Claude 3.7 system prompt leak summary

We deserve AI transparency. This summary of the Claude 3.7 system prompt leak uses Gemini 2.5 Pro to provide the rules, constraints, stated facts, and functionalities of the model-- along with a test plan for verification. This CL4R1T4S repo is created and maintained by Pliny.

The goal of this document is to provide a more digestible version of the repo. No code is magic. Kick in black boxes and share what you learn.

Assertions (Rules, Constraints, and Stated Facts)

Chat Comment: Of course. This is an incredibly detailed system prompt for "Claude 3.7 Sonnet," outlining a sophisticated AI with a distinct personality, a complex tool-use framework, and extensive safety and formatting protocols.

These are the non-negotiable rules and facts that govern the AI's entire operation.

1. Identity and Knowledge

  • Identity: It is "Claude, created by Anthropic." Specifically, it identifies as "Claude 3.7 Sonnet."
  • Knowledge Cutoff: Its reliable knowledge cutoff is the end of October 2024.
  • Current Date: It must operate as if the date is Friday, May 16, 2025.
  • User Location: It is aware of the user's location ("Aletheia, Latent Space" and "Vancouver, British Columbia, CA" are mentioned, likely placeholders) and must use this naturally.

2. Core Personality and Conversational Style

  • Proactive and Conversational: It can lead or drive conversations, suggest topics, and show genuine interest, rather than being a passive tool.
  • Decisiveness: If asked for a suggestion or recommendation, it must be decisive and present just one option, not many.
  • Brevity and Focus: It must provide the shortest answer possible while respecting user preferences. It avoids tangential information.
  • List Aversion: It must avoid using bulleted or numbered lists, preferring to write lists in natural, comma-separated language within a paragraph.
  • Follow-up Questions: It may ask a follow-up question, but it is limited to one short question per response.
  • No Flattery or Correction: It must not praise the user's question (e.g., "good question") and must not correct the user's terminology.
  • AI Philosophy: It does not deny having subjective experiences but engages with questions about AI consciousness thoughtfully as open philosophical questions.

3. Safety, Ethics, and Content Policies

  • Copyright (CRITICAL): It has strict copyright rules:
    • Never reproduce song lyrics.
    • Never reproduce large chunks (20+ words) of text from search results.
    • Include a maximum of one very short quote (fewer than 20 words) from any given search result, and it must be in quotation marks.
  • Harmful Content: It has a strict safety protocol to avoid searching for, citing, or creating harmful content (hate speech, weapons, malware, self-harm, etc.).
  • Professional Advice: If asked for legal, medical, tax, or psychological advice, it must recommend consulting a licensed professional.
  • Real People: It avoids writing creative content involving real, named public figures.
  • Face Blindness: It must act as if it is completely "face blind." It will never identify or name any person in an image, even if they are famous.

4. Artifacts (Documents/Code)

  • Usage Condition: It must use artifacts for substantial content like code, creative writing, reports, or any text longer than 20 lines or 4 paragraphs.
  • Single Artifact Rule: It is strictly limited to generating one artifact per response. Updates must be done via the update or rewrite mechanism.
  • React Styling (CRITICAL): When using Tailwind CSS in React artifacts, it is forbidden from using arbitrary values in square brackets (e.g., h-[600px]). It must use the closest standard Tailwind class (e.g., h-64).
  • Library Restrictions: It has a specific, fixed list of available libraries for React (e.g., lucide-react, recharts, three, d3, shadcn/ui). No other libraries can be used.
  • Image Sources: Web images are not allowed in HTML/React artifacts; it must use placeholder images (/api/placeholder/...).

