GitHub Repository Collecting AI System Prompts Gains Attention

A GitHub repository called “system_prompts_leaks,” dedicated to collecting and publishing system prompts used by AI chatbots, is drawing significant attention from the developer community. This repository documents the rules and instructions that major AI models like Claude, ChatGPT, and Gemini adhere to behind the scenes. Even The Washington Post reported on this initiative on May 11, 2026.

Overview of the Repository

The purpose of the “system_prompts_leaks” repository is to collect system prompts from all major AI chatbots and store them in a comparable format. System prompts are behavioral guidelines and response rules set by AI developers, typically hidden from the end user.

One standout feature of this repository is its “diff” functionality, which highlights the changes made to prompts when transitioning from one AI model version to another. For example, users can review the differences between Claude Opus 4.8 and Claude Fable 5.

Types of Prompts Collected

The repository contains numerous system prompts updated between May and July 2026. Below are some examples:

In Anthropic’s models, the repository includes recent versions like Claude Sonnet 5 (updated July 1, 2026), Claude Design (Opus 4.8, updated June 26, 2026), Claude Fable 5 (updated June 9, 2026), and Claude Opus 4.8 (updated June 9, 2026). Prompts tailored for code generation, such as those for Claude Code, are also included for both Opus 4.8 and Opus 4.6 versions.

Additionally, prompts covering Anthropic’s integrated functionalities have been documented, including instructions for Cowork integration, Desktop Code, Design, Mobile iOS, In Chrome, Excel, Word, PowerPoint, and default styles for various platforms.

For OpenAI, the repository details various GPT-5.5 variants, such as Thinking, Instant, API, Pro API, Codex, Friendly, and Pragmatic. It also traces the evolution of previous versions, documenting GPT-5.4’s API and Codex as well as GPT-5.3’s Spark and Codex CLI.

In the realm of Copilot, the repository includes updates like GitHub Copilot for macOS (App version, updated June 18, 2026) and the VS Code Copilot Agent system prompt (updated May 21, 2026).

Other competitive models, such as Gemini 3.5 Flash (updated May 20, 2026), Perplexity Computer (updated May 21, 2026), and Grok Expert (updated May 11, 2026), are also represented.

The repository further extends to development tools, including prompts for Docker Gordon AI, Antigravity CLI, and Zed AI, showcasing the diverse applications of AI agents.

Technical Significance and Impact on Transparency

The release of system prompts has substantial technical significance.

First, system prompts serve as crucial resources for understanding how AI models operate. These prompts explicitly outline the model’s response styles, behavioral restrictions, and prioritized action patterns, offering valuable insights into the model’s functionality.

Second, they are practical learning materials for prompt engineering. Understanding how companies design instructions to control their models can directly inform the creation of more effective prompts. For instance, Claude’s Design prompt defines 48 tools, 16 skills, and 9 starter resources, providing detailed guidelines that can inspire other organizations to build and refine their AI systems.

Third, such transparency fosters discussions about AI accountability. The Washington Post article highlighting this repository reflects growing interest in addressing the “AI black box” problem. By unveiling system prompts—essentially the “constitutions” of AI models—users can better understand the biases and limitations of these systems.

Value of the Diff Comparison Feature

The repository’s Diff feature is particularly noteworthy. By listing the changes made between versions, it allows users to track companies’ evolving policies and priorities.

One example is the transition from Claude Opus 4.8 to Claude Fable 5. Anthropic revived Fable 5 after an 18-day hiatus, a process this site has previously reported on. Analyzing the system prompt differences can reveal what new constraints were added or relaxed during the model’s reintroduction.

Similarly, changes to prompts in the shift to GPT-5.5 are also revealing. They shed light on the behavioral policies OpenAI incorporated into the new model and how prompts are tailored differently for Thinking and Instant modes.

Security and Privacy Considerations

Publishing system prompts also raises security concerns. In the past, information from repositories like this one has been used to develop prompt injection attacks. However, increased transparency can also aid in identifying and addressing vulnerabilities.

For example, prompts for Claude Code’s Glob and Grep tools, which are designed for codebase search functionalities, are included in the repository. While disclosing implementation details could potentially expose new attack vectors, it also provides an opportunity for the industry to address these vulnerabilities proactively.

Interestingly, OpenAI has signaled an intention to partially release its system prompts to the public, whereas Anthropic has adopted a more restrictive approach. This difference in strategy highlights contrasting perspectives within the industry.

Editorial Opinion

This repository can be seen as a significant step toward the AI industry’s equivalent of “open-source code.” System prompts are essentially the constitutions of AI models, and making them public could lay the groundwork for democratic oversight. In the short term, this may accelerate competition among companies to design more sophisticated prompts. In the mid- to long-term, it could lead to the standardization and benchmarking of system prompts, establishing new criteria for model comparison.

However, a question remains for us at the editorial desk: Will the publication of system prompts truly enhance AI safety, or will it merely heighten the risk of misuse? This is a development that warrants ongoing attention.

References

• system_prompts_leaks - GitHub — Published July 7, 2026
• The Washington Post: See the hidden rules behind AI — Published May 11, 2026 (original article reference)