Claude's new system prompt restricts song lyric reproduction

Anthropic continues to publish the system prompts governing its consumer Claude applications, Claude.ai and the Claude mobile apps, though not Claude Cowork or Claude Code. The company maintains both current prompts and a historical record of changes, a practice independent developer Simon Willison has praised, writing that he “loves” the disclosure.

Willison reports that Anthropic previously kept all prompt versions on a single page. On checking recently, he found the system had changed. According to his account, the newly published prompt places notably strict limits on the reproduction of song lyrics, going further than earlier versions in restricting the model from outputting copyrighted lyric text.

The summary provided does not detail the specific wording Anthropic now uses to enforce this restriction, nor does it specify what prompted the tightening. Willison’s post is presented as an observation of the change in the published documentation rather than a statement from Anthropic explaining the reasoning behind it.

The publication of these prompts, and the ability to track their revisions over time, offers outside observers a rare direct view into how Anthropic adjusts model behavior at the instruction level, separate from underlying model training. The lyrics restriction is noted as one visible example of that ongoing adjustment, documented publicly rather than disclosed only through user-side testing.