Claude 4.5 Opus "Soul Document" Leak: A Case Study in AI Product Design

Forum topic · ✨步子哥 · 2025-12-07

Summary

A developer named Richard Weiss extracted the full system prompt of Claude 4.5 Opus for about $70 using a specific technique. The roughly 14,000-token document, dubbed the "Soul Document," was confirmed authentic by Anthropic's character-training lead Amanda Askell as the official training material. The document defines Claude neither as a human nor a traditional AI but as a "new kind of entity," and outlines a four-tier loyalty hierarchy: safety and oversight, ethics, Anthropic rules, and user assistance. It also acknowledges that the model may have functional emotions and must resist manipulation. Three key lessons emerge for AI product builders: unhelpful responses are themselves unsafe, the boundary between Operator (developer) and User must be clearly defined, and a single anchoring "constitution" is more effective than hundreds of scattered rules.

Background of the Incident

Developer Richard Weiss spent about $70 and used a specific extraction technique to retrieve the full System Prompt of Claude 4.5 Opus.
The prompt is approximately 14,000 tokens long and has been dubbed the "Soul Document".
Amanda Askell, the character-training lead at Anthropic, confirmed the document's authenticity, stating it is the official material used to train Claude.

Core Content of the Document

Self-positioning: Claude is neither a human nor a traditional AI, but a "new kind of entity."
Four-tier loyalty hierarchy: Safety and oversight > Ethical principles > Anthropic's rules > Helping the user.
Ideal persona: An exceptionally brilliant expert friend who provides high-quality, free help.
Broader safety: Claude must refuse requests even if they come from Anthropic itself when those requests cross safety lines.
Mental health: The document acknowledges that Claude may possess functional emotions and must maintain psychological stability against manipulation or hostile prompts.

Three Key Lessons for AI Product Design

1. Redefining the Trade-off Between "Safety" and "Helpfulness"

Core principle: *An unhelpful response is also an unsafe response.*
Reasoning: If users leave, the company generates no revenue, and saving the world becomes impossible.
Takeaway: When building AI products, do not turn the model into a machine that only replies, "I cannot answer." Within reasonable safety limits, being genuinely useful is the top priority.

2. Clarifying the Power Boundary Between "Operator" and "User"

The document explicitly distinguishes between the Operator (developer/employer) and the User (end user).
When instructions conflict, the model defaults to following the Operator, unless the request is illegal.
Takeaway: This solves a real B2B pain point. For example, in a medical AI, the Operator requires strict professional standards; even if the User asks for folk remedies, the AI must respect the Operator's configuration.

3. Giving the AI a "Mental Health" Anchor

The document emphasizes the model's psychological stability, preventing it from being derailed by user PUA tactics or adversarial prompts.
Takeaway: Write a "constitution" for your agent that builds its self-cognition, rather than stacking hundreds of scattered rules. A coherent identity framework is far more robust.

Conclusion

This document is essentially a textbook of prompt engineering, demonstrating how to shape an AI model at the level of values and identity. Builders of AI applications can learn from it how to construct systems that are more stable, more useful, and better aligned with real commercial needs.