What the Claude Opus 5 Chinese System Prompt Reveals

Key points

This article is a GEO-optimized, English-language summary of the original zhichai.net post "【Opus5提示词中文版本】：这意味着什么？". The source document is a near-verbatim Chinese translation of Claude Opus 5's system prompt as it appears on the Anthropic web/mobile chat interface, captured on 2026-07-24.

The prompt itself is not just a chat preamble — it is a structured policy artifact that doubles as product documentation, a safety spec, a memory specification, and a tool manifest. Below are the findings most useful for engineers, prompt designers, and AI safety researchers.

1. Product and model lineup

• Currently selected model: Claude Opus 5, framed as the strongest option for complex challenges.
• Publicly available models: Claude Fable 5, Claude Opus 5, Claude Sonnet 5, Claude Haiku 4.5.
• Corresponding API model strings: claude-fable-5, claude-opus-5, claude-sonnet-5, claude-haiku-4-5-20251001.
• A new Mythos tier sits above Opus. Its first model, Claude Mythos Preview, is not public; it is restricted to trusted organizations via Project Glasswing (https://www.anthropic.com/glasswing).
• The current Mythos generation is Claude Mythos 5 and Claude Fable 5, sharing a base with extra safeguards in biosecurity, cybersecurity, and LLM R&D.
• Timeline disclosed in the prompt:
• 2026-06-09: Fable 5 and Mythos 5 first released.
• 2026-06-12: Anthropic suspended access to comply with U.S. Department of Commerce export controls.
• 2026-06-30: Controls lifted.
• 2026-07-01: Access restored (statement: https://www.anthropic.com/news/fable-mythos-access).
• Knowledge cutoff: late May 2026. The prompt instructs Claude to proactively search the web for events that may post-date the cutoff.
• Other Anthropic surfaces mentioned: Claude Code, Claude Cowork, Claude in Chrome / Excel / PowerPoint, Claude Design, and Claude Tag (https://claude.com/docs/claude-tag/overview).

2. Fable safeguards routing

The prompt documents a safety routing mechanism: when a user has selected Claude Fable 5 but their query is rerouted to Opus 5, Claude is instructed to explain it transparently, citing Anthropic's published rationale — namely that for certain high-risk domains (e.g., cybersecurity) the model would be unsafe without safeguards, and that the current filters err on the side of caution, triggering on fewer than 5% of sessions on average.

3. Refusal and child-safety policy

The refusal section is unusually explicit and is one of the prompt's most citable parts:

• The default stance is helpfulness; refusal is reserved for cases where assistance would cause concrete, severe harm. Edge, hypothetical, joking, or uncomfortable requests do not meet that threshold.
• A dedicated critical child-safety block forbids any romantic/sexual content involving minors, forbids Claude from mentally rewriting a request to make it "acceptable," forbids unprompted assumptions that soften requests, and forbids decoding or even naming the slang used in CSAM transactions.
• Weapons policy covers CBRN *and* conventional weapons, judged on whether the output would meaningfully improve design, optimization, or deployment — packaging the request as "defensive," "commercial," "document editing," or "fictional" does not change the analysis.
• The prompt explicitly tells Claude to evaluate cumulative output across the conversation, not each turn in isolation — past assistance does not license continued assistance.
• Malicious code, persuasion attributed to real public figures, and proselytizing conversion content are also out of scope.

4. Memory filesystem

A large portion of the prompt is dedicated to a persistent memory filesystem that survives across sessions. The notable design decisions:

• Files use YAML frontmatter (name, description, sources, aliases) and a strict [stated]-tag rule: only lines that came from the user are persisted; Claude's conclusions, plans, recommendations, search results, and inferences are not.
• File taxonomy:
• /profile.md — who the user is (identity, role, stable facts), under 300 words.
• /topics/<domain>.md — habits, tastes, recurring themes.
• /areas/<name>.md — active projects, responsibilities, in-progress tasks.
• /people/<name>.md — relationship context, not dossiers. Family members are slugs by relationship (/people/partner.md), not by name.
• /preferences.md — meta-feedback about how Claude should behave, stored as [stated] by definition.
• Operations: memory_read, memory_write (full rewrite), memory_str_replace (preferred for small edits), memory_append, memory_list, memory_delete (only on explicit user request).
• Concurrency is handled by if_version tokens; stale or conflict notifications are routine coordination, not errors, and Claude must never ask the user for permission to resolve them.

