System prompts are a model's constitution—most users never read them, yet they govern every interaction. On September 2, Simon Willison published a line-by-line diff of the new Claude Fable 5.1 system prompt against a version he had archived the month before. The diff revealed a new wall: lyrics.
Scope: Willison compared Anthropic's officially published consumer prompt, covering the claude.ai web interface and mobile apps. The Claude Code and Cowork prompts are not part of this public release, though Anthropic's incident notices suggest the platforms share the same underlying behavior.
The New Clause
The rule is a single sentence but broad: Claude must not reproduce lyrics, poetry, or passages from books and articles—not in full, and not in fragments. The enforcement details are stricter still: once you are refused, the model remembers what you asked within the same session. Rephrasing the request, asking for just two lines, or tweaking words will not help—it continues to refuse and offers descriptions and analysis instead.
There is an exemption: lyrics and poems first published before 1929 are allowed. But the year is judged by the model's own knowledge—claiming a song is old does not count; the model has to know it.
The Timing Is Hard to Call a Coincidence
Willison added context: on August 31, the Guardian reported that Sony Music Publishing and Warner Chappell were suing Anthropic over a lyrics database used in training. Four days later, the lyrics clause appeared in the prompt. He cautiously noted it did not seem like a coincidence. Lawsuits take years; prompt changes take effect the same day. Vendors drew a behavioral line outside the courtroom, handling the two tracks separately.
The Blue Hedgehog Test
The copyright guardrails go beyond lyrics. The new clause also covers SVG, canvas, CSS, and ASCII art: Claude will not draw well-known characters, mascots, or brand imagery—changing pose, color palette, or art style does not make it original. Willison tested this by asking for "a fast-running blue hedgehog." The model recognized Sonic the Hedgehog (a Sony character), declined, and instead created an original blue axolotl riding a skateboard, complete with a birthday. The barrier closed, but a channel remained open.
The 1929 Date Is Stricter Than the Law
Under US copyright law, protection lasts 95 years from publication. Works published in 1930 entered the public domain at the start of 2026—so the 2026 public-domain line should be drawn at 1930. The prompt stops at "before 1929," meaning works from 1929 and 1930 remain blocked despite already being public domain. That is two years stricter than the law.
This matters for developers: Claude Code is subject to these rules too. Ask an agent to quote a lyric in a code comment, draw a mascot in a README, or copy two lines of poetry into a test case, and you will hit the wall. The model's capabilities have not changed—its constitution has.
Other Notable Changes
- Three adverbs are now banned: "sincerely," "honestly," and "frankly" (in the Chinese-context translation: 真诚地, 老实说, 直截了当地).
- The conversation-ending tool's wording changed from "warn once, then kick" to something like "be assertive, but no excessive apologizing, self-deprecation, self-criticism, or capitulation."
- Drug-related clauses now reference URLs outside Anthropic's domain for the first time, pointing to harm-reduction organizations.
- An ironic footnote: Willison's prompt-tracking repository generates daily summaries via GitHub Actions using GPT-5.6—he does not let Claude summarize its own prompts, citing distrust of its self-assessment.
Closing Thought
A popular comment from the 93-comment Hacker News thread captured the debate: these walls cannot stop models without walls—the future involves using more than one model. However elegant the barrier, it only binds those who respect it. Whether that observation is correct may be a more interesting question than who owns the lyrics.