What it does This repository collects system prompts extracted from commercial AI products—ChatGPT, Claude, Gemini, Grok, Copilot, Perplexity, and others—through jailbreaks, prompt-injection tricks, and API responses. Each prompt is stored as a markdown file, dated, and organized by vendor. The repo also tracks older versions, raw dumps, and variant prompts for different modes (thinking, instant, API, etc.).

The interesting bit System prompts are the hidden scripts that tell an LLM how to behave, what tone to use, and what tools it can invoke. Seeing them laid bare is like reading the employee handbook for a robot: you learn that GPT-5.5 has separate “Friendly” and “Pragmatic” Codex personalities, or that Claude’s Excel integration carries its own mini-prompt. The Washington Post even used this repo for a feature in May 2026.

Key highlights

• Covers Anthropic (Opus 4.6–4.8, Sonnet, Claude Code), OpenAI (GPT-5.x family, Codex, o3/o4), Google (Gemini 3.x, Antigravity), xAI (Grok 4.x), Perplexity, and misc tools like Cursor and Zed
• Tracks variants: API vs. consumer, with-tools vs. without, raw vs. cleaned, and dated historical versions
• Recently updated entries are front-and-center in a table; everything else lives in collapsible details blocks
• Includes traffic badges showing weekly GitHub views and clones, which are themselves oddly high for a collection of text files

Caveats

• Extraction methods aren’t documented in the README, so provenance and reliability vary by prompt
• “Leaked” is a strong word for some entries; API system prompts and published behavior files are arguably just… observed
• No analysis or diff tooling provided—you’re on your own to compare Claude 4.6 with 4.8

Verdict Worth bookmarking if you build with LLMs, study prompt engineering, or just enjoy seeing how sausage is made. Skip it if you need verified, legally clean prompt data for production use.