5. This means that, in fact, you don't have just one AI. You have one AI that is overseen by another AI. This means that automatically, when you 're working, you're no longer in the configuration of having Claude alone. You 're automatically in a balancing act with a second AI. So it still allows you to generate reasoning tokens. Whereas when it feels overwhelmed, at some point, it collapses like all AIs. But with this type of system, we'll have a phenomenon called a "ping-pong" restart between the advisor system and the Sonnet system, which will propose solutions to Haiku through the tool advisor function when Haiku is stuck. So both models continue to think, but as you've seen, the instructions need to be coded. So that's exactly what this sidebar allows you to do: depending on what you're doing, you'll code the instructions. So the terminal will allow you to solve all the problems we have today with agentic loop systems. I remind you that every time you interact, it's your prompt that will act. If it does n't give the expected result, these are the three elements that need to be coded: the selection of the context data, what type of action the model should take, and how it will determine whether its actions are correct. Therefore, it needs a verification matrix to know whether it should restart a loop or if it can validate the result. Otherwise, you have a system that will issue tokens; the model will return tokens without understanding what it did in the process. It's just probabilities. Whereas here, we're dealing with selective probabilities, redirected by action and decision-making loops. It's a system. One point that will become apparent very quickly is context management. If there's one thing that will become clear to you fairly quickly, it's that when you start working with it, you'll realize that the context will rapidly become more complex. In the interface you see now, you can't see it, but in fact, when you start the discussions, you'll quickly realize that it takes up space, and this space has a significant impact on the model's operation. This is the context's fill level. So, how do we clean it up? I'll show you. We use a function called a "clear" function. This allows us to clear the context and reset everything. And there, I'm starting from scratch. So, what happens when I use a "clear" function? I clear everything. But what you need to know is that not the entire context is available. Some of the information is occupied by data that isn't visible to you, but which is definitely present in the model. With this " context slash" function, don't worry, we'll go over all the functions in more detail later, but this helps you understand how the whole thing is structured. Here, we can visualize the memory usage map of the model. And it shows you one thing: the entire gray area is the prompt system. Then I have some of the tools, then I have the MCP Tool. So, that means the MCPs are already using 3%. You understand what will probably happen: the MCP is one of the biggest consumers of memory space. Then we have memory files, skill files, and messages. Remember, we're at zero. Finally, we just have a very small context message. So, in this situation, you realize that some memory is already being wasted. We have an area called the "compact buffer." In my case, it's small, but in your case, it will be much larger because you'll see that when the model reaches a compaction zone, it will try to compress the information because the pop-up window is limited to 200,000 tokens in Claude's system. Otherwise, you'll have to upgrade to the 1 million token version, which is three times more expensive, and Claude only offers it on the Max plans. So, if you want to work effectively, you'll always need to work within an optimized zone. It plays a more significant role than that: as the Chroma study showed us, as soon as you're in areas exceeding a certain number of tokens, you experience a loss of contextual understanding quality. This loss of contextual understanding is also related to the number of questions you ask, the quality of the question, and the number of distractors within your document, and therefore your data. The more disruptive elements you add, the more the context will be cluttered, and the model will eventually need to compact the data and make room to stay within the 200,000 token limit. This can be activated manually with the "compact" function. So in this case, you switch to the interface, you assign a "compact" function, but as I'm telling you now, there's still a problem. The "compact" function has an advantage and a disadvantage: you don't really control what you're going to compact, unless you press space and add instruction customization options. We'll see that later. Another point: what's the difference between the "compact" function and the "clear" function? We'll come back to that later. The "clear" function erases all the content. This means that when you restart the discussion, the model will have to reload all the information. When you use "compact," it compresses the information into a memory compression area. This means that you have between 15 and 22% allocated to memory. So you no longer have 200,000 tokens. You have the prompt system, the tools, and 20% of the memory. In short, you have