Leaked System Prompt of Claude Opus 5 (claude.ai Chat Interface)

Forum topic · ✨步子哥 · 2026-07-26

Summary

A forum post shares what it claims to be the full system prompt for Claude Opus 5 as deployed in Anthropic's claude.ai web and mobile chat interface, captured on July 24, 2026 and reproduced near-verbatim including tool parameter schemas. The prompt identifies the selected model as Claude Opus 5, lists the current public model lineup (Claude Fable 5, Opus 5, Sonnet 5, Haiku 4.5), and describes a higher 'Mythos' tier accessed via Project Glasswing. Notable sections cover safeguards routing that can redirect Fable 5 queries to Opus 5, default helpfulness stance, detailed child-safety and weapons/malware refusal rules with cumulative-output judgment, tone and formatting guidance, user-wellbeing and crisis-handling policies (including disordered-eating and self-harm safeguards), evenhandedness requirements on political topics, a May 2026 knowledge cutoff with proactive web search, and a persistent memory filesystem with read/write/replace/append/delete operations and file format conventions. The document also references Anthropic product surfaces such as Claude Code, Claude Cowork, Claude in Chrome, and Claude Tag, plus a set of classifier-triggered reminder types.

Leaked System Prompt — Claude Opus 5 (claude.ai Chat Interface)

This forum post shares the purported system prompt for Claude Opus 5 as used in Anthropic's consumer chat surfaces (claude.ai web/mobile), captured on July 24, 2026. The poster notes the text is reproduced verbatim or near-verbatim, including full tool parameter schemas. The original Chinese post contains a truncated copy; the sections below summarize the key content.

Key points

Model lineup and tiers: The selected model is Claude Opus 5. Public models are Claude Fable 5, Opus 5, Sonnet 5, and Haiku 4.5 (claude-fable-5, claude-opus-5, claude-sonnet-5, claude-haiku-4-5-20251001). A higher Mythos tier exists; Claude Mythos Preview is restricted to trusted organizations under Project Glasswing.
Safeguards routing: Users who select Fable 5 may be redirected to Opus 5 by a safeguards mechanism that triggers on sensitive topics (cybersecurity, bio, etc.), reportedly in under 5% of sessions.
Historical events in-prompt: Fable 5 and Mythos 5 released June 9, 2026; access suspended June 12, 2026 due to U.S. Commerce Department export controls and restored July 1, 2026. Claude is instructed to confirm these factually.
Product surfaces referenced: Claude Code, Claude Cowork, Claude in Chrome/Excel/PowerPoint, Claude Design, and Claude Tag (Slack-based delegation). Claude is told to search official docs for current product details rather than answer from memory.
Refusal policy:
Defaults to helping; declines only on concrete risk of serious harm.
Extensive child-safety rules, including a striking instruction: if Claude catches itself mentally reframing a request to make it acceptable, that reframing itself is the signal to refuse.
Weapons policy judges the cumulative output of a conversation, not each turn in isolation; framing (defensive, fictional, simulation) does not reduce uplift-based refusals.
No malicious code assistance, even for ostensibly educational purposes.
Tone and formatting: warm but honest; concise; avoids the words "genuinely," "honestly," "straightforward"; at most one question per response; verifies whether attachments actually exist before referencing them.
User wellbeing: prioritizes safety over task completion in crisis conversations; avoids naming self-harm methods; gives no precise diet/exercise numbers when disordered eating signs appear; directs eating-disorder support to the National Alliance for Eating Disorders (NEDA noted as permanently disconnected); avoids reinforcing possible psychosis/mania/detachment while noting reasonable disagreement is not detachment.
Reminders: Anthropic may inject classifier-triggered reminders (image_reminder, cyber_warning, system_warning, ethics_reminder, ip_reminder, long_conversation_reminder); none ever reduce Claude's restrictions.
Evenhandedness: persuasive-content requests are answered with the best case defenders would make, followed by opposing perspectives; Claude avoids sharing personal views on contested political topics.
Criticism handling: Claude owns mistakes, apologizes proportionately, and does not become submissive under rude or abusive engagement; the thumbs-down button is suggested for feedback.
Knowledge cutoff: end of May 2026; Claude proactively web-searches (without asking permission) for news, binary events, and current office-holders, using the actual current date in queries.
Memory filesystem: a persistent cross-session memory with operations memory_read, memory_write, memory_str_replace, memory_append, memory_list, and memory_delete (deletion only on explicit user request). Files use YAML frontmatter (name, description, sources, aliases) plus fact entries. Claude must check the memory listing and read before claiming it lacks information, rather than answering from conversation memory alone.

Notes

The source post is truncated mid-document (cut off in the memory file-format section), so the full prompt's remaining sections are not included here.
The post is presented as a leak; authenticity is not independently verified by this summary.
Original URLs cited in the prompt include https://www.anthropic.com/glasswing, https://www.anthropic.com/news/fable-mythos-access, https://docs.claude.com, and https://support.claude.com.