asgeirtj/system_prompts_leaks

What this is

This repository collects the hidden system prompts used by ChatGPT, Claude, Gemini, Grok, Copilot, Cursor, Perplexity, and many other AI products, with versions dated back to May 2025 and new entries posted daily. The files are plain Markdown organized by vendor and product, so the repository works as a reference archive rather than executable software. It suits prompt engineers, AI product researchers, and developers who want to see the exact instructions shipped behind consumer and coding assistants.

Coverage includes both chat interfaces and agentic tools: Claude Code prompts for desktop, headless, and cloud modes; Codex prompts for plan and auto-review modes; Antigravity IDE and CLI prompts; and skill definitions for artifact types such as Slides and Dashboards.

Quick start

No installation or build step is required. Open the README index and pick a vendor section; each row links directly to a prompt file, for example the Claude Fable 5.2 prompt or Claude Code headless prompt. Read or download the Markdown file from GitHub's file view. To track new additions, watch the repository or check the "Most recent additions/changes" table at the top of the README.

Constraints and caveats

• Prompts are reproduced from observed model behavior and submissions; the README invites pull requests when a result differs, so versions may lag or differ from what a live model currently uses.
• Files are documentation, not runnable code. The repository's Python label does not indicate a packaged library or CLI.
• Vendor names and product versions are third-party trademarks; content may change or disappear as providers update their systems.
• Licensing is CC0-1.0, which allows copying and reuse, but the source of each prompt is not independently verifiable from the README.

HotRepo's take

HotRepo recorded daily snapshots from June 21 to November 13, 2026. The project sat around 44,135 stars on June 21, climbed to 55,577 by July 10, and reached 71,316 stars and 11,490 forks by November 13. The July window shows the strongest sampled single-day gains, including 1,704 stars on July 7. The November 13 entry is a Python day-list snapshot with a best position of 6; no all-language daily rank is recorded. The repository has no collected star curve or Hacker News discussion in this dataset, so the snapshot series is the main evidence of interest. That growth is consistent with a reference archive that gains traction whenever a major model or coding tool changes its instructions.