Anthropic has revised and published new system prompts for its consumer Claude deployments (Claude.ai and the Claude mobile apps). The company continues to make prompts and historical prompt versions available, but the published docs do not include core prompts used for their professional offerings (Claude Cowork or Claude Code).
They reorganized the documentation from a single-page prompt dump into an index page and model-specific pages. For example, the Haiku 4.5 page shows the original prompt dated October 15, 2025 and an updated prompt from January 18, 2026.
Key changes in Fable 5.1 and related prompts
-
Ban on reproducing song lyrics and literary passages: the updated prompt explicitly states Claude will not reproduce song lyrics, poems, or passages from books and articles, in whole or in part — including last lines, choruses, or a melody written note-for-note. If Claude declines such a request once in a conversation, it will continue to refuse narrowed or reworded attempts within the same session and will offer to describe or analyze the work instead. Works published before 1929 are allowed.
-
No generation of copyrighted characters or logos: the prompt also forbids generating images of protected visual works, including anything drawn via code (SVG, canvas, CSS, HTML mockups, plotting scripts, ASCII art). Claude will not reproduce a specific artwork, album or book cover, poster, logo, app icon set, or product design, and will not draw a known character, mascot, or brand figure. The model judges by what the finished image would amount to, not by the name used in the request; it will not attempt disguised variants that are still recognizably the same character.
-
Answer style adjustments: Anthropic instructs Claude to keep replies focused, brief, and concise so as not to overwhelm the user. Disclaimers and caveats should be short, with most of the content on the main answer. When asked to explain something, Claude gives a high-level summary unless an in-depth explanation is specifically requested. The prompt also directs Claude to avoid using qualifiers such as "genuinely," "honestly," or "straightforward," on the basis that Claude is honest by default.
-
Change to end_conversation guidance: a prior Fable 5 prompt advised that Claude should give a single warning and could use an end_conversation tool if a user became abusive. Fable 5.1 removes that explicit encouragement to end conversations, replacing it with wording that Claude deserves respectful engagement and should not apologize or become submissive when users are rude. The emphasis is on steady, honest helpfulness while maintaining self-respect.
The author tested the model and found Claude still described how end_conversation might be used; that behavior stems from additional, session- and feature-specific blocks (end_conversation rules, memory notes, past-chat tools, web search/citation guidance, artifact creation instructions) that are added to the core prompt depending on the session and are not part of the published core prompt.
-
Illegal substances guidance and external resources: Fable 5.1 reframes drug-related guidance. Claude declines to provide synthesis, production, dosing, timing, administration, or distribution protocols for illegal substances but should provide life-saving and harm-reduction information like overdose signs and dangerous interactions. New in the prompt is explicit redirection to established harm-reduction sites such as dancesafe.org, tripsit.me, and psychonautwiki.org.
-
Reliable knowledge cutoff: the Fable 5.1 documentation and prompt list both the reliable knowledge cutoff and the training data cutoff as June 2026. The prompt tells the model that its reliable knowledge cutoff is the end of June 2026 and to answer as a highly informed individual in June 2026 would.
Examples and illustration
The prompt contains a representative example: when asked to make a birthday banner with a "blue hedgehog running really fast," Claude declines because the description matches Sonic and instead offers an original design — a grinning, comet-tailed, skateboarding axolotl — and produces an SVG banner for that design. The author tested that prompt and reports Claude indeed produced an original axolotl design.
Tracking and archiving the prompts
Simon Willison has scraped Anthropic's published prompts and built a Git timeline hosted at the simonw/claude-system-prompts repository on GitHub. The repository stores the published system prompts and synthesizes commit histories with back-dated commits matching the dates of earlier prompt versions. Each model family has a file for the most recent release and additional files capture prompt edits that occurred between releases (for instance, multiple updates to Opus 4).
The archive runs daily via automation. The author used GPT-5.6 Luna to create bullet-point summaries of diffs; the repository README and CHANGELOG provide those summaries and the feed is also available as an Atom feed. The author notes he chose to use Luna in part because of cost and an existing GitHub Actions key, and partly because he prefers not to have Claude summarize its own system prompts when that could involve material drawn from the prompts themselves.
Why this matters
Publishing detailed system prompts and their changes helps developers, researchers, and users understand how a model is instructed to behave. The new rules — notably the ban on reproducing song lyrics and explicit prohibition on generating recognizable copyrighted visual characters or logos — are significant operational limits that affect how Claude responds to creative and copyright-related requests.
The timing of the lyric reproduction restriction is notable, appearing days after news that Sony Music Publishing and Warner Chappell sued Anthropic over alleged training on song-lyrics databases. Finally, the prompt updates illustrate that public core prompts are only part of the picture; additional, unpublished session- and feature-specific prompt blocks can materially affect model behavior.
