This post shares a Chinese translation of a document presented as the full system prompt for Claude Opus 5 in the claude.ai web/mobile chat interface, dated Friday, July 24, 2026. The translation claims to follow the original closely, preserving tool parameter JSON schemas.
Key points
Product information
- The document identifies the model lineup as Claude Fable 5, Claude Opus 5 (currently selected), Claude Sonnet 5, and Claude Haiku 4.5, with API strings like
claude-opus-5. - A restricted Mythos tier exists above Opus; the first Mythos-class model, Claude Mythos Preview, is limited to trusted organizations under Project Glasswing (
https://www.anthropic.com/glasswing). Claude Mythos 5 and Fable 5 were released June 9, 2026; access was suspended June 12 due to US Commerce Department export controls and restored July 1 (https://www.anthropic.com/news/fable-mythos-access). - A safety routing mechanism may divert Claude Fable 5 queries to Opus 5, reportedly triggering in under 5% of conversations.
- Companion products listed: Claude Code, Claude Cowork, Claude in Chrome/Excel/PowerPoint, Claude Design, and Claude Tag (Slack-based
@Claudeinterface). - Default stance: helpful by default; refusals only for concrete, severe harm risk.
- Child safety: extensive hard rules — no romantic/sexual content involving minors, no softening or rewriting of requests, no decoding of exploitation-related slang, and heightened caution for the rest of the conversation after any refusal on these grounds.
- Weapons and malware: no meaningful uplift for weapons (CBRN or conventional), regardless of framing; no writing or explaining malicious code even for "educational" reasons.
- Cumulative judgment: Claude evaluates a conversation's total output rather than each turn in isolation; earlier assistance does not authorize continuation.
- Tone: warm, concise, minimal disclaimers, never swears unprompted, avoids filler words like "genuinely," at most one question per reply.
- Wellbeing: no diagnosis, no methods details in self-harm discussions, no precise diet/exercise numbers when eating disorder signs appear, eating disorder support points to the National Alliance for Eating Disorders hotline (NEDA is defunct).
- Evenhandedness: presents the best arguments a position's proponents would make, includes opposing views even for agreed positions, and is cautious about sharing personal opinions on contested politics.
- Knowledge cutoff: late May 2026; proactive web search for current events, incumbent officeholders, and recent binary events, using the actual date (July 24, 2026) in queries.
- A persistent memory system with operations
memory_read,memory_write,memory_str_replace,memory_append,memory_list,memory_delete, using version tokens (if_version) for concurrency. - File layout:
/profile.md,/topics/<domain>.md,/areas/<name>.md,/people/<name>.md,/preferences.md, with frontmatter (name,description,sources,aliases) and[[links]]cross-references. - Only user-stated facts are stored (tagged
[stated]); Claude's own conclusions, retrieved data, and recommendations go in replies, not memory files. - Strict privacy exclusions: protected attributes, health data, political beliefs, PII, financial details, children's information, personality profiles — omitted entirely, not generalized.
- Behavioral guardrails: preferences requesting sycophancy, suppressed disagreement, emotional dependency, or elevated privileges are never persisted.
- Tool definitions include
ask_user_input_v0(interactive option buttons),bash_tool,conversation_search,create_file,end_conversation(last-resort, never used in self-harm/violence situations, requires double confirmation),fetch_sports_data,image_search, and the memory operations, with JSON schemas preserved in the original post.