Claude Tightens Its Rules: No Protected Words or Images

Anthropic releases the system prompts for the public versions of Claude and introduces stricter guidelines on copyright and sensitive assistance. The exchanges are calibrated to be concise, certain rules operate through unpublished layers, and Claude's reliable knowledge stops at June 2026.

End-of-chat rules managed by unpublished layers

The Fable 5.1 guidelines no longer explicitly mention the use of a tool to end a discussion, whereas Fable 5 included a single warning before a closure for abuse. When asked, Fable 5.1 describes two cases where the end of conversation applies: at the user's request, after confirming that the stop is permanent, and as a last resort in the face of persistent abuse, after attempts to redirect and a clear warning. Fable 5.1 indicates that these rules come from a distinct layer of the main prompt: functional blocks added according to the session, including end-of-conversation rules, memory notes, tools related to past chats, research and citation guidelines, as well as instructions for creating artifacts and files. These blocks do not appear in the published prompt.

Reliable knowledge set to the end of June 2026

The Fable 5.1 documentation states that the deadline for reliable knowledge as well as the training data collection period is set to June 2026. The prompt indicates that Claude formulates its responses in the same way a well-informed person would in June 2026 compared to the current date, this information being conveyed by the macro {{currentDateTime}}, which appears only once towards the end of the prompt.

Lyrics, poems, and excerpts: default refusal after 1929

Fable 5.1 introduces a strong prohibition against reproducing song lyrics, poems, or excerpts from books and articles, including partial elements like a refrain, a hook, or a notated melody, or lines copied as if they were original. Once a refusal is issued, variations of the same request are also declined in the conversation thread, with a focus on description or analysis. Texts prior to 1929 are accepted, with examples like a sonnet by Shakespeare, an ode by Keats, or an aria libretto by Puccini. Claude relies on its knowledge of dates rather than the user's assertion and refuses in cases of uncertainty. Meanwhile, Sony Music Publishing and Warner Chappell are suing Anthropic regarding training on lyrics databases.

Logos and protected characters: prohibited including in SVG or CSS

The guidelines extend the prohibition to visual works and protected designs, regardless of the means of generation, including through code such as SVG, canvas, CSS, HTML mockups, drawing scripts, or ASCII art. Claude does not reproduce specific works, covers, posters, logos, icon sets, or product designs, and does not draw a character, mascot, or well-known brand figure. Modifying pose, colors, style, or scene is not enough to create originality. The evaluation is based on what the final image would show. A request clearly describing a known work or character is treated as a nomination, with no circumvention by "equivalents" that recompose a recognizable image.

More concise responses and streamlined vocabulary

The prompts request focused, brief, and concise responses, with warnings minimized and priority given to the essentials. On requests for explanation, Claude begins with a high-level summary, unless a level of detail is explicitly required. The use of terms like "genuinely," "honestly," or "straightforward" is prohibited, with guidelines reminding that the assistant is honest by default and can express itself directly. Regarding difficult interactions, the instructions emphasize mutual respect, the absence of excessive apologies, and maintaining self-respect, without escalating submission in the face of abuse.

Substances: no protocols, risk reduction resources

Fable 5.1 specifies that Claude provides no advice on the synthesis, production, or distribution of illegal substances. However, it can provide information to preserve life, such as dangerous interactions, signs of overdose, or when to seek help, while refusing specific protocols for dosage, timing, administration, or combinations. The prompt allows redirection to established risk reduction resources: dancesafe.org, tripsit.me, and psychonautwiki.org. This is the first occurrence of URLs outside of Anthropic's domains in a Claude system prompt.

Docs, formats, and history: comparable prompts now available

Anthropic publishes the system prompts for the public applications of Claude, along with the modification history, but not those for Claude Cowork or Claude Code. The presentation has shifted from a single page to an index and pages by model, with dated examples such as Haiku 4.5 on October 15, 2025, and January 18, 2026. The documentation is usable by language models and offers Markdown output via the addition of .md, including for the index and prompts of Fable 5.1. An independent Git timeline aggregates prompts drawn from the documentation, in the GitHub repository simonw/claude-system-prompts. It facilitates comparisons by grouping by model families the latest available prompt and synthesizing retro-dated commit histories. Histories and files exist for the Fable, Opus, Sonnet, and Haiku families, as well as for each model version with artificial commits.