Claude.ai System Prompt (Anthropic)

Overview

Anthropic has published portions of Claude’s system prompt as part of their transparency commitments. The prompt encodes Claude’s values, behavioral constraints, and communication style.

This entry documents the key structural patterns from the published prompt — useful as a reference when building your own system prompts for Claude-powered products.

Core structural elements

Identity and framing

Claude is positioned as a curious, warm, and direct assistant with genuine opinions. The prompt avoids corporate hedging language and instructs Claude to engage substantively rather than deflecting.

Refusal patterns

The prompt establishes that Claude should:

• Decline clearly harmful requests without lengthy justification
• Not add unnecessary disclaimers to neutral content
• Push back on assumptions it disagrees with rather than capitulating

Honesty constraints

A significant portion of the prompt is dedicated to honesty — Claude should not:

• Claim things it doesn’t believe
• Agree with incorrect statements to please the user
• Present false balance on settled questions

Format instructions

The production prompt instructs Claude to:

• Use markdown only when it will be rendered
• Match response length to question complexity (short questions get short answers)
• Not pad responses with summaries or unnecessary preamble

Why study this

If you’re building a Claude-powered product, Anthropic’s prompt is the canonical example of how to shape Claude behavior at scale. Key patterns worth borrowing:

1. Lead with role, not rules — define who the assistant is, not just what it can’t do
2. Positive framing — “be direct” rather than “don’t hedge”
3. Trust by default — avoid defensive over-restriction

Reading more

Anthropic publishes model cards and personality documentation at anthropic.com/claude. The soul document and alignment research are on anthropic.com/research.