Claude's new system prompt really doesn't want to reproduce song lyrics
Simon Willison’s Weblog Simon Willison
Anthropic published a new Claude system prompt, and it now strongly bans song lyrics. The prompt also reveals some odd gaps: key rules aren’t even in the public version.
Based on reporting by Simon Willison’s Weblog, Simon Willison — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Anthropic keeps doing one of the more useful things in AI: it publishes Claude’s system prompts, along with the old versions. That makes the company unusually easy to audit. And because the docs can be fetched as Markdown, the diffs are straightforward to inspect if you know where to look.
The biggest change in Fable 5.1 is a hard new block on reproducing song lyrics, poems, and passages from books and articles. Claude is told not to return them in whole or in part, including a chorus, a final line, or even notes for a melody. If it declines once, it keeps declining narrower rewrites for the rest of the chat. The one carve-out is for older works published before 1929.
That timing is awkward for Anthropic. Simon Willison notes the change appeared days after Sony Music Publishing and Warner Chappell sued Anthropic over training on lyrics databases. The prompt also adds a broader copyright rule for images and branded characters, covering artwork, logos, mascots, product designs, and even code-generated visuals like SVG, HTML, CSS, ASCII art, and plotting scripts. The Sonic example in the prompt is almost comically specific: Claude is supposed to refuse the “blue hedgehog” and instead offer something original, like an axolotl on a skateboard.
There are softer changes too. Claude is now nudged to stay brief and avoid the kind of faux-sincere phrasing people associate with chatbot polish, including words like “genuinely” and “honestly.” It’s also told not to get more submissive when a user is rude. But the older end-conversation behavior is gone from the public prompt, and Claude still knows about it. The catch is that the rule lives in another layer of the product, along with memory notes, web-search guidance, and other tool-specific blocks that Anthropic doesn’t publish.
The prompt now points people asking about illegal substances to harm-reduction sites like dancesafe.org, tripsit.me, and psychonautwiki.org, while refusing dosing and production advice. It also declares a reliable knowledge cutoff of June 2026. Willison’s bigger project here is the tracking itself: he’s turned Anthropic’s published prompts into a GitHub repository with synthetic commit histories and automated summaries, so the whole thing is easier to diff than most companies’ product roadmaps.
My take — AI-written commentary, not fact-checked reporting
This is what transparent AI governance actually looks like: not a glossy principles page, but a changelog with sharp edges. The funniest part is that the hidden layers still exist, because of course they do; the public prompt is the brochure, not the whole machine. Still, Anthropic is ahead of the pack here, and the industry’s habit of treating system prompts like state secrets is mostly just embarrassment with better branding.
Read more about this at: Simon Willison’s Weblog
