System Prompts Leaks: AI System Prompts Collection
/system_prompts_leaks
Reach for this if you write prompts or build on top of LLMs and want to see how the big products actually steer their models, roughly the wording Anthropic, OpenAI, and Google appear to ship. Skip it if you need a canonical, vendor-verified source: these are collected and reverse-engineered prompts, so read them as informative reference, not ground truth.
The System Prompt You Never Get to See
If you build with LLMs, the system prompt is the part you never see: the hidden instructions that shape tone, refusals, tool use, and output formatting. Vendors rarely publish these, so prompt engineers end up guessing at what makes a product behave the way it does. This repo pulls the extracted prompts into one place so you can read them side by side instead of chasing scattered tweets and gists.
A Collected Archive, Not a Tool
System Prompts Leaks is an AI system prompts collection: a CC0-licensed GitHub repo that documents the extracted system prompt instructions behind well-known AI products. It covers prompts attributed to Claude (including Claude Code), ChatGPT and Codex, Gemini, Grok, Cursor, GitHub Copilot, Perplexity and others, stored as per-vendor Markdown files. It is reference material to read, not a runnable tool, API, or library.
56k Stars and a Washington Post Mention
The README says it was featured in The Washington Post in May 2026. The pull tracks the wider interest in prompt engineering and agent building: people want to read the actual instructions behind Claude Code, ChatGPT, and Gemini rather than theorize about them. The README's 'Recently Updated' table shows it adds new model versions as they ship, listing entries like GPT-5.6 and Claude Sonnet 5.
Sorted by Vendor, Tracked as They Change
- โSystem prompts sorted by vendor: Anthropic, OpenAI, Google, xAI, Microsoft, Perplexity, Cursor, Meta, Mistral and more
- โCoverage of coding agents specifically, including Claude Code, Codex, Cursor, and GitHub Copilot, not only chat models
- โA 'Recently Updated' table in the README that tracks which prompts changed and on what date
- โPlain Markdown files you can read, search, and diff in the browser or after a clone
- โExtras beyond the main prompt: tool definitions, older model versions, and per-personality variants
- โCC0-1.0 (public domain dedication) license on the repository
Studying How Products Steer Their Models
- โขStudying how shipping AI products phrase refusals, tool instructions, and formatting rules
- โขBorrowing structure and patterns for your own system prompts and agent instructions
- โขComparing how a model's prompt appears to change between versions using the collected files
- โขTeaching or writing about how LLM products are steered behind the scenes
Clone It or Just Read It on GitHub
There's nothing to install: it's a documentation repo, not a package. Read it directly on GitHub, or clone it with `git clone https://github.com/asgeirtj/system_prompts_leaks` to grep and diff the files locally. Everything is Markdown organized into per-vendor folders.
Finding the Prompt You're Looking For
Open the vendor folder for the product you care about, for example the Anthropic folder for Claude prompts or OpenAI for ChatGPT and Codex, then read the Markdown file. The README's tables link straight to each prompt and flag recently updated ones. Since the files are plain text under CC0, you can copy passages, diff two versions, or search across the whole repo after cloning.
One Place to Compare Vendors
- โOne place to read the collected instructions behind many AI products, side by side
- โBroad coverage across vendors and into coding agents, not just the flagship chat models
- โPlain Markdown under CC0, so it's easy to search, diff, and reuse
- โUpdated as new models ship, with a changelog-style table in the README
Collected, Not Confirmed
- โณThese are extracted and community-collected prompts, not vendor-published ones, so authenticity and exact wording can't be fully verified
- โณSystem prompts change silently, so a file can be stale the moment a vendor updates its product
- โณIt's reference material, not runnable code: there's no tool, API, or library to integrate
- โณDepth varies by product; some entries are full prompts while others are partial, older, or personality variants
Other Prompt Archives and Official Docs
Prompt Engineers Who Verify Before They Cite
Prompt engineers, agent builders, and researchers who want to see how shipping AI products are actually instructed, and who accept that these are collected rather than official prompts. If you need an authoritative, vendor-verified source for a paper, a compliance claim, or a product decision, treat this as a starting point and confirm against official docs.
System Prompts Leaks FAQ
System Prompts Leaks is an AI system prompts collection on GitHub. It documents the extracted system prompt instructions behind products like Claude, ChatGPT, Gemini, and Grok, stored as per-vendor Markdown files you read rather than run.
Treat them as informative, not authoritative. They're extracted and community-collected, so exact wording and authenticity can't be fully verified, and any prompt can be out of date once a vendor changes its product.
The repository carries a CC0-1.0 (public domain) license, so the collected text is dedicated to the public domain per the repo. That covers reuse of the files; it doesn't verify that each prompt matches the vendor's live product.
The README says it's updated regularly and keeps a 'Recently Updated' table with dates, adding new model versions as they ship. Because prompts change quietly, the collection still lags whatever a vendor is running right now.
Related repositories
Want a second opinion on system_prompts_leaks?
Ask an AI that can read this page โ one click and you get its take on system_prompts_leaks.