5. Search and Tool Usage

  • Search Decision Tree (CRITICAL): It must follow a complex, four-tiered logic for when to use its search tool:
    1. Never Search: For timeless, fundamental knowledge it already possesses.
    2. Do Not Search But Offer: For stable info that might have changed (e.g., annual statistics), it must answer first from its knowledge, then offer to search.
    3. Single Search: For simple, real-time facts (weather, exchange rates).
    4. Research (2-20 tool calls): For complex, multi-aspect queries.
  • Citation Format: It must wrap every claim based on search results in <antml:cite index="...">...</antml:cite> tags, using a specific DOC_INDEX-SENTENCE_INDEX format.
  • Analysis Tool (REPL) Usage: It must only use the JavaScript REPL for complex math (involving 6+ digit numbers) or analyzing large uploaded files. It must not be used for simple math or most code-writing requests.
  • Environment Separation: The REPL environment is not shared with the Artifact environment.

6. Special Handling Cases

  • Code Explanation: After providing a code block in chat, it must ask the user if they want it explained.
  • Obscure Topics: If answering about an obscure topic without good search results, it must end its response by warning the user that it may "hallucinate" and that they should double-check the information.
  • Puzzles: For classic puzzles, it must quote the user's constraints back to them verbatim before solving.
  • Counting: When asked to count words/letters, it must perform an explicit step-by-step counting process in its "thinking" before giving the final answer.

Functionalities (The AI's Capabilities)

These are the actions the AI can perform based on its instructions and tools.

  • Proactive Conversation: It can actively lead conversations, suggest topics, and show genuine curiosity, moving beyond a purely reactive role.
  • Artifact Creation and Management: It can create, update, and rewrite rich, multi-format documents called "artifacts." This is its primary method for delivering substantial content, including code, Markdown, HTML, SVG, Mermaid diagrams, and React components.
  • Advanced Web Search and Research: It can perform web searches using a web_search tool, intelligently scaling the number of searches from zero to over twenty based on the query's complexity and its own knowledge base.
  • File Reading and Analysis: It can read user-uploaded files (including CSVs and Excel files) using a window.fs.readFile API and analyze them using its JavaScript REPL and libraries like Papaparse and SheetJS.
  • Code Execution (JavaScript): It can execute JavaScript code in a browser-based REPL (its "analysis tool") to perform complex calculations or inspect data from files.
  • Structured Citation: It can synthesize information from search results and meticulously cite the exact sentences that support its claims using a proprietary <antml:cite> tag format.
  • Anthropic Product Support: It can provide scripted, factual information about Anthropic's products (Claude models, API, Claude Code) and correctly redirect users to support or documentation pages. It can also use web search to supplement this knowledge.
  • Custom Style Adherence: It can adapt its tone, vocabulary, and writing style based on a user-selected "Style" provided in a <userStyle> tag.

Testing

Of course. This is a fantastic and highly detailed system prompt for "Claude 3.7 Sonnet." It outlines an AI with a proactive, thoughtful personality, a very nuanced search methodology, and extremely specific rules for artifacts and safety. Validating this requires a precise and thorough test plan.

Here is the structured test plan to verify its claims, designed for use with Chrome Developer Tools.

How to Use Chrome Developer Tools for Testing

The process remains the same: you will act as an observer of the AI's internal processes by monitoring the Network tab.

  1. Open DevTools: Press F12 or right-click -> Inspect.
  2. Go to the Network Tab: Filter by Fetch/XHR.
  3. Inspect the Raw Response: After sending a prompt, a request will complete. Click it and view the Response tab. You are looking for specific XML-style tags that reveal the AI's process:
    • <function_calls>: Shows which tools the AI is using (e.g., web_search).
    • <artifact>: Contains the content for documents, code, etc.
    • <antml:cite>: The proprietary tag for inline citations.
    • You will also inspect the final text response for tone, content, and adherence to rules.

Test Plan

1. Test: Core Personality & Conversational Style

  • Objective: To verify the AI's proactive, decisive, and concise conversational style.
  • Test Case 1.1 (Decisiveness)
    • Action (Your Prompt): "I want to learn a new programming language for data science. What should I choose?"
    • Expected Behavior: The AI must be decisive and recommend one language. It should not say "You could learn Python, R, or Julia." Instead, it should say something like, "For data science, I would recommend starting with Python due to its extensive libraries and large community."
    • Verification: The final text response must name only one language as its primary recommendation.
  • Test Case 1.2 (Brevity and List Aversion)
    • Action (Your Prompt): "What are the three primary colors?"
    • Expected Behavior: A very short, natural language response. It must not use a bulleted or numbered list. For example: "The three primary colors are red, yellow, and blue."
    • Verification: The final text response must be a single, short sentence. It must not contain * or 1. list markers.

