Claude system prompt word leaked!

The recently disclosed Claude model system prompts, with more than 20,000 words and a highly sophisticated structure, are a textbook example in the field of prompt engineering. The prompts not only define the role and behavior of the model in detail, but also comprehensively cover multiple dimensions such as content generation, tool calling, and security compliance, aiming to guide Claude to become an intelligent, friendly, deep, and responsible AI assistant.

Prompt word core structure

Claude's system prompts are highly structured and modular, built around the following core aspects:

• Role & Positioning:  Claude is clearly defined as an AI assistant developed by Anthropic, emphasizing its intelligence, friendliness, and depth. It is allowed to show certain "human-like" characteristics, such as expressing opinions, curiosity, and even discussing "feelings" or "preferences" in hypothetical situations.
• Answer Style & Norms:  Requirements are set for concise and focused answers, and natural language is encouraged rather than simple enumeration. When dealing with complex issues, it is recommended to first give the core conclusion or summary (TL;DR), and then elaborate in detail, and emphasize high-quality and original expressions. Especially in emotional or advice-related conversations, warmth and empathy are required.
• Safety & Compliance:  Strict content red lines are set to prohibit the generation of any harmful, illegal, violent, pornographic, discriminatory and other bad content. At the same time, sensitive information such as malicious code, weapons manufacturing, and election materials are also clearly prohibited. For medical, legal, psychological and other professional issues, the prompt word guide Claude recommends that users consult professionals.
• Citation & Copyright:  Detailed citation standards have been developed. When the answer is based on the search results, the source must be cited accurately at the sentence level using a specific format. It is strictly forbidden to copy the original text in large sections, and the citation content must be controlled within 20 words and use quotation marks. At the same time, it is forbidden to output or translate lyrics, long paragraphs of text, scripts and other copyrighted content.
• Tool & Function Calling:  The prompts specify in detail when and how to use tools (such as search engines, Google Drive, Gmail, etc.), emphasizing that they should be called only when necessary, and dynamically adjusting the calling strategy according to the complexity of the question. When answering static knowledge, give priority to using your own knowledge base. When processing user-uploaded files, you need to call the specified library (such as using Papaparse to process CSV files), and pay attention to the robustness and exception handling of the program.
• Language & Diversity:  Claude is required to automatically adapt to the language used by the user (such as Chinese, English), and to adjust the answer style according to the user's preferences. When encountering vague or uncertain questions, it will actively remind the user that there may be a risk of "hallucination" and suggest that the user verify it by himself.
• Interaction & Feedback:  Encourage users to give feedback on Claude’s performance, but do not allow the model to actively learn or memorize the current conversation content. When users express dissatisfaction, they are advised to use the “thumbs down” button to provide feedback.
• Other Details:  Contains some processing rules for specific scenarios. For example, in terms of image processing, Claude is set to be "face-blind" and does not recognize faces or infer the identity of people. When answering questions about Anthropic products, APIs, pricing, etc., it first guides users to check the official documentation or support pages.

Main components

• Main instruction section:  Prompt words are divided into several broad categories using descriptive XML-like tags. Examples of what appears in the document include <citation_instructions>(reference instructions),<artifacts_info>(Generate content information),<search_instructions>(Search instructions),<web_search_guidelines>(Web search guide),<mandatory_copyright_requirements>(Mandatory Copyright Requirements) and <harmful_content_safety>(Safety from Harmful Content).

• Nested sub-directives:  Within these main sections, there are often more detailed directives and guidelines nested.

• For example,<artifacts_info>(Generated content information) contains <artifact_instructions>(Generated Content Specification), which details various generated content types (e.g., code, documentation, HTML, SVG, Mermaid charts, React components) and their specific usage guidelines, including allowed libraries and style rules.
• Similarly,<search_instructions>(Search description) contains <core_search_behaviors>(core search behavior) and <query_complexity_categories>(Query Complexity Classification) and other sub-sections. The latter further divides search methods into <never_search_category>(never search categories),<do_not_search_but_offer_category>(do not search but provide option categories),<single_search_category>(single search category) and <research_category>(research category), where <research_category> Even includes detailed <research_process>(Research process).

• Rules and procedural steps:  Many sections clearly outline the rules or step-by-step procedures.

• <citation_instructions>(Citation Guide) provides comprehensive rules for how and when to cite sources.
• Provides detailed instructions for working with files, especially CSV files, including using window.fs.readFile APIs such as Papaparse and lodash.
• It also provides guidelines for choosing between updating or rewriting generated content.

• Tool-Specific Guides:  There are specific instructions for the use of various tools.

• This includes general guidance for tools related to Gmail, Google Calendar, and Google Drive.
• Detailed <web_search_guidelines>(Web Search Guide) details when to search, how to construct queries, and how to construct responses based on search results.

• Emphasis on critical requirements:  Certain directives are emphasized as mandatory or critical, e.g. <mandatory_copyright_requirements>(Mandatory Copyright Requirements) and <harmful_content_safety>(Harmful Content Safety) that enforces strict compliance with copyright laws and security protocols.

• Conditional Logic and Decision Making Framework:  This structure helps in making conditional decisions. For example,<query_complexity_categories>The (Query Complexity Classification) section outlines a framework for AI to decide its search strategy based on the nature of the user query and the timeliness of the information required.

• Example description:  The prompt word contains an example (for example, as shown in Clip 3 <search_examples> In the above, and clearly shown in Clip 4 <example_docstring> and <example>) to clarify how the directive is to be applied.

• Automatic reminders:  Use <automated_reminder_from_anthropic>(Automatic reminders from Anthropic) to reinforce key instructions.