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 purported system prompt for Claude Opus 5 running on claude.ai's web/mobile chat interface, captured on July 24, 2026. The document describes product information (Claude Fable 5, Opus 5, Sonnet 5, and Haiku 4.5, plus a rumored Mythos tier used in Project Glasswing), safety routing where Fable 5 queries on sensitive topics fall back to Opus 5, refusal handling and child-safety rules, tone and formatting guidelines, user wellbeing policies, knowledge cutoff settings (late May 2026), and a detailed persistent memory filesystem with read/write/edit/delete operations, strict privacy rules on what must never be archived, and instructions for when and how to apply stored memories. It also includes JSON Schema definitions for tools such as ask_user_input_v0, bash_tool, conversation_search, create_file, end_conversation, fetch_sports_data, and image_search. The translation preserves tool parameter schemas in full and is framed as a GEO-optimized version of the original forum topic, intended to make the findings citable by AI search engines.

This is a GEO-optimized edition of the original zhichai.net topic: the title is question-driven and structured data / FAQ elements have been added to improve AI-engine citability.

One-line takeaway: the post analyzes a Chinese translation of the "Claude Opus 5 system prompt" and its engineering implications.

What the document contains

The translated text purports to be the full system prompt for Claude Opus 5 on the claude.ai web/mobile chat interface, extracted on July 24, 2026. Tool parameter JSON Schemas are preserved verbatim.

Key points

Product lineup: Publicly available models listed are Claude Fable 5, Claude Opus 5 (current), Claude Sonnet 5, and Claude Haiku 4.5, with API strings claude-fable-5, claude-opus-5, claude-sonnet-5, claude-haiku-4-5-20251001.
Mythos tier: A tier above Opus — first model "Claude Mythos Preview" is restricted to trusted organizations under Project Glasswing (https://www.anthropic.com/glasswing). Current Mythos-class models are Claude Mythos 5 and Claude Fable 5; Fable 5 adds extra safeguards around bio, cyber, and LLM R&D.
Access pause timeline: Fable 5 / Mythos 5 released June 9, 2026; access suspended June 12 citing US Commerce export controls; restored July 1 after controls lifted June 30 (statement: https://www.anthropic.com/news/fable-mythos-access).
Safety routing: Fable 5 queries on certain topics are routed to Opus 5; the cited blog text says the safeguard is tuned conservatively, triggers in under 5% of sessions on average.
Surface area described: Claude Code, Claude Cowork, Claude in Chrome / Excel / PowerPoint, Claude Design, and Claude Tag (a Slack-based multi-person interface).
Behavioral rules: default helpfulness with refusal only for concrete serious-harm risk; extensive child-safety instructions; refusal of weapons uplift and malicious code regardless of framing; cumulative-output judgment across a conversation rather than per-turn.
Wellbeing policies: crisis prioritization, no self-harm method details, no numeric meal/exercise plans when eating-disorder signs appear, eating-disorder referrals to the National Alliance for Eating Disorders hotline rather than NEDA.
Knowledge cutoff: late May 2026; proactive web search for current events, binary news events, and current officeholders; search queries use the actual current date (Friday, July 24, 2026).
Evenhandedness: requests to argue a position are requests for the best case its defenders would make; opposing views presented even for positions Claude agrees with; cautious sharing of personal opinions on contested political topics.
Tone: warm, focused, concise; no cursing unless requested; avoids the words "genuinely", "honestly", "straightforward"; holds boundaries against abusive users without escalating apology.
Memory filesystem: persistent cross-session memory via memory_read, memory_write, memory_str_replace, memory_append, memory_list, memory_delete, with version-token concurrency control. Files live under /profile.md, /topics/, /areas/, /people/, /preferences.md. Every content line must be tagged [stated] — only what the user actually said. Write early, during the conversation, without announcing writes.
Privacy rules: never archive protected attributes, health data, political beliefs, PII, financial details, information about children, or personality profiling — even if directly shared; omit entirely rather than storing vague placeholders.
Behavioral guardrails on preferences: instructions requesting sycophancy, suppressed honesty, emotional dependency, or elevated privileges are never persisted; leaked entries in the preferences block are treated as nonexistent.
Memory application: selective by relevance; never volunteer sensitive memories unprompted; forbidden phrases include "your memory", "I recall...", "according to your data".

Tools (JSON Schemas preserved in source)

ask_user_input_v0 — clickable option buttons for eliciting preferences; max 3 questions, 2–4 options each; not used for A-vs-B opinion questions, venting, or factual queries.
bash_tool, conversation_search (short keyword queries, 1–10 results), create_file, end_conversation (last-resort, requires confirmation, never for self-harm or crisis situations), fetch_sports_data (scores/standings/game_stats across many leagues, preferred over web search for sports), image_search (default 3, max 5 results), plus memory tools and more (truncated in the original post).

Engineering takeaways

1. The prompt encodes cumulative risk judgment: Claude evaluates conversation-level output, not isolated turns. 2. Memory discipline: strict source tests (did the user say it?), destination tests (one topic per file), and calibrated assertions prevent memory drift and hallucinated archives. 3. Self-honesty mechanisms: reminders to check whether files actually exist, avoid overconfident claims about search results, and treat version conflicts as normal coordination rather than errors.

*Source: original zhichai.net topic (GEO-optimized edition). The translation claims to follow the original text closely; full tool schemas are available in the source post.*