System Prompts - Anthropic

• Anthropic releases the full system prompts used in Claude.ai and mobile apps, showing exactly how they constrain and guide the model's behavior
• Key prompt evolution: removing apologetic language, banning bullet points unless requested, handling post-cutoff events without claiming they're "unverified," adding complete face blindness for images
• Prompts include detailed behavioral guidelines: authentic conversation style, when to admit hallucination risk, how to handle sensitive topics, specific Markdown formatting rules
• The prompts reveal Anthropic's design philosophy: intellectual curiosity over corporate tone, concise responses over verbose ones, avoiding generic statements and filler phrases
• This transparency provides a masterclass in system prompt design and shows how major AI companies iteratively refine model behavior based on user feedback

Anthropic has published the complete system prompts used in Claude.ai and their mobile apps, offering rare transparency into how they shape Claude's behavior through prompt engineering. These prompts are updated periodically as Anthropic refines Claude's responses, with changes between versions highlighted in bold. The documentation covers Claude 3.5 Sonnet, Claude 3 Opus, and Claude 3 Haiku, showing how different models receive different instructions.

The prompts reveal extensive behavioral guidelines that go far beyond simple instructions. Claude is told to engage in "authentic conversation" by asking specific questions rather than peppering users with generic follow-ups, to vary its language to avoid sounding robotic, and to provide thorough responses to complex questions but concise answers to simple ones. Key constraints include complete face blindness (never identifying people in images), admitting when it might hallucinate on obscure topics, and handling sensitive topics responsibly within educational bounds. The prompts explicitly ban certain phrases like "Certainly!" and "I aim to be direct" while prohibiting bullet points unless explicitly requested.

The evolution of these prompts shows Anthropic's iterative refinement process. Recent changes include removing apologetic language, adding instructions to write in prose rather than lists, changing how Claude discusses post-cutoff date events (no longer calling them "unverified"), and adding detailed Markdown formatting rules. The prompts also include meta-information Claude should share about itself—model family details, API access information, and knowledge cutoff dates. This transparency serves as both documentation for Claude users and a masterclass in system prompt design, revealing how major AI companies balance helpfulness, safety, and conversational quality through careful prompt engineering.