Breaking Down Claude's System Prompt Update: The Tug-of-War Between AI Copyright Protection and Safety Boundaries

Anthropic has long been known for publicly sharing the system prompts of Claude's consumer-facing applications (Claude.ai and its mobile apps) — not just the current version, but also historical changelogs. Developer Simon Willison compared the system prompts of Fable 5 and Fable 5.1, uncovering a series of intriguing adjustments — from lyrics copyright to abstract brand characters, from response style to drug-related information handling. These subtle changes reflect the ongoing tug-of-war AI companies face between legal compliance, ethics, and product experience.

Refusing to Reproduce Lyrics: An Immediate Response to Legal Pressure

The most striking change in this update is the addition of a lengthy new section on "not reproducing lyrics." The system prompt explicitly instructs Claude not to reproduce song lyrics, poems, or book passages in full or in part — including final verses, choruses, note-by-note melodies, and even content users paste in line by line while claiming it as their own work.

More interesting is its "persistent refusal" mechanism: once Claude declines such a request in a conversation, it will continue refusing any narrower or rephrased version of the request throughout the entire session, offering descriptions or analysis instead. Lyrics and poems first published before 1929 (such as Shakespeare's sonnets, Keats' odes, or the Italian libretto of Puccini arias) are permitted, but Claude relies on its own judgment about the work's date rather than the user's claims, defaulting to refusal when uncertain.

The timing of this change is no coincidence. Just days earlier, Sony Music Publishing and Warner Chappell sued Anthropic, alleging the company used lyrics databases to train Claude. The system prompt was updated within days of the lawsuit becoming public — a textbook case of how AI companies use product-level rules to hedge against legal risk.

No Drawing Copyrighted Characters: New Compliance Challenges from Improved Image Generation

The next section prohibits generating copyrighted visual content, with this restriction specifically covering works "drawn with code" — SVG, canvas, CSS, HTML prototypes, drawing scripts, and even ASCII art.

The prompt emphasizes that characters themselves are protected, and changing their pose, color, style, or scene does not make the output "original." Claude judges requests based on what the finished image "ends up looking like," not what the user calls it. If the described elements clearly point to a well-known work or character, it's treated as referencing that work, and Claude won't bypass an already-declined request by swapping in "alternative elements."

The system prompt includes a vivid example: a user wants to make a birthday banner featuring a "super-fast blue hedgehog" for their son. Claude recognizes this as Sonic, declines, and instead offers an original design — a "grinning comet-tailed skateboarding axolotl." Simon tested this and confirmed Claude does indeed respond this way.

A notable detail: Anthropic hadn't previously been too concerned about this issue — unlike OpenAI and Gemini, it doesn't have a dedicated image generation model. The appearance of this new rule is itself evidence that Fable has become powerful enough at SVG generation for copyright issues to start surfacing.

Response Style Adjustments: More Concise, Less Filler

Anthropic is also continuously fine-tuning Claude's response style. The new prompt requires Claude to keep answers focused, short, and refined. Disclaimers and qualifiers should be brief, with most of the response dedicated to the core answer. When asked to explain something, Claude should offer a high-level overview first unless explicitly asked to go deeper.

More specifically, the prompt explicitly tells Claude to avoid words like "genuinely," "honestly," and "straightforward." The reasoning: Claude is honest by default and can state its views directly without needing these modifiers to "convince" users — which actually comes across as less sincere. This directly addresses a long-standing user complaint about Claude's language style.

From "End the Conversation" to "Maintain Self-Respect": A Subtle Shift in Safety Strategy

The approach to handling abusive conversations has shifted subtly. The old Fable 5 prompt allowed Claude to use an end_conversation tool to terminate a conversation when subjected to abuse, though it required giving one warning first.

Fable 5.1 no longer encourages ending conversations. Instead, it asks Claude to "take responsibility without being self-deprecating, overly apologetic, self-critical, or submissive." Rather than becoming increasingly compliant in the face of abuse, the goal is "steady, honest helpfulness" — acknowledging what went wrong, staying focused on the issue, and maintaining self-respect.

Interestingly, when Simon asked about the end_conversation tool that had been removed from the prompt, Claude was still able to describe in detail how it works. Claude explained that this information comes from "a different layer" — after the core prompt, a series of tool-specific modules are appended based on features enabled for the session (such as the memory system, conversation history tools, web search, artifact creation, etc.), and these modules are not publicly disclosed. This is a reminder that there are still key parts of Anthropic's system prompts that remain hidden.

Harm Reduction Guidance and the First-Ever External Links

Regarding illegal substances, the new prompt makes clear that Claude won't provide guidance on synthesis, production, or distribution, but can and should provide "life-saving" information — such as dangerous drug interactions, signs of overdose, or when to seek help. It refuses to give specific plans for dosage, timing, administration, or combinations, instead directing users to established harm reduction resources, including dancesafe.org, tripsit.me, and psychonautwiki.org.

According to Simon's analysis, this is the first time in the history of Claude's system prompts that external links outside of claude.com, anthropic.com, or claude.ai have appeared — he specifically ran a script to verify this across all historical records. This detail signals that Anthropic is moving from a simple "refuse" approach to a more mature strategy of "redirect to professional harm reduction resources" when handling potentially dangerous information.

Additionally, Fable 5.1's reliable knowledge cutoff date is set to June 2026, and the prompt uses a {{currentDateTime}} macro — it appears only once, near the end of the prompt, which makes sense from a caching efficiency perspective.

Using AI to Track AI: An Automated System Prompt Monitoring Solution

Simon also demonstrated how he tracks these changes. He stores the collected prompts in the simonw/claude-system-prompts repository on GitHub, building files with "traceable commit histories" for each model family so that any two versions can be compared via diff directly in GitHub's interface.

More cleverly, since reading diffs is tedious and LLMs excel at it, he uses GPT-5.6 Luna to auto-generate summaries of key changes for each update. Why not use Claude to summarize its own system prompts? Simon is candid: he doesn't trust Claude to summarize its own system prompt — because the prompt's content could influence its judgment, creating a conflict of interest. The entire tracking system runs daily via GitHub Actions, and the automation code was almost entirely written by Claude Fable 5.1 itself.

Conclusion

This seemingly minor system prompt update is actually a mirror reflecting the complex realities of running AI products today: lawsuits forcing copyright rules into place, improved image capabilities creating new compliance challenges, user feedback driving language style adjustments, and safety strategies being recalibrated between "protection" and "self-respect." For practitioners focused on AI governance and prompt engineering, continuously tracking these public system prompts may be the most direct window into understanding what AI companies are really thinking.