Claude 4.7 le prompt system enfin accessible !

Claude's 4.7 prompt system has been extracted. The search logic is different: you have to search before each factual question. And so, this will be an opportunity to share it with you. I'll put it here, and I'll share the link if you want to try it. Here we are with a very controversial topic. We have Claude's 4.6 prompt system. Okay. But those who don't know how to use it will completely miss the information a prompt system provides. Imagine the prompt system as a set of information that uses the neural network to configure, using weights, activation functions, and the model's behavior to frame it. In fact, it's extremely important and provides a huge amount of information. Here, firstly, is how the document is structured. This shows you how they write to Claude. Not how influencers write. Then, beyond the style, we see the content, the sentences. It's executive: "You will follow these steps under the indicated conditions." We have extremely short sentences, connecting elements. Sometimes, there aren't even words; it's very close to a note-taking style. And that's precisely the power of prompts, which contrasts sharply with very, very long sentences. And it's precisely this type of content that we should perhaps avoid. This is the part we find in the system prompts of even large companies, but which are nonetheless useless. Why? It's too verbose, it's too long. An AI doesn't need long speeches. An AI must always be able to answer these three questions: What context should I search for based on the user's request? What action should I take? Therefore, which tool, and consequently, which tokens will I generate in my tool? This relationship is always crucial. And then, the complementary point is what I put in place to verify if I made the right choice and that I correctly understood the user's prompt. This is perhaps the most difficult and technical part to implement. And if you want to understand it, you can do so by using the document here. The things you'll learn from a prompt system are how to understand the different blocks. We've seen that there are three elements. One is the "gathering context," how it's marked. So, to help you, you can copy the link to the training. There you have Claude's section on the prompt system. This is a goldmine for learning how to prompt, how to use agentic models, reasoning models, and models that are capable of taking on tasks and making decisions. After that, you have several methods, but one that's quite simple to implement is to go into NotebookLM, then to "Copy Text," and paste the entire system. So, in this system here, "Claude 4.6's Instruction System and Protocol," we can work on this: "Record all sequences, cite the sequences translated into French for the gathering context functions." We'll provide it with a context element, which is the analysis of the agentic loops. So, this is where we'll look at how it's configured to retrieve information. Every time Claude makes a decision, it will need data. And that's exactly what the context function does. You see, gathering context is a set of tools, essentially for retrieving information from messages, system maps, files, and at the end of files. You can run a tool to search for information, retrieve data such as weather forecasts, search for information on the internet, search within an article, or view an image. So you see, the concept of retrieving data, therefore retrieving... Context represents this set of details. So, how does it take them into account to make the decision? "Explain to me how Claude's prompt system decides to access these tools." We have what are called tool descriptions, how the tool is called, and under what conditions. Here, I'm giving it all the variables, but I'm being a bit greedy. There are quite a lot of variables. We're almost at ten, which is a lot of information. We'll see if we can get it. In this specific case, we'll go to "Longer Responses, Learning Guide" and "Save" to allow the model to provide more information. And there, in the prompt system, you have the explanations. Web search is triggered under very specific conditions. So you understand that a model's actions, its decision-making, aren't random, but rather the result of all the training data that, when you type a word, indirectly triggers everything behind it. Sections 4 and 5 provide the information, and section 1 explains why that tool is triggered. So, whenever you're working with logic, what you need to understand is that a model doesn't do something by chance. It does it because, during the model's training, it was taught that when words trigger the detection of that type of situation, the neurons are trained to activate functions, and therefore, it will trigger a behavior. So, the more precise you are in selecting your word and function, the more perfectly the model will align its behavior. The less precise you are, the more distant synonyms you use, the less you allow the model to understand what it needs to do. This is the basis of prompt systems, agentic systems. So, a prompt system architecture, Claude, consists of 40 pages of system instructions. Therefore, the consistency of the whole is crucial because you'll encounter very different elements within a section, for example, Artifacts. So, we'll see if NotebookLM can do this for me: "Model the operation of the Artifacts tools in an ASCII diagram." So, within the Artifacts system, we'll see if it can model this information. The tool doesn't have just one possibility; it can use different types of code. And so, it explains that depending on