System Prompts Leaks is a GitHub repository that collects the system prompts of commercial AI assistants — the hidden instructions a model is handed before a user's first message — and stores them verbatim as plain Markdown files. According to its own description it covers 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, Gemini 3.1 Pro and Antigravity; and xAI's Grok and Grok Bot, plus Cursor, Kimi and others. It is for prompt engineers, developers building products on these models, and researchers or journalists who would rather read the actual instructions than infer them from a chatbot's behaviour.
What it does
The repository is an archive, not a program. Each captured prompt lives in its own file, and the files are grouped into folders by the company that ships the model — an Anthropic folder with a nested claude-code directory, an OpenAI folder with a Codex subfolder, an xAI folder, and so on. The text is stored as captured, without paraphrase or commentary wrapped around it, so what you read is meant to be the instruction block itself rather than someone's summary of it.
The README keeps a "Most recent additions/changes" table at the top, listing what was added, the date it was added, and a direct link to the file. Recent rows include the Claude Code headless prompt for Fable 5.1 (September 5, 2026), Codex GPT-6-Astra (September 4, 2026), Claude Fable 5.1 (September 1, 2026), Grok 4.6 (August 29, 2026) and Grok Bot (August 21, 2026). That table is the fastest way to see whether anything has changed since your last visit.
The collection has drawn a large audience: 65,075 stars and 10,694 forks at the time of writing. It has also been used as a primary source by newsrooms and analysts. The README points to a Washington Post interactive story built on prompts from the repository (May 11, 2026) and a live data dashboard built from its files by CEPS' AI World (July 10, 2026).
How it works
There is nothing to install and nothing to run. You browse the folder tree on GitHub, open a Markdown file such as Anthropic/claude-fable-5.1.md, OpenAI/Codex/gpt-6-astra.md or xAI/grok-4.6.md, and read it in the browser. GitHub lists JavaScript as the repository's main language, but the substance is text.
The project was created on May 3, 2025 and is still being pushed to — the most recent push in these materials is September 9, 2026 — with the description promising regular updates. It carries a "PRs welcome" badge, so new captures arrive through pull requests from contributors rather than from any vendor. The license is CC0-1.0, a public-domain dedication, which means the collected text can be copied and reused without attribution conditions imposed by the repository itself.
One thing worth being clear about: these are third-party captures. The materials here describe what was collected and when it was added, but they do not describe a verification procedure for any individual file. Treat each document as a snapshot someone extracted, not as something published or confirmed by the company that wrote it.
Getting started
Open the repository and start with the recent-additions table, then drill into the vendor folder for whichever product you care about. If you want to work with the prompts seriously, clone the repository so you can grep across all of them at once and use git log and git diff on a single file to watch how one product's instructions changed between versions — that diff view is often more revealing than any single prompt.
When to use it / when not
It earns its place when you are writing system prompts of your own and want to see how well-funded teams phrase tool descriptions, formatting rules, refusal language and persona framing. It is equally useful for teaching, for reporting, and for comparing how differently two vendors talk to the same underlying capability.
It is the wrong source when you need authority. It is not official documentation, and it is not an API contract: vendors change these instructions silently, so any file may describe a version that no longer ships. Do not build production behaviour on the assumption that a prompt here is current, and do not go looking for model weights, SDKs or runnable code — none of that is the point.
Anyone whose work depends on how commercial assistants actually behave should keep this repository bookmarked. It is the closest thing to a public reading room for the instruction layer that sits between a model and its users, it is maintained often enough to stay interesting, and its public-domain license removes the friction of quoting it. Read it as evidence gathered by enthusiasts rather than as a specification, and it is one of the more genuinely useful reference collections in the field.