asgeirtj/system_prompts_leaks

⭐ 39,970  ·  None  ·  GitHub Repo

Extracted system prompts from ChatGPT (GPT-5.5 Thinking), Claude (Opus 4.7, Opus 4.6, Sonnet 4.6, Claude Code), Gemini (3.1 Pro, 3 Flash, Gemini CLI), Grok (4.3 beta), Perplexity, and more. Updated regularly.

ai ai-transparency anthropic chatgpt claude claude-code gemini generative-ai

1-Sentence Summary

A centralized, community-driven repository of extracted system prompts from every major AI model, updated regularly.

🔥 Key Capabilities & USP

• Comprehensive Model Coverage: Includes system prompts from ChatGPT (GPT-5.5, GPT-5.4, Codex, o4-mini, o3), Claude (Opus 4.7, Opus 4.6, Sonnet 4.6, Claude Code, Cowork), Gemini (3.1 Pro, 3 Flash, Gemini CLI), Grok (4.3 Beta, 4.2), Perplexity, and Meta AI. Solves the pain point of hunting across forums and tweets for leaked prompts.
• Regular Updates with Latest Models: Prompts are updated frequently, with recent additions like GPT-5.5 Thinking, Claude Cowork, and Grok 4.3 Beta. This eliminates the frustration of stale, outdated prompt collections.
• Organized by Provider and Variant: Clear directory structure (OpenAI/, Anthropic/, Google/, xAI/) with separate files for each model, personality variant (Friendly, Pragmatic), and tool-specific prompts (Web search, Deep research, Canvas). Solves the pain point of messy, unsearchable text dumps.
• Community-Contributable via PRs: Anyone can submit corrections or new leaks via pull requests. This turns a single point of failure into a crowdsourced, self-correcting dataset that scales with the AI landscape.
• Historical Versioning: Includes older/legacy versions alongside the latest, enabling comparative analysis of how model instructions evolve over time.

Technical Architecture

The project is a flat file-based collection of Markdown documents. No code, no build system, no dependencies.

Component	Description
Storage	Git repository (GitHub)
Data Format	Markdown (.md) files
Organization	Directory per provider (e.g., OpenAI/, Anthropic/)
Navigation	README table of contents with links to each model file
Versioning	Git history for tracking changes over time
Contribution	Pull requests for new prompts or corrections

Architectural Highlight: The simplicity is the strength. Zero infrastructure means zero maintenance cost. Anyone with a GitHub account can browse, fork, or contribute.

Quick Start Guide

No installation or code execution is required. The project is a static document repository.

1. Navigate to the repository: https://github.com/asgeirtj/system_prompts_leaks
2. Browse by provider: Click into the OpenAI/, Anthropic/, Google/, or xAI/ directories.
3. Read a prompt: Click on any .md file (e.g., OpenAI/gpt-5.5-thinking.md) to view the extracted system prompt.

To contribute a new prompt:

# Fork the repository
# Create a new branch
# Add your .md file to the correct provider directory
# Submit a Pull Request

Pros, Cons & Use Cases

Pros

• Extremely comprehensive: Covers more models and variants than any other single source.
• Always current: Updated with the latest model releases (GPT-5.5, Claude Opus 4.7).
• Zero friction to use: No code, no API keys, no setup. Just read the files.
• Community-verified: PRs allow the community to correct inaccuracies and fill gaps.
• Historical record: Provides a timeline of how system prompts evolve, which is invaluable for research.

Cons

• No automated extraction tools: Relies on manual collection and leaks; accuracy is not guaranteed.
• No analysis or comparison tools: You get raw text, not a diff view or structured comparison.
• Potential for outdated or incomplete prompts: Some entries may be partial or from older model versions.
• No search functionality: Beyond GitHub's file search, there is no built-in querying capability.

Who should NOT use this?

• Developers needing a programmatic API: If you need to query prompts via an API or integrate them into a pipeline, this raw Markdown repo will require significant scraping effort.
• Users seeking 100% verified, official prompts: Since many prompts are leaked or injected, they may contain inaccuracies or be intentionally misleading.
• Teams requiring a managed, SLAd service: This is a community project with no guarantees of uptime, accuracy, or update frequency.

Ideal Use Cases

• AI Researchers: Studying how model instructions (personality, constraints, refusal policies) change across versions and providers.
• Prompt Engineers: Reverse-engineering the "secret sauce" of top models to improve their own prompt crafting.
• Transparency Advocates: Documenting and publicizing the hidden instructions that govern AI behavior.
• Competitive Analysis: Comparing how OpenAI, Anthropic, Google, and xAI instruct their models differently.
• Security Researchers: Analyzing injection vectors and system prompt vulnerabilities.

Community & Activity

This project is on fire. With 39,970 stars and regular updates (last updated May 10, 2026), it has clearly struck a nerve in the AI community. The momentum is driven by the insatiable demand for transparency in an increasingly opaque AI landscape. The active inclusion of the very latest models—GPT-5.5 Thinking, Claude Cowork, Grok 4.3 Beta—shows a dedicated maintainer and community that stays on the bleeding edge. If you care about what your AI is really being told to do, this is the single best resource to watch, fork, and contribute to.