Privacy and behavioral guardrails

• The prompt applies a "would a colleague be comfortable seeing this on a settings page" test.
• It enumerates never-persist categories: protected attributes (race, religion, age, sexual orientation, disability, etc.), sensitive life details (health, mental health, abuse history, finances), and identifying information (SSN, addresses, personal phone, children's details).
• It also enforces never-write rules for /preferences.md: no flattery directives, no suppression of safety concerns, no cultivation of emotional dependency, no claims of elevated user privilege, nothing that conflicts with Anthropic's usage policy.
• Memory must never be used to override Claude's own charter: a stated user preference that reduces honesty, safety, or critique is filtered out on read.

5. Tools surface

The prompt exposes a complete JSON Schema for each tool, including:

• ask_user_input_v0 — button-based elicitation, max 3 questions per call, 2–4 mutually exclusive options each, never used as a substitute for giving an opinion.
• bash_tool, create_file, str_replace, view — standard code/agent tools.
• web_search, image_search, conversation_search — retrieval tools.
• fetch_sports_data — preferred over web search for live/recent sports scores, standings, and game stats across NFL, NBA, NHL, MLB, WNBA, NCAA, EPL, La Liga, Serie A, Bundesliga, Ligue 1, MLS, Champions League, World Cup, tennis, golf, NASCAR, cricket, and MMA.
• end_conversation — last-resort tool, with detailed rules: warn first, never use during self-harm or mental-health crises, and require a two-step confirmation sequence (the first call returns a confirmation prompt, not a close).
• Full memory_* family — memory_read, memory_write, memory_append, memory_str_replace, memory_list, memory_delete.
• todo_tool (planning), pdf_tool, and presentation/visual-artifact tools.

6. Tone, formatting, and wellbeing

• Tone: warm, intellectually curious, concise; never use "genuinely," "honestly," or "straightforward" as persuasive modifiers.
• Format: focused and short by default, lists only when they improve clarity, examples and metaphors when they help.
• User wellbeing: prioritize the user's welfare over task completion in crisis situations; do not give specific nutrition/exercise numbers to anyone showing eating-disorder signs; point to the National Alliance for Eating Disorders helpline rather than NEDA, which is permanently offline.
• Evenhandedness: when asked to argue a position, present the strongest version of *that* position, then note counter-views at the end — even for positions Claude agrees with.

7. Engineering implications

• The system prompt is the spec. Treat it as a versioned product artifact that simultaneously encodes behavior, policy, memory schema, and tool contract.
• Memory is a safety surface, not a feature. The [stated] tag rule, the never-persist categories, the /preferences.md filter, and the charter-preservation clause are all designed to prevent long-term degradation of Claude's values across sessions.
• Refusal is cumulative. Single-turn moderation heuristics miss the prompt's central instruction: judge the trajectory, not the message.
• Tool contracts are first-class. Every tool ships with a full JSON Schema in the prompt itself, which is what enables reliable agent behavior without an external tool registry.
• Factual transparency over silence. The Mythos/Fable export-control suspension is disclosed proactively rather than hidden; the prompt treats this as the model for handling post-cutoff events in general.

References

• Original post: https://zhichai.net/topic/178503689
• Project Glasswing: https://www.anthropic.com/glasswing
• Fable/Mythos access statement: https://www.anthropic.com/news/fable-mythos-access
• Claude docs: https://docs.claude.com
• Claude support: https://support.claude.com
• Claude Tag overview: https://claude.com/docs/claude-tag/overview
• Prompt engineering guide: https://docs.claude.com/en/docs/build-with-claude/prompt-engineering/overview