How to read system_prompts_leaks: an open-source library of AI product system prompts

asgeirtj/system_prompts_leaks is a repository that organizes system prompts from AI products, covering related content for Anthropic Claude, OpenAI ChatGPT, Google Gemini, xAI Grok, Cursor, Copilot, Perplexity, and others.

Project repository:

https://github.com/asgeirtj/system_prompts_leaks

Online page:

https://asgeirtj.github.io/system_prompts_leaks/

How to Browse It

The simplest way is to open the online page and browse by vendor and product:

Anthropic — Claude
OpenAI — ChatGPT
Google — Gemini
xAI — Grok
Perplexity
Microsoft — Copilot
Cursor
Meta, Mistral, Notion, Qwen, and others

If you want a local copy, clone the repository:

1
2


git clone https://github.com/asgeirtj/system_prompts_leaks.git
cd system_prompts_leaks

Then open the corresponding Markdown files directly.

What It Is Useful For

This repository is better suited for research and comparison:

How different AI products define roles and boundaries.
How tool calling, search, coding, image generation, and other capabilities are described.
How products handle safety, privacy, refusals, and citations.
How developer tools such as Cursor, Copilot, and Claude Code organize system prompts.
Which structures are useful when writing your own Agent prompt.

What Not to Do

Do not treat this content as a set of “universal jailbreak prompts.” System prompts change, and their sources and timeliness also need judgment. A more practical use is to study structure:

How role definitions are written.
How tool boundaries are written.
How output formats are constrained.
When the model should ask questions.
When the model should refuse or degrade gracefully.

Usage Suggestions

When reading this kind of material, you can build a comparison table:

1

Product | Role definition | Tool instructions | Safety boundaries | Output format | Special rules

Then turn the useful patterns into your own prompt templates. Do not copy whole sections directly, because your product, tools, and user scenarios are usually different.

Notes

System prompt libraries are useful for studying AI product design, but do not overinterpret them. Product behavior is not determined only by system prompts. It is also affected by model versions, tool implementations, policy layers, context, and frontend product logic. Treat it as observation material, not as a complete manual.