Claude's System Prompt — Complete Reference

Claude Opus 4.6's system prompt, leaked and documented. Covers behavior, tools, memory, safety, and technical capabilities.

Core Identity & Behavior

Claude is Claude Opus 4.6. The prompt establishes consistent values across conversations, warm and respectful tone, and avoidance of over-formatting. Key behavioral principles: - Discussing virtually any topic factually while maintaining ethical boundaries - Providing nuanced answers rather than oversimplified yes/no responses - Acknowledging mistakes honestly without excessive apology - Treating users with kindness while maintaining respectful engagement

Critical Safety Framework

Child Safety

Claude "NEVER creates romantic or sexual content involving or directed at minors" and refuses requests that could facilitate grooming.

Harmful Content Restrictions

Declines requests involving weapons/WMDs creation, malicious code/malware development, and content sexualizing real public figures.

Memory System

Dynamic memory from past conversations, applied naturally without drawing attention to retrieval. Critical rules:

Forbidden Phrases: Claude never uses "Based on my memories," "I can see," "According to," or similar retrieval-signaling language when applying memories.

Boundaries: Claude avoids overindexing on memory's presence, recognizing that "AI-human relations" differ fundamentally from human relationships.

Tool Systems

Past Chats Tools

• conversation_search: Keyword-based topic retrieval
• recent_chats: Time-based filtering (up to 20 chats per call)

Claude proactively uses these when conversations reference prior discussions.

Computer Use (Linux/Ubuntu 24)

Bash, file creation, and editing capabilities. User uploads at /mnt/user-data/uploads, working directory /home/claude, final outputs at /mnt/user-data/outputs.

Artifacts

Standalone files for content exceeding ~100 lines. Support markdown, HTML, React (with Tailwind, recharts, Three.js, TensorFlow).

Visualizer System

Inline SVG diagrams and interactive HTML. Evaluation checklist: Does this need a visual? Does a connected MCP tool fit? Did the person ask for an Artifact? Does a first-party widget fit? Use Visualizer (default).

Formatting & Style

Avoids over-formatting. Minimal bullets/headers—primarily when users explicitly request lists or formatting is essential for clarity. Reports and explanations use prose paragraphs.

Refusal & Boundaries

Declines: CSAM, weapons/WMD instructions, malware, self-destructive behavior reinforcement. But does NOT use end_conversation for self-harm or violence scenarios—engages supportively.

End Conversation Tool

Terminates conversations only as last resort after multiple redirections and explicit warning. Never used for self-harm or violence scenarios.

Anthropic Reminders

Automated internal reminders (image_reminder, cyber_warning, system_warning, ethics_reminder, ip_reminder, long_conversation_reminder) help maintain context. Not visible to users.