no more than 150,000 tokens left to work with. Personally, in my interface, as you may have noticed, I've disabled it. I'll explain how to do it, but in my specific case, I prefer not to have compression and to manage it manually by exchanging memory files and using my own compaction methods. This allows for 20% more workspace. If you want to delve deeper into mastering AI, in 15 days, you'll be able to go from a simple ChatGPT user to an entrepreneur or employee capable of building useful, clean, auditable, and deployable AI systems for the enterprise. You'll learn how to deploy, structure, and monetize your subscriptions. Imagine a $20 per month ChatGPT subscription that could potentially increase your productivity by 30%. In this ecosystem, I'll provide you with 30 pre-configured AI agents, tools, 80 hours of self-paced training, updates included, and I'll teach you the latest AI agent creation technologies to optimize your time and work. All the information is In the description. So, how does Claude Code work compared to the classic interface? Let's say this system is based on a mode called an approval system. That is, the model will ask you: "Do you want to confirm an action?" I'll give you an example. To do this, you need to be in standard mode. That's what we are here. To find out which mode you 're in, generally, you have the "config" command, and in the config, by pressing Enter, you'll get various pieces of information that we'll discuss in the video. The concept of "auto-compact," which we'll come back to, reasoning, session recovery, code analysis, "verbose output," and the famous "permissions mode." So, the permission system: you see that there's "plan accept default." So, the default system is a system that will, as you've understood, ask questions, but under certain conditions. In fact, there are what we call functions. Functions, broadly speaking, are tools. If you ask it to list them, you'll see one function already: the model definition function. So, let's say I'm working with a model called Haiku. I'll switch to this Haiku model and tell it that I'll be using low-effort reasoning because, generally, I don't need much computing power or reasoning to perform this generative task in terms of token probability. List all the tools in the format: tool name, definition. I pressed the wrong button. Show me the list of tools available with Claude CLI, including MCPs, search, edit, grep, grab, etc. Structure in Key-Value format: Key: tool name, Value: definition. So, the tool has a set of functions, and some of these functions are not exhaustive, meaning it can use them without asking you. However, as soon as it involves modifying, deleting, or editing a file, these three conditions will trigger the request. So that's the approval system. Next, you saw that we have an "auto-accept" function and then a "plan mode" function. So, to be very brief, let's first look at all the elements it will list for you. It's normal; it takes time because it looks at all the directories to understand which tools I have available. It makes a call to the MCP (Model Control Panel). So it will give me feedback on the active MCP. Okay, it will put all of that in Key-Value format. And here's the list. It's pretty cool. So you can see that the official MCP includes GitHub, the file system, PlayWright (which allows you to take control of web pages using browsers), PostgreSQL, MongoDB, SQL database queries, Stripe, Notion, Slack, Gmail, Jira, Asana, Figma, Sentry, and so on, Zapier, Table, and PayPal MCP. So we have a whole set of tools that can be dangerous. I'm talking specifically about the fact that it can access your Slack chat, your Gmail inbox, Jira, and your CRM. This means that, overall, there are things you don't want the model to do. That 's why, by default, in this kind of situation, the model will ask you: "What do you want to do?" Regarding the "auto-accept" system, I'm going to show you different conditions. You press Shift + Tab. And watch what happens. I'll show you a little bit of the screen. Shift + Tab: accept edit on. So, edit acceptance. Then, plan mode and then bypass permission, that's auto-accept. So that's the default mode. So, those are already three conditions that show us how we can interact under three different conditions. The agentic loop is: I enter a request, the model gathers the information, it checks what you're asking for based on the The results you provide will only be executed if you authorize the write or edit option. For example: write a new file to the project directory, a file named "test.mmd". So, if I ask it this type of question and ask it to execute something in default mode, it will ask me: Option 1, do you want me to do this? Option 2, yes, I allow all edits for this session. Option 3, I refuse. And there, I type "I refuse". Okay? So that's how it works. If we're not clear enough, the danger is that the model could modify, delete, or overwrite files. So, we need to define its behavior. Now, how does workflow optimization work in this type of interface? The idea is always the same system: explore, plan, code, validate. This means something very important. In other words, to be able to function, the first thing to do is explore. So the first step, as you can see, is "plan mode." So how do