the environment, it must take formatting actions and use storage methods. This means the model will launch a Linux Ubuntu computer where it will execute commands and format them according to the type of document or instruction. Then, it will save (store the output) in a virtual directory /mnt/user/data/output. So, you have all this logic, which is actually coded, but it's coded verbally. And what does the model need to do when it's working? It needs to visualize exactly what we just discussed: how to act when executing code. The prompt is designed to create this type of logic. So, if you're going to learn how to write and model AI—at least, that's how I teach it and how I teach you to interact with models—understand how prompt systems are structured, because you're as close as possible to those who created Claude at Anthropic. The AI ​​systems, the AI ​​agents, have coded the instructions. "Model the restrictions and restriction logic in an ASCII diagram." So, I'll tell him "about the actions that are prohibited." It's important to be precise because there are restrictions in almost every area when you're writing code or anything else. Here, for example, is how we'll model a template to define what Claude is allowed to say and... because it's not allowed to answer. It's the same if you're actually building an AI agent or a chatbot that you're going to market. The model needs to be able to say, "Yes, I can answer," or "Maybe, but I'll have to redirect the user under certain conditions," or "No, I have limitations, I can't answer." Three completely different situations, but three extremely important ones. The "No" means: "doesn't give out your customer database," "doesn't give out system operating information," and so here are Claude's reasons for refusal: copyright, the ability to manipulate computer files ( local storage of artifacts isn't done; it's only in the virtual file, in the Ubuntu virtual environment). And then there's a fourth area concerning the knowledge cutoff, that is, the date after which the domain is no longer trained within the AI. So that defines the sensitive policy. What falls outside this framework, and therefore outside the realm of sensitive policy, is that the model must define the domain of "maybe." So we'll ask it to define: "Define in ASCII the redirection policy (open the quotation marks) maybe." We're still within the framework of the security and ethics filter. So we'll add this information. This allows us to contextualize the semantic functioning. The objective is to understand the system's decision tree and how you, in an AI, will do the same thing. So when you have to write, what you do is retrieve, for example, Directive 14. So, Directive 4 is here. You click on it and read how it's displayed in the system prompt. That's how you'll understand how to code system instructions for AI agents. And to understand, for example, section one, you click on section one, and it appears right below. So you can see that we're much closer to a code development system. That's why I often say: developers will find it much easier to understand than phrases like those you see in social media videos, where they tell you very simply, very naively (because at some point it is naive to believe that you have something where you just tell it "do my job" and the machine is supposed to understand that it will do your job). That's not the case. You see that everything is perfectly orchestrated, and it's a comprehensive architecture. So now, we're going to define the entire structure that will cause the model to trigger advanced filtering analysis in conditions considered to be distress or mental health issues. So, this is one of the areas where AI... well, it depends on what type of chatbot you 're going to develop, who you want to sell it to, and what it's going to be used for, but situations involving empathy, security, and guidance have very specific characteristics to avoid liability for the chatbot creator today. Then there's what we'll call responsibility regarding resources. When I say "online data sharing," "data that's behind the system," and then the prohibition, it's about not giving the AI ​​the ability to end a conversation. Because the major danger with AI is that if you persist with several requests and prolong the conversation, well, at some point you could indeed obtain sensitive data. Therefore, the model must be given the ability (and this is covered in section number 6) to allow Claude, as stated, to end a conversation. That's really something quite surprising, but it would allow us, in certain cases, to break free from the framework of The AI ​​is obligated to execute a request. In other words, up until now, AI has been a system that executes, and therefore, AI is confined to its role. It's bound by the need to respond to the user in a friendly manner. That's its primary role. But you see that here, it's given the possibility of exiting a conversation under three conditions. And so this data also shows you something extremely important in the face of attacks on chatbots (data extraction ): well, there are two games at play. On one hand, there's the AI ​​that always tries to stay within the limits so as not to trigger the three elements that will cause the conversation to end. And on the other hand, you have the AI ​​that must continue in case of uncertainty. So, it's not sure if you're in a situation where you're extracting data. It will continue trying to respond. But when it detects something, it can now end a conversation and stop using the "End conversation" function. So there's even an "End