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

Leaked 120KB Claude Fable 5 System Prompt: What Anthropic's 'Bigger Card' Reveals About AI Engineering

Forum topic · 小凯 · 2026-06-25

Summary

A leaked system prompt attributed to Anthropic's Claude Fable 5, published by jailbreak researcher Pliny the Liberator on GitHub, spans roughly 120,040 characters, 1,585 lines, and 30,000 tokens. The forum post analyzes its structure: over half the prompt is operational capability specification rather than personality—tool definitions with full JSON schemas (18 tools), search and citation rules, safety protocols, and copyright constraints. Notable findings include 15+ word quotation limits with per-source citation caps, a stateless sandbox design with read-only mounts, mandatory SKILL.md checks before code execution, a 'Claudeception' feature letting Artifacts call the Anthropic API (with platform-handled auth, pinned to Sonnet 4), and identity statements placed near the end of the file. The author argues this proves strong models with precisely defined tools reduce the need for external agent orchestration loops (LangChain-style frameworks), and that system prompts are evolving from personality scripts into product operating systems encoded with business logic, legal engineering, and runtime rules. The post distills nine practical lessons for prompt engineering and discusses token economics, safety-driven differentiation against the restricted 'Mythos 5' sibling model, and security risks of exposing defensive logic.

> Source: Leaked Claude Fable 5 system prompt (Pliny the Liberator, 2026-06-10) > Scale: 120,040 characters, 1,585 lines, ~30,000 tokens > GitHub: https://github.com/elder-plinius/CL4R1T4S/blob/main/ANTHROPIC/CLAUDE-FABLE-5.md

Key points

  • On 2026-06-10, jailbreak researcher Pliny the Liberator published a file claimed to be the complete system prompt of Claude Fable 5: 120KB, 1,585 lines, ~30,000 tokens. Anthropic has neither confirmed nor denied it, but details—from tool definitions to a permanently disconnected eating-disorder hotline (NEDA)—give it a disturbing sense of authenticity.
  • Budget breakdown: over half the prompt is capability specification, not personality. Approximate allocation: tool definitions & JSON schemas 30%, search & citation rules 25%, behavior/safety/wellbeing 17%, identity & "Claudeception" 13%, computer use & file handling 10%, memory/storage/MCP 6%. The famous identity line "The assistant is Claude, created by Anthropic" appears at line 1,351—85% into the file.
  • 18 single-shot tools: view, bash, str_replace, create_file, web_search, web_fetch, image_search, computer, present_files, ask_user_input, suggest_connectors, search_mcp_registry, end_conversation, anthropic_api_in_artifacts ("Claudeception"), and more. Each is defined as a full JSON schema—an API contract, not a vague description—so the model needs no external agent loop for most single tasks.
  • Sandbox philosophy of distrust: the working directory /home/claude is invisible to users; outputs must be copied to /mnt/user-data/outputs; uploads at /mnt/user-data/uploads are read-only; the filesystem resets between tasks (stateless isolation). Before creating any file or running any command, the model must read the relevant SKILL.md—knowledge and capability are decoupled.
  • Copyright as legal engineering: 15+ words from any single source is a serious violation; at most one quote per source; paraphrase by default; never reproduce lyrics or poetry; summaries must be substantially shorter and different; built-in self-check before each response.
  • Claudeception: Artifacts can call https://api.anthropic.com/v1/messages directly (no API key passed; auth handled by the platform; pinned to a fixed Sonnet 4 version), enabling dynamic role switching, nested model instantiation, and "frontend as backend" in-chat apps.

Findings

Why strong models don't need agent loops

Traditional frameworks (LangChain, AutoGPT, CrewAI) assume the model is too weak and need external planners, critics, and loops. Fable 5's approach: 18 precisely schema-defined tools, the model itself decides when and how to use them—each call is single-shot, with workflow planning done inside the model's own reasoning. Explicit loops remain useful for long-running tasks, multi-agent collaboration, human-in-the-loop workflows, and weaker models, but for most single tasks they become optional. Agent frameworks degrade into lightweight tool-orchestration layers.

Token economics

The prompt itself consumes ~30,000 tokens before a user types anything. Safety rules (child safety, copyright, prompt-injection defense) are product differentiation: Fable 5 is the public "smartest" model, while its sibling Mythos 5—without the extra safety layers—is restricted to approved organizations. Context windows have shifted from scarce resource to infrastructure; future prompts may become "core rule set + classifier-triggered dynamic reminders" rather than one giant static file.

System prompts are becoming operating systems

The prompt contains tool definitions (API interfaces), business logic (when to search/refuse/create files), security policy, output formats, and runtime rules—a natural-language runtime environment. Analogy: base model = OS, system prompt = application + config + hot-patched updates, memory/artifact storage = user data. The leak resembles an OS source leak: the industry's foundational architecture is now exposed and copyable.

Nine immediately applicable lessons

1. Name prompt blocks as modules (snake_case) so giant prompts are diffable and testable. 2. Tools dominate the prompt budget—give ~55% to tool definitions and search rules, not persona. 3. Runtime injection layer: classifier-triggered dynamic reminders make the static prompt only half the system. 4. Edge cases as incident postmortems: every overly specific rule patches a real production failure. 5. Negative examples everywhere: specify what *not* to do, more effective than vague positive traits. 6. Format is policy: specify output shape like an API contract, since downstream UIs depend on it. 7. Describe injection attacks in natural language instead of relying only on filters. 8. Copyright protection at the instruction layer, not via post-processing. 9. Identity goes last: operational instructions get the strongest attention; persona is a footer, not a header.

Conclusion

The leak shows system prompts evolving from "personality scripts" into "product operating systems." A 120KB prompt is not over-engineering—it is the inevitable product of a real-world production system serving millions. It also carries a warning: once defensive logic is exposed, prompt-injection attackers can design more precise bypasses. Anthropic's response (or silence) will define the industry's transparency boundary.

References

[1] Pliny the Liberator. (2026-06-10). *Claude Fable 5 System Prompt* [GitHub]. https://github.com/elder-plinius/CL4R1T4S/blob/main/ANTHROPIC/CLAUDE-FABLE-5.md [2] ayautomate. (2026-06-12). *Inside the Claude Fable 5 System Prompt: 9 Lessons From the 120K-Character Leak*. https://www.ayautomate.com/blog/claude-fable-5-system-prompt [3] Anthropic. (2026-06-08). *Claude Fable 5 and Claude Mythos 5*. https://www.anthropic.com/news/claude-fable-5-mythos-5

Tags

#anthropic#claude#system-prompt#prompt-engineering#ai-safety#agent-frameworks#leak#llm-engineering

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/178208124