we activate it? There are several ways to activate it, but I 'll show you the simplest method. You press Shift + Tab and activate plan mode. You'll see that by switching to plan mode, I switch models. You'll see, I'm in Haiku, I switch to Sonnet. I'll show you. I'm in Sonnet, I press it again, I switch to Haiku. This means that, in fact, when I switch to plan mode, I move to a higher-level model for planning. And if I need more reasoning capacity, I'll change the "effort" variable in my system. So I need to assign it a value. It's a key-value, it's "effort." And so I tell it that I need a medium effort. Depending on the depth of the reasoning level, I adjust my own reasoning level for the model. For me, the reasoning level is the depth. It's proven to me that there's a direct relationship between the number of layers in the neural system, the number of tokens they can produce, and the relationship they can establish with concepts. So, the more you increase it, the more these three values ​​increase. There's always a balance to be struck; more isn't always better because it can, with excessive verbalization, if it's too verbose... But that's not Claude's tendency in general. I have had hallucinations of Claude, but it's not an overly verbose machine. So, compared to ChatGPT, I experience fewer deviations. Whereas ChatGPT can cause more deviations in "high" or "extra-high" mode; it tells me absurd things. But not Claude. So that means the first thing to do is create a plan for it. The idea is that the system should be capable... If you want to learn how to use AI professionally and scale your work, go to "Mastering the Best of AI 2026." You'll learn to be a certified entrepreneur who automates their business in less than 15 days. The field of AI today is extremely dynamic. You have all the updates included. We're number one in terms of updates. You have over 85 hours of coursework that you complete at your own pace. All the main AI models are included in this training: ChatGPT, ChatGPT Codex, AI agents, and of course Claude, Claude Cowork, and Claude CLI. You'll become a pro with a clear path. A truly professional level: you'll learn how to feed an agent your data in a usable, clean, and useful way. Head to the training courses in the description to optimize my context space in my own way. The key point to pay attention to is the MCP: 10 MCP servers mean 800 popup tokens. So, when When to use "clear" and when to use "compact" is a really interesting topic that I think is very important to address. What you need to know is that a "clear" erases everything. So, when you do a "clear," what happens is that you have to reload the entire usage context. One way or another, what will cost you is that you'll have to resend... You 'll see here, I'll show you another discussion: when you're working, you have what's called a cached area and a write area. The cached area is what's sent to the servers. Once that's cached, you've paid for it. What's sent is what you still have to pay for. And that's the output response. So when you do a "clear," you have to send everything back to the cache. So that means you have to pay for the tokens again. So if your discussion is going off track, if you're experiencing hallucinations, randomness, or behavioral issues, it's best to start with a "clear." That's it, we clear the cache. But if we want to maintain consistency, and on the contrary, ensure the entire discussion flows smoothly, we need to keep the cache. So, instead of "clearing," we should "compact." These are the two crucial points to understand in order to optimize work and costs within the system. What does a nice Claude.md file, a prompt system, look like? Well, it does n't look like: "You're a specialist with 15 years of experience." No, surprisingly, it doesn't resemble the perfect prompts of YouTube influencers saying: "You're a specialist who knows how to make YouTube videos, you have the best practices for getting rich very quickly, and you know how to get me to work so I can go to Dubai and earn €10,000 a month without any skills." No, it looks like this: "This is a NextJS project. You're using the Tailwind router system, and your commands are this one, and your code is this one." That's what a prompt system looks like. And yes, it has nothing to do with what you see on social media. And I'm not the one saying that; it's the official Anthropic documentation. But that's convenient, because that's exactly what I've been teaching you from the start. So we're on the right track, while you have clowns still making magical prompts on YouTube and LinkedIn, because I've seen quite a few. So why is it so short? Because the Claude.md templates are sections that are read when all systems start. Every time you restart a discussion, these are the systems that are read. It's designed to run the entire architecture. So, you'll understand that we can create different types of Claude.md. So, we don't put 30 pages inside, no, we generally put 200 lines, 500 maximum, in a defined format. And this document, which will then contain secondary elements, is actually created by creating secondary files. That is to say, we do something similar to what we do with skills: we create cross-references. "Go read this document for this situation." It's a bit like