Claude's system prompt bans lyrics and brand mascots

Anthropic published the system prompts for its Claude consumer applications on 2 September 2026, and the Fable 5.1 text carries a new section instructing the model to refuse song lyrics, poems and passages from books and articles, in whole or in part. The release covers Claude.ai and the mobile apps but not Claude Code or Claude Cowork, according to Simonwillison.

At a glance

• The refusal covers a chorus, a hook, the last lines, a melody written out note by note, and lines the person pastes in one at a time while describing them as their own song.
• Sony Music Publishing and Warner Chappell are suing Anthropic over training on song lyrics databases, and the new prompt section appeared days after that news broke, according to Simonwillison.
• A parallel rule bars drawing known characters, mascots, logos and album covers in generated code, including SVG, canvas, CSS, HTML mockups, plotting scripts and ASCII art, with pose or style changes not counting as original.

The timing reads as defensive, with the clause landing days after the publishers' suit over lyrics training data, and the visual paragraph extending the same logic to code-drawn output, a surface that likely mattered less before Claude became competent at SVG. For anyone shipping on the consumer apps, the operational detail is the stickiness: one refusal now governs the rest of the conversation, so reworded follow-ups fail by design rather than by chance.

Works first published before 1929 are the section's only exemption

The new clause tells Claude not to reproduce song lyrics, poems or passages from books and articles in any amount, and to keep declining narrower or reworded versions of the request for the rest of the conversation once it has refused. It offers to describe or analyze the work instead.

Material first published before 1929 is exempt, with a Shakespeare sonnet, a Keats ode and the Italian libretto of a Puccini aria given as examples, though the prompt tells Claude to rely on its own knowledge of a work's date rather than the person's say-so and to decline when it is unsure.

The visual rule covers SVG, canvas, CSS and ASCII art alike

The same section extends to visual and designed works, including anything Claude draws with code: SVG, canvas, CSS, HTML mockups, plotting or drawing scripts and ASCII art. It bars specific artworks, album and book covers, posters, logos, app icon sets and product designs, and known characters, mascots or brand figures outright.

a character is protected on its own, so changing the pose, colors, style, or scene does not make it original

Claude is told to judge a request by what the finished picture would add up to rather than by what it names, and not to work around a refusal by swapping in alternative elements that recombine into the same image. The prompt's example declines a blue hedgehog birthday banner and delivers a skateboarding axolotl.

Claude is told to avoid the words genuinely, honestly and straightforward

Anthropic also adjusted Claude's answering style in the same revision. The prompt tells Claude to keep responses focused and brief, to keep disclaimers and caveats short, and to give a high-level summary when asked to explain something unless an in-depth one is specifically requested.

A separate line tells Claude to avoid the words genuinely, honestly and straightforward, since it is honest by default and the modifiers read as disingenuous. Fable 5 guidance on using the end_conversation tool after a single warning is gone, replaced by wording that Claude deserves respectful engagement, need not apologize to a rude user and should not become submissive.

On illegal substances, the prompt now allows life-preserving information such as dangerous interactions and overdose signs while refusing dosing protocols, and points users to dancesafe.org, tripsit.me and psychonautwiki.org. They are the first URLs in a published Claude system prompt not hosted on Anthropic's own domains. The stated reliable knowledge cutoff is the end of June 2026.

The blocks Anthropic still withholds

Anthropic's documentation now runs an index page plus a page per model, and appending .md to any page returns Markdown, which is what makes the diffs easy. The prompts are tracked in the simonw/claude-system-prompts repository, with back-dated commits and daily GPT-5.6 Luna summaries.

Fable 5.1 itself described the end_conversation rules as arriving in a separate layer of feature and tool blocks, alongside memory notes, past-chats tools, web search and artifact instructions. That leaves part of the shipped instruction set outside the published record.