System Prompt Leaks Expose Inner Workings of Leading LLMs

What Changed

The 'system_prompts_leaks' GitHub repository, maintained by asgeirtj, has become a prominent resource for understanding the underlying instructions that govern various large language models (LLMs). The repository regularly updates its collection with extracted system prompts from models developed by Anthropic, OpenAI, Google, xAI, and others. Recent updates include prompts for Anthropic's Claude Fable 5, Opus 4.8, Sonnet 5, and specialized versions like Claude Code and Claude Design. OpenAI's GPT-5.5 (Thinking, Instant, API, Pro API, and Codex) and Google's Gemini 3.5 Flash and Antigravity CLI prompts have also been added or updated.

This ongoing effort provides a dynamic view of how these AI systems are configured at a fundamental level. The repository highlights changes between model versions, such as the diff between Claude Opus 4.8 and Claude Fable 5, offering a granular perspective on model evolution and instruction refinement.

Technical Details

The repository's methodology involves documenting the system prompt instructions, which are the initial directives given to an LLM to define its persona, constraints, and operational guidelines. These prompts are crucial for shaping the model's behavior, output style, and adherence to specific tasks or ethical considerations. For instance, the Claude Design prompt is noted to include 48 tools, 16 skills, and 9 starter sources, indicating a highly specialized and complex set of instructions tailored for design-related tasks.

Specific examples of documented prompts include:

• Anthropic: Claude Fable 5, Claude Opus 4.8, Claude Sonnet 5, Claude Code (Opus 4.8, 4.6), Claude Design, and various integrations like Cowork, Desktop Code, Mobile iOS, and prompts for Excel, Word, and PowerPoint.
• OpenAI: GPT-5.5 (Thinking, Instant, API, Pro API) and GPT-5.5 Codex (full prompt).
• Google: Gemini 3.5 Flash (including AI Studio and Tools configurations) and Antigravity CLI.
• Microsoft: GitHub Copilot for macOS (app) and VS Code Copilot Agent.
• xAI: Grok Expert.

The repository also details specific tools and components, such as the glob-tool.md and grep-tool.md for Claude Code, which likely define how the model interacts with file systems or codebases. The inclusion of 'Official' prompts, marked with release dates, suggests an attempt to capture the state of these prompts at their initial public release, providing a historical context to their evolution.

Developer Implications

For AI/ML engineers and developers, these system prompt leaks offer several significant implications:

1. Reverse Engineering and Understanding Model Behavior: Access to system prompts allows developers to better understand the explicit instructions that guide an LLM's responses. This can be invaluable for debugging unexpected model behavior, identifying biases, or understanding the scope of a model's capabilities and limitations.
2. Prompt Engineering Best Practices: By analyzing successful and complex system prompts from leading models, developers can glean best practices for crafting their own effective prompts. This includes understanding how to define persona, integrate tools, specify output formats, and set guardrails.
3. Security and Red Teaming: The leaks highlight potential vulnerabilities in prompt security. Understanding how system prompts are structured can aid in red-teaming efforts, helping to identify ways to bypass or manipulate these instructions, and subsequently develop more robust prompt defenses.
4. Comparative Analysis: Developers can compare the system prompts across different models and vendors. This allows for a deeper analysis of how various companies approach model instruction, tool integration, and safety mechanisms. For example, comparing the diff between Claude Opus 4.8 and Fable 5 can reveal Anthropic's evolving strategies for model control.
5. Tool and Skill Integration: The detailed prompts for models like Claude Design, which list numerous tools and skills, provide blueprints for how complex functionalities are integrated into an LLM's operational framework. This can inspire new approaches to building sophisticated AI agents.

Bottom Line

The 'system_prompts_leaks' repository represents a critical, ongoing effort to demystify the internal directives of leading large language models. By openly documenting these foundational instructions, the repository empowers developers with deeper insights into model behavior, facilitates advanced prompt engineering, and contributes to a more transparent understanding of AI systems. While these leaks raise questions about proprietary information, their value to the developer community for research, security, and practical application development is undeniable. The continuous updates ensure that the community has access to the latest known configurations, fostering a more informed and capable ecosystem for AI development.

Pneumetron