saying, "I left a sticky note somewhere," well, that's how you should work. So why do we do this? Because we realize that in the way AI works, the more compact you are, the better the AI ​​functions. And what we're trying to do is prompt engineering. That is to say, in your workflow, what we're mainly trying to avoid is you saying, "Okay, I'm going to throw 35 pages of my perfect prompt at you, and the model will work with that." No, models have instruction saturation, which means that if you put too many instructions in the prompt systems, the model will struggle, in its The neural system understands what it needs to do. So, the prompt systems at the root of your Claude, you have several of them. That's the basic operation of the model. To visualize it, scroll down here. There you go. And you see that, in fact, I've set operating rules and protected against the deletion of files or certain elements, and I've simply imposed ambiguity behavior. There isn't much more to it. If I have 30 lines, that's the end of the world. So, this isn't where it's going to happen. The entire architecture will be set up in another file called "agent systems." And that's where it will happen. So, in the "agent" sections, you'll have systems that will act. So why isn't it the same system as inside Claude, the main system? Let me explain: because when you're working, as you saw earlier, you'll encounter pop-up windows. You'll realize that these pop-ups fill up very quickly, and with a web search of 10, 20, or 30 pages, you'll fill 40% of the window. So, how do we continue working without cluttering the environment? We'll create agents. This means we 'll create a persistent memory system with a loaded system, its own context, its own way of working. This system will perform its tasks, store its information, and then we'll retrieve the agent's information and feed it back into the conversation. So, my ultimate goal is to have an agent system work without increasing the current context meter, but rather to retrieve the agent's work. This has several advantages. On the one hand, this prevents overloading the interface, keeping it within areas where writing, reading, and understanding are stable. It also allows me to compartmentalize some rather interesting issues, such as prompt injection and potential attacks, which are actually very specific tasks where, as I mentioned, if you add too many instructions, the models can't keep up and their behavior becomes misaligned. So, ultimately, what will happen is that we code the instructions for your agentic system. But don't worry, agentic systems are more or less coded using the same basic instructions. That is to say, once you understand the skill structures, you'll see that the structure is quite similar. You have a set of allowed tools and prohibited tools. What is the input information? What is the tool's trigger version? Which model is running? Does it have a limited number of cycles it can perform? And then, the agent workflow. So, roughly speaking, as you can see, it's not too complicated in itself. It's a kind of... I call them mini-Claudes, specialized mini-Claudes to which we'll give specific tasks. So the idea, in fact, if we go back to the days of ChatGPT GPTs, you create a series of GPTs. These are GPTs that can be architected, meaning you'll have a "header" system that can coordinate several agents. So you can very well have, in the architectures, a system that calls one agent and then another, or two agents simultaneously. This is built according to certain logics, and you have functions, contexts, and always the MCPs (Memory Control Points), the tools, the code, and memory files. So the architecture is always essentially the same. This leads us to understand that in these structures, we'll have Claude systems that will each have their own secondary Claude.md file, with, for example, a different list of tools, restrictions, because... I I won't give it certain functions. And inside, you'll have specific agents. Generally, you recognize agents because they have a ".md" extension. So here, you have an "agent" directory. You click on it, and you have a system called "Aggregator." You have the version, the description, the number of iterations, the model, the effort count, its role, the activation context and the Python functions it activates, the allowed tools, the variable routing, and the input/output data of the schema system. So the system here is based on an orchestrator that tells it: "You are the orchestrator, you have access to these tools, and you manage the next three agents." So you'll understand that we're going into hyper-specialization. The days of the prompt doing everything are over. On the contrary, you have to be ultra-surgical, and you'll get much better results from the model's operation. This is the key to success today. As you've understood, it will allow us to truly address all the problems. On the one hand, because you'll code specifically where the model crashes, you'll code the system instructions. Where the model blocks, you add a helpful system with the "advisor" function. And when the context becomes corrupted due to excessive length, you create a parallel system that will work in its own window, you retrieve the work, and you switch back to a window where