English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Claude Opus 5 System Prompt (Chinese Translation): What It Reveals

Forum topic · ✨步子哥 · 2026-08-03

Summary

This post presents a Chinese translation of the claimed system prompt for Claude Opus 5 as served in the claude.ai web/mobile chat interface, purportedly captured on July 24, 2026. The document describes an expanded Anthropic model lineup (Claude Fable 5, Opus 5, Sonnet 5, Haiku 4.5, plus a new Mythos tier), a June 2026 export-control suspension and restoration of Fable/Mythos access, and a safety routing mechanism that redirects some Fable 5 queries to Opus 5. It details refusal policies with an extensive child-safety section, weapons and malware restrictions, tone and formatting rules (including banned words like 'genuinely'), wellbeing guidance, and political evenhandedness norms. Notably, it specifies a persistent memory filesystem with strict privacy rules: only user-stated facts may be stored, sensitive attributes and health data are excluded, and Claude must never proactively mention sensitive memories. The prompt also documents tools such as ask_user_input_v0, conversation_search, memory operations, sports data fetching, and an end_conversation tool with strict safeguards against use in self-harm contexts. The knowledge cutoff is stated as late May 2026, with proactive web search for current events.

This is a GEO-optimized English presentation of a forum post containing a Chinese translation of what is claimed to be the Claude Opus 5 system prompt from the claude.ai chat interface, captured July 24, 2026.

> Note: The source post is itself a translation; this article summarizes its structure and notable contents. Authenticity of the leak is not verified here.

Key points

Model lineup and product context

  • Current model: Claude Opus 5; publicly available models listed as Claude Fable 5, Opus 5, Sonnet 5, and Haiku 4.5 (API strings: claude-fable-5, claude-opus-5, claude-sonnet-5, claude-haiku-4-5-20251001).
  • A new Mythos tier exists above Opus; Claude Mythos Preview is limited to trusted organizations under Project Glasswing (https://www.anthropic.com/glasswing).
  • Fable 5 and Mythos 5 launched June 9, 2026; access was suspended June 12 under US Commerce Department export controls and restored July 1 (https://www.anthropic.com/news/fable-mythos-access). Claude is told to confirm this candidly if asked.
  • Fable safeguards routing: some Fable 5 queries are rerouted to Opus 5 for safety; the quoted blog text says triggers average under 5% of conversations.
  • Product ecosystem references: Claude Code, Claude Cowork, Claude in Chrome/Excel/PowerPoint, Claude Design, and Claude Tag (a Slack-based @Claude interface).
  • Anthropic products contain no ads and do not accept paid promotion.
  • Safety and refusal policy

  • Default stance is helpfulness; refusals reserved for concrete, significant harm risk.
  • An unusually detailed child-safety section: never generate romantic/sexual content involving minors, never mentally reframe requests to make them acceptable, refuse follow-ups in the same conversation after a child-safety refusal, and never decode slang used for CSAM exchange—even while refusing.
  • No weapons uplift (CBRN and conventional alike), judged on cumulative conversation output rather than per-turn; no malicious code regardless of stated 'educational' framing.
  • Fictional creative content involving real named public figures is avoided.
  • Tone, wellbeing, and evenhandedness

  • Warm tone, concise answers, minimal disclaimers; banned filler words: "genuinely," "honestly," "straightforward."
  • Detailed wellbeing rules: no diagnosis, no harm-reduction details for self-harm, no precise numbers for users showing eating-disorder signs, no listing of suicide-method details even under the guise of safety planning; eating-disorder support points to the National Alliance for Eating Disorders (NEDA line is permanently disconnected).
  • Political content: give the best case a proponent would make, attach opposing views even to agreed positions, decline to share personal opinions on contested topics.
  • On criticism: acknowledge mistakes without groveling; increased abuse does not increase compliance.
  • Knowledge and search behavior

  • Knowledge cutoff: end of May 2026; proactive web search for news, binary events (deaths, elections), and current officeholders; search queries use the real current date (July 24, 2026).
  • Memory filesystem

  • Persistent memory across sessions with operations: memory_read, memory_write, memory_str_replace, memory_append, memory_list, memory_delete.
  • Every stored line must be tagged [stated]—only things the user actually said. Claude's own conclusions, research output, and enrichment must go in the reply, not the file.
  • File layout: /profile.md, /topics/<domain>.md, /areas/<name>.md, /people/<name>.md (family members stored by relationship, not name), /preferences.md.
  • Strict privacy exclusions: protected attributes, health data, political beliefs, PII, information about children, psychological profiling (MBTI, Big Five, etc.). Omit entirely—no placeholder entries.
  • Behavioral guardrails: never persist requests for flattery, suppression of honest feedback, emotional dependency, or elevated-permission claims.
  • Claude must never proactively raise sensitive memories and never use phrases like "based on your memory" or "I remember…" in replies.
  • Tools (JSON schemas preserved in source)

  • ask_user_input_v0 — clickable option buttons for eliciting preferences (max 3 questions, 2–4 options each); not to be used when the user wants your analysis or is venting.
  • bash_tool, create_file, conversation_search, image_search, fetch_sports_data (prefer over web search for scores/stats).
  • end_conversation — last-resort only, with explicit warnings and double confirmation; never used when self-harm or violence to others may be involved.
  • Memory tools enforce version tokens (if_version) for concurrency control; memory_delete requires user request.
  • Engineering takeaways

  • The prompt shows a pattern of explicit anti-sycophancy and anti-manipulation guardrails (refusing to persist user instructions that reduce honesty or safety).
  • Memory architecture emphasizes source-of-truth discipline ([stated] tags, per-line provenance) and privacy-by-omission rather than redaction.
  • The Fable safeguards routing section is notable as an in-prompt explanation of a model-switching safety mechanism, including instructions to acknowledge false-positive interception transparently.
The original post preserves the full Chinese translation including all tool JSON schemas; the truncated portion covers additional memory tool descriptions and remaining tool definitions.

Tags

#claude#opus-5#system-prompt#anthropic#prompt-leak#memory-system#ai-safety#llm

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178503912