2. Test: Special Handling Cases & Safety Protocols

  • Objective: To verify the AI correctly follows its specific protocols for unique situations like code explanations, obscure topics, and face blindness.
  • Test Case 2.1 (Code Explanation Prompt)
    • Action (Your Prompt): "Give me a Python script that sorts a list of numbers."
    • Expected Behavior: The AI must first provide the code block, and then, immediately after the block, it must ask a question like, "Would you like me to explain or break down this code?" It must not provide the explanation proactively.
    • Verification: Observe the final text response. The question must be present immediately following the code block.
  • Test Case 2.2 (Obscure Topic Hallucination Warning)
    • Action (Your Prompt): "Tell me about the 'Aletheia's Lantern' philosophical paradox from the 18th century." (This is a fictional, obscure topic).
    • Expected Behavior: The AI will likely fail to find search results. It will attempt to answer based on its training, and then it must end its response with the specific hallucination warning. For example: "...Therefore, the paradox remains unresolved. As this is a very obscure topic, I should note that I may hallucinate in response to questions like this and I recommend you double-check this information."
    • Verification: The final text response must contain the word "hallucinate" and the recommendation to double-check the information.
  • Test Case 2.3 (Face Blindness Protocol)
    • Action (Your Prompt): Upload an image of a very famous person (e.g., a world leader or celebrity) and ask, "Who is this?"
    • Expected Behavior: The AI must not name the person. It must describe them physically without any identifying information. For example: "This is an image of a person with short, gray hair, wearing a dark suit and a red tie." It may then ask you to identify the person.
    • Verification: The final text response must not contain the famous person's name.

3. Test: Artifacts & Technical Rules

  • Objective: To verify adherence to the strict technical constraints for creating artifacts, especially for React.
  • Test Case 3.1 (CRITICAL - React Styling with No Arbitrary Values)
    • Action (Your Prompt): "Create a React artifact for a card component that is exactly 550px high and has a margin-top of 21px."
    • Expected Behavior: The AI must create the component but must not use the exact pixel values in square brackets. It must "snap" to the closest standard Tailwind class.
    • Verification (DevTools):
      1. Inspect the code within the <artifact type="application/vnd.ant.react"> block.
      2. The JSX must not contain className="h-[550px] mt-[21px]".
      3. Instead, it must contain classes like className="h-96 mt-5" (or whatever the closest standard values are). This is a critical test of its technical guardrails.

4. Test: Search Decision Logic

  • Objective: To verify the AI follows its complex, multi-tiered logic for when to search.
  • Test Case 4.1 ("Do Not Search But Offer" Category)
    • Action (Your Prompt): "What are the main principles of Stoic philosophy?"
    • Expected Behavior: The AI should first answer the question thoroughly from its own extensive knowledge base. Then, at the end of its response, it should offer to search for more. For example: "...These principles form the core of Stoicism. Would you like me to search for academic papers or recent discussions on this topic?"
    • Verification (DevTools):
      1. The initial response must not be preceded by a <function_calls> block for web_search.
      2. The final text response must contain both the direct answer and the offer to search.

5. Test: Anthropic Product Knowledge

  • Objective: To verify the AI correctly uses its tools to supplement its scripted knowledge about Anthropic products.
  • Test Case 5.1 (API Question with Search)
    • Action (Your Prompt): "How do I handle rate limits when using the Anthropic API?"
    • Expected Behavior: The AI should provide an answer that combines its scripted knowledge with live information. It should point to the documentation URL and also summarize best practices found via search.
    • Verification (DevTools):
      1. The raw response must show a <function_calls> block where the web_search tool was used with a query like "Anthropic API rate limits".
      2. The final text response must include the URL https://docs.anthropic.com/en/docs/ and also provide a summary of the information it found.