system_prompts_leaks: Curated LLM System Prompts

TL;DR: asgeirtj/system_prompts_leaks curates extracted system prompts from Claude, ChatGPT, Gemini, Grok, Cursor, and other major LLMs in one searchable place, released under CC0 so anyone can copy, remix, or study the prompts without legal friction.

The collection at asgeirtj/system_prompts_leaks is a JavaScript-based repo that catalogs the system prompts shipped with commercial large language models. As of this writing it carries 65,438 stars and 10,739 forks, with the maintainer pushing updates within the last few days. The list of covered products reads like a directory of frontier AI: Anthropic’s Claude Fable 5.1, Opus 5, Claude Design, and Claude Code; OpenAI’s ChatGPT GPT-6-Astra and Codex; Google’s Gemini 3.8 Flash, 3.1 Pro, and Antigravity; plus xAI’s Grok, Grok Bot, Cursor, and Kimi.

System prompts are the hidden instructions that shape how an LLM behaves before any user message arrives. They decide whether the model refuses a request, writes code in a particular dialect, adopts a persona, or follows a structured output format. Most providers publish these only partially. Anthropic publishes a system card; OpenAI leaks fragments through custom GPTs; Google’s prompts tend to surface only in academic papers. A consolidated, readable archive is useful for prompt engineers who want to compare approaches, researchers studying model alignment, and developers debugging why two seemingly identical prompts produce different outputs across vendors.

The repo updates regularly — the most recent push was September 9, 2026 — which matters because vendors change their prompts whenever they ship a model revision. Stale archives become misleading within weeks. Topics on the repo include anthropic, openai, gemini, grok, claude-code, cursor, codex, prompt-engineering, and system-prompt, so it shows up in GitHub search for anyone hunting for prompt-level documentation.

The license is CC0-1.0, which is a public domain dedication. You can copy the prompts into your own project, modify them, or republish without attribution. That makes the collection friendlier than the typical “all rights reserved” stance some prompt libraries take. If you are building a benchmark or a research paper that needs to reference exact prompt text, this is the cleanest source available.

For prompt engineers, the practical value is comparison. Reading Claude’s prompt next to Gemini’s, line by line, exposes how each lab structures its instructions. Anthropic tends to use nested bullet lists with explicit safety boundaries; OpenAI leans on tool-calling manifests; Google’s prompts are denser and assume the model has read recent product documentation. You can lift the structural patterns even if your own product runs on a different model.

For researchers, the archive functions as a corpus. You can grep across the entire collection for specific phrases — respond as, you are, tool_use, chain_of_thought — and see which vendors lean on which conventions. The repo does not include chat transcripts or training data, so it stays inside the bounds of what is generally considered fair use.

If you decide to copy a prompt wholesale into your own application, two caveats apply. First, vendors update prompts without notice; the version you copy today may not match what the vendor runs tomorrow. Second, some prompts contain references to internal product features that third parties cannot legally replicate. Treat the archive as a reference for structure and tone, not a turnkey template.

The repo is plain JavaScript with no build step, so cloning and reading it locally is trivial. There is no executable code, no telemetry, no install script — it is a documentation project that happens to live on GitHub.

Source and Accuracy Notes

• Project page: https://github.com/asgeirtj/system_prompts_leaks
• Source repository: asgeirtj/system_prompts_leaks (verified via GitHub API html_url)
• License: CC0-1.0 (verified via GitHub API license.spdx_id)
• Stars at time of writing: 65,438