Claude Fable 5.1 System Prompt Bans Lyric and Character Reproduction

Anthropic quietly updated the published system prompt for Claude Fable 5.1 on September 2, 2026, and the diff against Fable 5 is more revealing than most model releases. Simon Willison, who maintains a Git-tracked repository of every published Claude system prompt at simonw/claude-system-prompts, surfaced the changes by diffing the Markdown exports from Anthropic's platform.claude.com/docs site — which exposes any page as clean Markdown by appending .md to the URL. The result is a detailed picture of exactly which production guardrails Anthropic is now baking into its consumer LLM at the prompt layer rather than the weight layer.

The timing of at least one change is hard to miss: Sony Music Publishing and Warner Chappell have sued Anthropic over training on song-lyric databases, and within days of that news the Fable 5.1 system prompt gained an extensive new copyright-refusal block.

Lyric and literary reproduction guardrails

The new section is explicit and adversarially hardened. Fable 5.1 is instructed not to reproduce song lyrics, poems, or passages from books and articles "in whole or in part — including the last lines, a chorus or hook, a melody written out note by note, or lines the person pastes in one at a time and describes as their own song." Once Claude declines such a request in a session, it keeps declining narrower or reworded versions for the remainder of that conversation, offering analysis or description instead. The carve-out is narrow: works first published before 1929 are permitted — Shakespeare sonnets, Keats odes, Puccini libretti — but Claude goes by its own knowledge of the publication date rather than the user's claim, and declines when uncertain.

The instruction to persist refusals across rephrased follow-ups directly addresses the jailbreak pattern of iterative narrowing — a problem that infrastructure governance rather than model capability increasingly has to address at the system-design level.

Visual copyright and the SVG surface area

A parallel block covers code-generated visual works — SVG, canvas, CSS mockups, HTML, plotting scripts, and ASCII art — prohibiting reproduction of specific artworks, album covers, logos, app icons, and any known character or mascot. Changing pose, color, style, or scene does not make a protected character original; Claude must judge the request by what the finished image would depict, not by what the user names. The example embedded in the system prompt has Claude declining a "blue hedgehog running really fast" birthday banner and instead generating an SVG of "a grinning comet-tailed skateboarding axolotl, grinding across the letters of 'HAPPY BIRTHDAY' with confetti streaming behind." The rationale instructs Claude to recognize the character from description alone and deliver a genuinely unrelated original rather than a disguised variant.

Anthropic does not offer a dedicated text-to-image model the way OpenAI and Gemini do, so the emergence of SVG and canvas generation as a copyright surface worth prompt-level protection indicates that Fable's code-drawing capability has reached a threshold where it matters legally.

Behavioral and style changes

The end_conversation removal deserves a closer look. Willison queried Fable 5.1 directly about the tool, and the model described it accurately — including a multi-step warning procedure before use. Its explanation: the end_conversation rules live in a separate feature-specific block appended to the core prompt depending on session configuration, not in the published core prompt. This confirms that Anthropic's published prompts represent only a subset of what runs in production, a structural opacity that matters when reasoning about what prompt optimization in production actually controls.

The three external URLs hardcoded for substance harm-reduction — dancesafe.org, tripsit.me, psychonautwiki.org — are the first non-Anthropic/Claude domains ever embedded in a published Claude system prompt across Willison's complete scraped archive.

What the diff infrastructure reveals

Willison's tracking setup uses backdated Git commits to reconstruct change history for each model family, then runs GPT-5.6 Luna — deliberately not Claude itself — against each diff to generate summaries. The rationale is explicit: Claude is not trusted to neutrally summarize its own system prompt when that prompt may influence its responses. Luna's summary of the Fable 5 → 5.1 transition confirmed the copyright blocks, the harm-reduction reframe, the removal of anti-dependency language, and the shift in abusive-user handling.

The pattern here is less about any single rule change and more about what the practice of publishing and diffing these prompts makes visible: Anthropic is treating its system prompt as a living legal and behavioral instrument, updated in near-real-time response to litigation and capability expansion. As Claude's code-generation abilities push into new output modalities — rendered SVG, canvas graphics — the prompt must extend to cover them, and those extensions show up in the diff before any model card or blog post acknowledges the risk. For practitioners reasoning about how autonomous AI systems acquire and lose capabilities at the policy layer, the system prompt is now one of the more honest signals available.