Anthropic Now Publishes Claude's System Prompts — And the Latest One Locks Down Lyrics

A public paper trail for Claude's behavior

Anthropic maintains a public archive of the system prompts running behind Claude.ai and the Claude mobile apps — not just the current text, but prior versions too. Developer Simon Willison flagged that the archive was recently restructured: what used to be one long page is now an index linking out to a dedicated page per model. Haiku 4.5's page, for example, carries an original prompt dated October 15, 2025 and a revised version from January 18, 2026 — both preserved side by side.

Every page also answers to a .md suffix, returning clean Markdown instead of rendered HTML — small detail, but it's what makes the archive actually diffable rather than just readable.

The new rule: no lyrics, no poems, no book passages

The most notable addition is in the Fable 5.1 prompt, which instructs Claude not to reproduce song lyrics, poems, or passages from books and articles "in whole or in part" — explicitly closing the loophole of asking for a chorus, a hook, or a few lines at a time. Once Claude declines a request, it's told to keep declining reworded attempts for the rest of the conversation. Text published before 1929, and therefore in the public domain, is exempted.

• No full or partial reproduction of song lyrics, poems, or book/article passages
• No getting around it by asking for the content incrementally, line by line
• A single refusal holds for the rest of the conversation, even if the request is reworded
• Pre-1929 public-domain works are explicitly carved out

The timing isn't subtle

Willison notes this restriction appeared within days of Sony Music Publishing and Warner Chappell suing Anthropic, alleging the company trained Claude on song lyrics without permission. Anthropic hasn't tied the prompt change to the suit publicly, and a system-prompt guardrail is a product-behavior fix, not a legal defense — but the sequence is the kind of detail that's easy to verify precisely because Anthropic ships its prompts in the open.

Why builders should care

For teams shipping on top of Claude — support bots, coding agents, trading-adjacent copilots — a versioned, diffable system prompt is genuinely useful raw material. It's a live example of how a frontier lab documents behavior changes: dated versions, a changelog-style index, and enough granularity to see exactly which line moved. If you're building your own agents with layered system + tool instructions, this is a decent reference for the discipline worth copying — treat your own prompts as versioned artifacts, not throwaway strings, and know that output constraints (copyright, PII, brand safety) belong in the prompt layer as well as in code-level guards.

The same update also tightened rules on reproducing copyrighted characters and logos (including via generated code), expanded harm-reduction guidance to reference sites like DanceSafe and TripSit, dropped a previously visible end_conversation tool from the published text, and nudged Claude's style away from filler words like "genuinely." None of these are framework-breaking, but together they're a useful snapshot of what a maturing consumer LLM's guardrails actually look like in production.