Anthropic publishes the system prompts for its Claude consumer apps, meaning Claude.ai and the mobile apps, along with the historic changes. Simon Willison has been tracking this for months with a homemade Git timeline. Now he’s laid Fable 5 and Fable 5.1 side by side, and the most interesting difference jumps right out.
Fable 5.1 gets a hefty new section about song lyrics. Claude no longer reproduces lyrics, poems, or passages from books, in whole or in part, not the last lines, not a chorus, not a line someone pastes in on its own and calls their own song. Once Claude has declined the request, it keeps declining reworded or narrower versions for the rest of the conversation. Works first published before 1929 are still fine, a Shakespeare sonnet for instance. That this section shows up days after the news that Sony Music Publishing and Warner Chappell are suing Anthropic over training on lyrics databases is, Willison figures, no coincidence. Neither do I.
Right after it comes a second new section that forbids drawing protected works. And not just with a text-to-image model, but with anything Claude draws in code: SVG, canvas, CSS, HTML mockups, even ASCII art. No known character, no logo, no album cover. Changing the pose, color, or style doesn’t make it original. The prompt includes a charming example: a user wants a birthday banner for his son with a fast blue hedgehog on it. Claude answers that that’s Sonic, and offers an original speedster instead, a grinning axolotl on a skateboard. Willison tried the prompt, and sure enough, the model refuses Sonic and delivers the axolotl.
A few more changes are worth a mention. Anthropic has taught Claude to avoid words like “genuinely,” “honestly,” and “straightforward,” because they come off as disingenuous. The old rule about issuing a warning and then ending a conversation over persistently abusive behavior is gone. Instead Claude is meant to stay calm and matter-of-fact without becoming submissive. There’s also a new section on illegal substances that, for the first time, includes external links, to harm-reduction sites like dancesafe.org and tripsit.me. Willison notes these are the first non-Anthropic URLs ever to appear in a Claude system prompt.
The real kicker comes at the end. Willison asked Fable 5.1 about the deleted end_conversation tool, and the model described it in detail, even though it no longer appears in the published prompt. The explanation: the core prompt gets extended with feature- and tool-specific blocks that vary by session, and Anthropic doesn’t publish those. Which means what we see isn’t the whole prompt. It never was, but it’s worth keeping in mind.
Sources: Simon Willison: Claude’s new system prompt really doesn’t want to reproduce song lyrics · Anthropic: System prompts · The Guardian: Anthropic sued over song lyrics