Claude's Fable 5.1 system prompt now refuses song lyrics and copyrighted characters in generated SVG

Anthropic publishes the system prompts used by the Claude web and mobile apps, and the current Fable 5.1 prompt adds two rules that were not in Fable 5. One covers song lyrics and book passages. The other covers pictures Claude draws with code. Simon Willison compared the two versions on 2 September and wrote up the differences.

The lyrics rule is direct. The published prompt says Claude "does not reproduce song lyrics, poems, or passages from books and articles, in whole or in part," and it names the ways people get around such a rule: asking for the last lines, asking for a chorus or hook, asking for the melody note by note, or pasting lines in one at a time and calling them their own work. Once Claude declines in a conversation, the prompt tells it to keep declining reworded versions for the rest of that conversation. Works first published before 1929 are allowed, and the prompt tells Claude to go by what it knows of the date rather than the user's claim about it.

The rule that reaches code

The second rule is the one that will surprise people who build with the API's sibling products. The prompt extends the same limit to "visual and designed works, including anything Claude draws with code," and lists SVG, canvas, CSS, HTML mockups, plotting scripts and ASCII art by name. It says Claude does not draw a known character, mascot or brand figure at all, and that changing the pose, colours, style or scene does not make the result original. It also tells Claude to judge the request by what the finished picture would add up to rather than by the words used, so describing a character without naming it counts as naming it.

The prompt carries a worked example of the refusal: asked for a birthday banner with "a blue hedgehog running really fast," Claude declines Sonic in one sentence and offers a skateboarding axolotl instead.

Three smaller changes sit alongside these. The prompt now tells Claude to avoid the words "genuinely," "honestly" and "straightforward," on the reasoning that those modifiers "come off as disingenuous." The old instruction to give one warning and then use the end_conversation tool has been replaced with guidance to stay on the problem and maintain self-respect. And the stated reliable knowledge cutoff is the end of June 2026.

These prompts apply to Claude.ai and the Claude mobile apps. Anthropic does not publish the prompts for Claude Code or Claude Cowork, so a team working through those tools cannot diff what changed underneath them the same way.

The practical point is that a refusal boundary moved without a model version changing under most people's feet. If a product asks Claude for icon sets, marketing banners, chart styling that mimics a known brand, or anything a reviewer would recognise, the same request that worked in Fable 5 can now come back declined. That is a runtime behaviour change with no API deprecation notice attached to it, which is exactly the kind of change an eval suite catches and a manual spot check does not. Anthropic keeping the historic prompts online is the useful part here: the diff is available to anyone who wants to know what moved.

Source

Anthropic: Claude Fable 5.1 system prompt, and Simon Willison's comparison of the Fable 5 and 5.1 prompts.