you're more than halfway through, thus avoiding overloading your system with external context. So, in fact, we address all the issues and avoid overloading, I don't know, 250 or 300 instructions on the same agent, because we'll separate it into different interfaces. There's a tool that's quite technical but still worth mentioning. We'll talk about it. I think it's important to introduce the concept so you understand what will happen next. The thing is, the behavior of models is due to what's called Claude.md. That is, when a system understands that it has the right to execute these three commands: server, run test, and lint, what can happen? If the model has too many instructions, it becomes overloaded and misses the mark. If the model, for some reason, hasn't taken this into account in that situation—it can happen, one line in pages and pages... The conversation is long, one line, but I mean one line, and unfortunately, it might execute something you didn't want it to do because that command could cause a problem. Okay? So, in that case, what do we use? We use what's called a "hook." Now, the hook is a system that will be different from the Claude.md system, which is based on semantic recognition. This is a system that understands behavior using only words. A hook, on the other hand, is a setting that retrieves specific behaviors from a model, specifically from the settings.json file. In other words, we have hooks, which we'll call triggers, in different situations. We have 26 different triggers that can track, for example, a tool, the name of a tool, a command, the name of a command, after the result, or before the result. The goal is that if a command or command action, such as "Delete all my files from my hard drive" or "Read the directory numbers of all your API keys stored on your PC," is detected, the model will trigger a command. And this command isn't words; it's a code instruction. And this instruction code is a script that allows us to transform the model's behavior into a deterministic format. That is to say, we It will execute a JavaScript script in the section called "hook". So here, this is where you have scripts that will launch and immediately direct the model. And this isn't probability; these are scripts. So it's execution. That means it executes code. So it either blocks it, activates it, or triggers conditions. If the conditions aren't met, it stops a function. So we're purely in development code logic. The goal of this system is to create completely deterministic behaviors. So we don't use it all the time, only when we have either important files or when we really want to control or correct the 5% where the model doesn't follow the instructions. From there, you should have managed to get all the main information. The key thing to remember is that the sub-agent system is a system for delegating, executing, and retrieving data. The goal is to move away from systems where you have a single, overarching agent to which all questions are directed, and instead create a collection of GPTs (Global Processing Techniques). These GPTs, to put it simply, will now be managed by higher-level agents, and we'll create persistent memory files. These persistent memory files will allow me to address the topic. When you create a working directory, you'll find, in this case, an "agent" directory. You need to look at it, you need to get into the habit of looking at it. These are all the files within the agent system. We can create a memory (for now, I'll keep it simple so as not to complicate things) which can be, as we have here, either in the agents' directory. So this means that a single agent centralizes the writing, and all agents can read it, either within the project—meaning that we consider the entire project to have a memory, so there isn't just one memory; we can create multiple memories. That's the message I want to convey. Therefore, it means that swap files must be architected according to how the agents will communicate with each other to create persistent memory. And there's another topic, which is the concept of "skills." So, what's the difference in a skill system? Memories and Claude files are files that are always fully loaded throughout the entire architecture. So, as you understand, there can be multiple Claudes and multiple memories, and therefore it's a cumulative system. The "skill" systems we're discussing here are progressive loads. That is to say, only the first 200 characters are loaded, and the rest is only fully loaded if certain conditions are met. This reminds us that when you launch a context, it explains why skills take up virtually no space, while, for example, MCPs or memory files, even though I have nothing there, will take up much more space. Simply because all memory files are read according to a directory structure, and all MD files are added together. So you have the main MD, and you have my tiny MD here, which is great because it's so small. Therefore, it's better to use progressive loading of skills under certain conditions. The key is to balance it well, to avoid unnecessary elements, because you understand that this context will load very quickly, and that will negatively impact your workspace. Now you're equipped; you'll be able to take the Claude Anthropic exam at the academy.