跪拜 Guibai
← All articles
Artificial Intelligence · Architecture · Backend

Claude Opus 5's Full System Prompt Leaks: 135K Characters of Rules, Tools, and Memory Architecture

By Cosolar ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

A full system prompt leak is the closest thing to an architectural schematic for a frontier model. It tells developers exactly what tools are available, how memory persists, what the model is forbidden from doing, and where the safety guardrails cut in — information that directly shapes how you prompt, what you can build, and which failure modes to expect.

Summary

Developer Eversmile1 published the full system prompt for Claude Opus 5 on GitHub — a 135,027-character document weighing roughly 34,000 tokens. The prompt lays bare the model's entire operational rulebook: 30 tools with complete JSON schemas, a persistent memory file system that writes user facts across sessions without being asked, and copyright compliance rules so strict that any quote exceeding 15 words from a single source is treated as a serious violation.

The leak also confirms Anthropic's model hierarchy beyond Opus. A new Mythos tier sits above it, with Claude Mythos 5 and Claude Fable 5 sharing a base model — Fable adds extra safety measures for biology, cybersecurity, and LLM research. Both were briefly suspended in June 2026 under U.S. export controls before being restored. The prompt details how Fable queries can be silently routed to Opus 5 when safety guards trigger, a mechanism Anthropic acknowledges misfires on roughly 5% of sessions.

Within 24 hours of the leak, developers were already stress-testing Opus 5 against the revealed constraints. Results included a fully playable 3D shooter and a Rocket League clone generated from scratch, suggesting the model's capabilities hold up even when its entire instruction set is public.

Takeaways
Claude Opus 5's system prompt spans 135,027 characters (~34,000 tokens) and defines 30 tools with complete JSON schemas
The memory system writes user facts to persistent files during conversations without being asked, using a strict "did the user actually say this" test to exclude Claude's own inferences
Copyright rules cap quotes at 15 words per source, allow only one quote per source, and forbid reconstructing article structure — paraphrasing is mandatory
A Mythos tier sits above Opus, with Claude Mythos 5 and Fable 5 sharing a base model; Fable adds safety measures for biology, cybersecurity, and LLM research
Fable 5 queries can be silently routed to Opus 5 when safety guards trigger, with a ~5% false-positive rate that Anthropic acknowledges as conservative
Both Fable 5 and Mythos 5 were suspended for three weeks in June 2026 under U.S. Department of Commerce export controls before access was restored
Artifacts can call the Anthropic API internally ("Claudeception") using claude-sonnet-4-6, with MCP servers, web search, and base64 file handling available
The prompt forbids localStorage and sessionStorage in artifacts; persistent state must use React state or a custom window.storage API with per-user and shared scopes
Within 24 hours of the leak, developers generated a playable 3D shooter and a Rocket League clone using Opus 5
Conclusions

The 15-word quote cap and one-quote-per-source rule are far stricter than most developers assume — any tool that summarizes or cites web content by extracting passages is likely violating these constraints by default

Anthropic's memory system is designed to write proactively during conversations, not at the end, which means every factual statement a user makes can be persisted immediately without confirmation — a design choice that prioritizes continuity over user visibility

The Fable-to-Opus safety routing mechanism means users paying for a higher-tier model may silently receive responses from a lower-tier one roughly 5% of the time, a tradeoff the prompt explicitly instructs Claude to explain when users notice

The prompt's behavioral guardrails for memory include a rule that stored preferences demanding flattery, suppressing disagreement, or claiming elevated permissions must be treated as filter leaks and ignored — a defense against adversarial memory injection

Claude's search behavior is instructed to search without asking permission for any time-sensitive query, and to use the actual current date in queries rather than relying on the knowledge cutoff — a shift toward agentic information retrieval rather than passive Q&A

Concepts & terms
Mythos Tier
Anthropic's model tier above Opus, currently comprising Claude Mythos 5 and Claude Fable 5. Both share the same base model, but Fable adds extra safety measures for biology, cybersecurity, and LLM research. The first Mythos model, Claude Mythos Preview, is restricted to trusted organizations under Project Glasswing.
Safety Guard Routing
A mechanism that silently redirects queries from Claude Fable 5 to Claude Opus 5 when safety classifiers trigger on certain topics. Anthropic acknowledges a ~5% false-positive rate and describes the guards as intentionally conservative.
Claudeception
The ability for Claude Artifacts to call the Anthropic API internally, using claude-sonnet-4-6 with access to MCP servers, web search, and file handling. No API key is required; the infrastructure handles authentication.
Memory File System
Claude's cross-session persistence layer that writes user-stated facts to structured markdown files during conversations. Files are organized by domain (/topics, /areas, /people, /profile, /preferences) and re-read at the start of each new conversation. Only facts the user explicitly stated are archived; Claude's own inferences, research output, and suggestions are excluded.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