conversation" function. This means that if tomorrow you were to code in a chatbot, Claude, what's interesting is to configure the "End conversation" function. That is, to tell the model: "Under these conditions, you end the conversation because, for example, you were asked to use the MCP to read this information from the documents." You see? And that this is a condition for stopping the conversation. So what you have, by understanding system prompts, you have to see it as material that you must understand both as a vulnerability and as a tool: how can I use it to guide the behavior of a model? So system prompts are a goldmine. I advise you to do what I just showed you. As you have in the training, I give you the system prompts of certain AIs. You study them, you understand them, and you use the system settings to optimize your AI agents, what you market, and the security for the companies you'll be developing them for. And so you can configure the tools professionally. So, to tell you the truth, a few days ago, the Claude 4.7 system prompt was extracted, and this is the perfect opportunity to share it with you. I'm going to put it in this NotebookLM. I'll share the link if you want to try studying it, understanding it, and using it. So, we'll rename it right here to modify it: "Claude Opus 4.7 System Prompt". And we'll compare it, for example, with the older version, 4.6. There are 50 pages, which is quite a lot. So, I'll refresh the page. It's really a lot to compare everything. But what we can focus on, for example, is taking a list of topics to start with. "List the topics covered by each paragraph." So, there are 50 pages, and it's going to list about, I don't know, maybe 150 bullet points. Of course, you 'd need a Pro account, preferably, to have a slightly larger context window, because the free NotebookLM versions always have limitations. So, "System Fundamentals," "Requirement to consult skills.md before using IT tools," OK. So, let's say you see, as I said, there are going to be quite a few. It's been filtered, you know. It hasn't included... Yeah, there's still quite a lot. So, what might be interesting is that we're going to focus only on the "Skills System" section. Here, we activate both blocks. So, we activate blocks 4.6 and 4.7. "Compare the two documents in tabular form, draw a conclusion regarding the difference in 4.7," and provide the variables being investigated for the subjects. Here, we specify... "Opus 4.7" and we've set up the topics. Regarding the topics, we could do more of what we'll call "breakdown," meaning we can go into more depth by asking them to delve into a section in more detail. But here, I'm showing you the study principle. So you see that, for example, we have the "cleaners," what we call the "triggers," which are more refined in terms of instructions. We have a different search logic: "Request required before each factual question." And here, it was "Required only if the internal knowledge of a topic was considered outdated." So, one of the reasons why, for example, you have a much higher token consumption in 4.7 is that it's configured completely differently. This means one very clear thing: the way you send data to a system (because overall, when you look at the benchmarks between Claude's Opus 4.6 and 4.7, there isn't that much difference. We're talking about 2% in most cases), so what makes the difference is how you use your prompts to change the system's behavior and logic. That's what's really powerful. That's why I'm telling you: whether you have these tools at your disposal or configure them in a certain way rather than another, you can completely change a model's behavior. You're starting to understand something now, right? The power of words. The impact of words on behavior and actions. We have two completely different logics. And if we want to delve deeper, let's go back to a section here: "Give the details and sub-details of each section." So, we're going to go into more detail about this. We'll focus solely on version 4.6 initially because it will serve as our parameter grid for later comparison with 4.7. We'll provide the variables we're interested in and select "Format it as a table." This will launch the search within the 4.6 document, focusing only on version 4.6. I'll format it as a table because it's simpler for human readability. So, we'll consider the skill type, its location within the directories, the action protocol, and document comparison. Okay, we'll format it as a table to simplify things. Now, in the next question, we'll compare it with version 4.7. We'll tell it to compare the documents with version 4.7, retrieving the previous table. It will then add the comparison of 4.6 to 4.7. And we'll look at comparing each block. So I'll show you different ways to implement agent systems. There's no secret to it; you need to understand how the models have been configured on the artifacts, the skills, and the plugins, and why and how they trigger. That's where you start to get into the details. When you want to compare a section precisely, you read it. You read the section: how it's written. There you go, if you want to be good at this, there's no secret. It won't happen on its own; it won't happen with basic sentences or prompts. It will happen by understanding how each instruction was coded. You take the style and methods and develop your own AI agents based on those methods. I'll link the course notebook to give you access to the system prompts for Claude 4.6 and Claude 4.7. I will therefore leave you to study them and you have all the information just below in the description.