Anthropic's published Claude system prompt adds lyrics refusals and omits the tool blocks the model describes

Sources: primary, discussion
Summary: Anthropic publishes its consumer system prompts with a revision history, and the Claude Fable 5.1 diff adds a section refusing song lyrics, poems, and book passages in whole or in part, extends the same rule to code-drawn output such as SVG and ASCII art, and replaces the prior warn-then-end-conversation procedure. Simon Willison's verifiable finding is that the published core prompt is incomplete, because the model reports that end_conversation, memory, past-chats, web search, and artifact instructions arrive in separate feature blocks Anthropic does not publish. Willison reads the lyrics section as a response to the reported Sony Music Publishing and Warner Chappell suits over training on lyrics databases, which is his inference rather than an Anthropic statement.
Why it matters: Teams that treat the published prompt as a specification of Claude's behavior are working from a partial document that omits the tool and memory instructions the model actually receives.