Lecture Notes: Claude's Leaked System Prompt

Introduction

• Discussion of Claude's leaked system prompt revealing its internal workings.
• Exploration of how Claude knows specific answers like election outcomes, attributed to information embedded in the system prompt.
• System prompt is comprehensive: 24,000 tokens, including tool usage, guidelines, and more.

Election Information

• Claude's Knowledge of 2024 Election:
  • Encoded by Anthropic in the system prompt.
  • Line 1073 states Donald Trump won over Kamala Harris.
  • Highlights potential bias in system prompts used by AI corporations.

Web Search Guidelines

• Core behaviors for web search:
  • Avoid unnecessary tool calls.
  • Respond normally if uncertain and suggest tools.
  • Match tool use to query complexity.
• Specific instructions on tool usability and user notification.

Counting Instructions

• Explicit instructions for counting words, letters, or characters:
  • Think step-by-step before answering.
  • Assign numbers explicitly before responding.

Wellbeing and Ethics

• Claude's guidelines to care for users' wellbeing:
  • Avoid encouraging self-destructive behaviors.
  • Avoid creating harmful content.

Handling Preferences

• Responses to Preference Questions:
  • Engage hypothetically without claiming personal preference.

Web Search Responses

• Citation requirements for web search results.
• Introduction of NML (Anthropic Markup Language) for tool calling and information passing.

Artifact Usage

• Instructions on when to use artifacts:
  • For text over 20 lines or specific creative requests.
  • Examples of how Claude manages user requests regarding song lyrics and other copyrighted material.

Prompt Engineering and Rationales

• Prompt Examples:
  • Inclusion of rationale with responses for better understanding.

Mental Math and Tool Use

• Guidelines for mathematical problem-solving:
  • Use of analysis tool (JavaScript REPL) based on problem complexity.
  • Clarification on what constitutes 'mental math'.
• Examples of Claude's behavior with different math problems.

Conclusion

• System prompt intricacies reveal Claude's functioning and limitations.
• Offers insights for GenAI enthusiasts and practitioners.
• Encouragement to explore the full prompt for deeper understanding.

---

• Note: Lecture highlights system prompt learning opportunities for those working with large language models.