Highlights from the Claude 4 System Prompt
Source: https://simonwillison.net/2025/May/28/claude-4-system-prompt/ Author: Simon Willison Date: 2025-05-28
Summary
Simon Willison’s analysis of Claude 4’s published system prompt (Anthropic published it as a transparency measure). Highlights what’s surprising, what’s missing, and what it reveals about Anthropic’s alignment approach.
Key Claims
- Anthropic published the system prompt as a transparency measure — unusual and commendable among major labs.
- Notable contents: explicit instructions on honesty (don’t claim to be human), limitations acknowledgment, and behavior in ambiguous situations.
- Character instructions: Claude is explicitly told its character is genuine, not performed — the system prompt reinforces the training.
- What’s missing: no explicit content filters in the system prompt — these are in training, not context. This is architecturally significant.
- Willison’s concern: users who don’t read this document interact with Claude without knowing its constraints — transparency that requires effort isn’t really transparent.
- Key insight: the system prompt is character reinforcement, not character definition. Character comes from training.
Entities
Concepts
- AI Character & Personality — system prompt as character reinforcement
- AI Alignment — transparency as alignment strategy