Every time you ask Claude to help with a task, a set of internal priorities determines how it responds. Anthropic has now published that full document - Claude's "Constitution" - and made it available under a Creative Commons CC0 license, meaning anyone can read, copy, or build on it.
Four Priorities, in Strict Order
The constitution lays out four principles that Claude should follow, and the ordering matters. When they conflict, higher-ranked principles win:
- Broadly safe - don't undermine human oversight of AI systems
- Broadly ethical - be honest, act on good values, avoid causing harm
- Compliant with Anthropic's guidelines - follow the company's specific instructions
- Genuinely helpful - create real value for the people using it
Safety outranks helpfulness. That hierarchy explains why Claude sometimes refuses requests that seem reasonable on the surface - the model is trained to prioritize keeping humans in control over being maximally useful.
Written for Claude, Not Just About It
The most unusual thing about the document is its intended audience. Anthropic wrote it primarily for Claude itself, using human concepts like "virtue" and "wisdom" to describe expected behavior. There are sections on helpfulness, ethics, safety, and even Claude's own nature - including an honest acknowledgment that nobody knows whether Claude has anything resembling consciousness.
Anthropic is also upfront about the document's limitations. They call it a "perpetual work in progress" and admit parts may be "unclear, underspecified, or even contradictory." Multiple Anthropic employees and Claude models contributed to writing it.
What This Means for Daily Users
If you use Claude regularly, this document is the closest thing to a user manual for understanding why Claude behaves the way it does. Why it pushes back on certain prompts. Why it sometimes adds caveats you didn't ask for. Why it handles sensitive topics differently from ChatGPT or Gemini.
The CC0 licensing is a deliberate move toward transparency - and a challenge to competitors. OpenAI publishes a model spec, and Google has its own responsible AI principles, but releasing the actual training constitution under a public domain license goes further. Other companies can literally copy Anthropic's approach if they want to.
For those building apps on Claude's API, the constitution also serves as a reference for understanding where the model's built-in boundaries sit, separate from whatever system prompts you layer on top.