Open-source project
asgeirtj/system_prompts_leaks avatar
asgeirtj/system_prompts_leaks

System Prompts Leaks: A Raw Archive of AI's Hidden Instructions

GitHub describes it as Extracted system prompts from Anthropic - Claude Fable 5, Opus 5, Claude Design, Claude Code. OpenAI - ChatGPT GPT-5.6-Sol, Codex. Google - Gemini 3.5 Flash, 3.1 Pro, Antigravity. xAI - Grok, Cursor, Copilot, VS Code, Perplexity, and more. Updated regularly.. The repository metadata lists JavaScript as its primary language. The metadata lists the CC0-1.0 license. This article stays within the project description and details documented in the GitHub repository README.

71,476 stars11,507 forksJavaScriptCC0-1.0

At a glance

What is it?
This repository collects verbatim system prompts from major AI products, offering a unique look at the rules that shape model behavior. It is a reference archive, not a tool, and its value depends on how current and accurate each capture is.
Who is it for?
Adopt this repository if you need a central, openly licensed reference for AI system prompts, whether for research, journalism, or building your own agent prompts. Skip it if you require guaranteed accuracy or production-ready documentation, since captures are unofficial and may contain errors.
Can I use it commercially?
Yes. CC0-1.0 is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
Is it still maintained?
Yes. The repository received new commits within the last day.
What is it written in?
Mainly JavaScript, according to GitHub's language statistics.

Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.

From daily snapshots

Star history

  • asgeirtj/system_prompts_leaks71,476 ★ · +8,640
Star history
Repository2026-08-132026-11-15
asgeirtj/system_prompts_leaks62.8k71.5k

8,640 Stars in the last 94 days, measured from the daily GitHub snapshot. Compare with other projects

Editorial analysis

What This Repository Actually Provides

This is not a software library. It is a collection of text files, each holding what the maintainer describes as leaked system prompts captured verbatim from AI products. The README lists captures from Anthropic's Claude Fable 5, Opus 5, and Claude Code, OpenAI's ChatGPT 5.6 Sol and Codex, Google's Gemini 3.5 Flash and 3.1 Pro, plus xAI's Grok, and integrations like Cursor, Copilot, VS Code, and Perplexity. The intended audience is anyone who wants to see the hidden instructions that shape AI behavior before the first user message. That includes journalists, researchers, and developers building their own prompt-based systems. The Washington Post and CEPS' AI World are cited as having used the repository for interactive stories, which suggests practical value beyond casual curiosity.

The Capture Mechanism: Verbatim, But How?

The README describes these as the full verbatim system prompts and invites readers to open a PR or get in contact if they got a different result. Without a verifiable capture method, you cannot assess whether each file is complete, whether it includes injected reminders or hidden tool definitions, or whether it reflects a specific model version or a regional variant. The presence of separate files for 'Claude Opus 4.6' and 'Claude Opus 4.6 No tools' indicates that the maintainer distinguishes between configurations, but the reader must trust that distinction. For a security researcher, this is a red flag: you cannot verify authenticity without a reproducible method. For a journalist, it is a caveat to state in any article. The repository's value rests on the assumption that the captures are genuine, and that assumption is unverified.

Organized by Vendor, Model, and Component

The directory structure mirrors the product landscape. Anthropic prompts live under 'Anthropic', with a claude-code subfolder and an artifact-types folder. OpenAI prompts are under 'OpenAI'. Google, xAI, Meta, Kimi, and Perplexity each have their own top-level folders. Within each, files are named by model or product, such as 'claude-opus-5.md' or 'claude-sonnet-5.5.md'. This layout makes it easy to find a specific prompt, but it also means that versioning is implicit in the filename. If a model is updated, the maintainer must add a new file or overwrite the old one. The 'Most recent additions/changes' table tracks new and updated entries with dates (for example 'Kagi Assistant (updated)'), but only for the latest changes; for older files, check the commit history on GitHub.

Getting the Files: Git Clone and Browse

There is no installation step because there is no code to run. The repository is a set of Markdown files. To use it, you clone the repository or browse it directly on GitHub. A standard clone command is 'git clone https://github.com/asgeirtj/system_prompts_leaks.git'. Once cloned, you can read any file with a text editor or use grep to search across all prompts. For example, to find every mention of 'copyright' across all files, you would run 'grep -ri copyright .' from the repository root. That command works because the files are plain text. There are no configuration keys, no build steps, and no dependencies. The primary language listed is JavaScript, but that likely refers to the repository's metadata or a minor script, not to the content. This simplicity is both a strength and a limitation: anyone can read the files, but there is no built-in way to compare versions or track changes over time.

Licensing: CC0-1.0 and Its Implications

The repository is licensed under CC0-1.0, which is a public domain dedication. That means you can copy, modify, and redistribute the prompt files without attribution, and you can use them for commercial purposes. This is a significant advantage over the original prompts, which are likely protected by the AI vendors' terms of service. However, CC0 does not give you any rights to the underlying models or the products themselves. You can use the text of a prompt to inform your own system design, but you cannot use it to claim affiliation with Anthropic or OpenAI. The license also means there is no warranty. If a prompt contains a factual error or a harmful instruction, the maintainer is not liable. For a developer who wants to build a competitor or a research project, CC0 removes legal friction, but it does not remove the risk of violating the vendors' terms when you obtained the prompt in the first place.

A Genuine Limitation: Staleness and Accuracy

The most serious limitation is that prompts change frequently, and this repository is only as current as its last push. The README's update table shows entries from October and November 2026, but it also lists older models like Claude Opus 4.6 and Opus 4.7. If you are using a prompt to replicate behavior for a current product, you must check whether the file matches the version you are targeting. There is also no guarantee that a prompt is complete. The README describes the prompts as complete and verbatim, and some entries come with tools and skills as separate files, but you cannot verify from the README alone that a given file is whole. A prompt that is cut off mid-sentence could mislead a reader about the full instruction set. For anyone relying on these prompts for security research or prompt injection testing, this staleness is a real risk. You cannot assume a file from March 2026 reflects the current model behavior.

The Alternative: Official Documentation and Your Own Captures

The direct alternative is to rely on official documentation. Vendor documentation is the official alternative, though the repository's existence suggests that official coverage is incomplete. Another alternative is to capture prompts yourself using API access or a controlled environment. That approach gives you verifiable, timestamped data, but it requires technical effort and may violate terms of service. The key difference is control: this repository offers convenience and breadth, while your own captures offer accuracy and provenance. For a one-off research question, the repository may be sufficient. For a production system that needs to track prompt evolution, you would need a custom pipeline that records prompts over time. The repository does not offer that; it is a snapshot archive, not a monitoring tool.

Editorial conclusion

Adopt this repository if you need a central, openly licensed reference for AI system prompts, whether for research, journalism, or building your own agent prompts. Skip it if you require guaranteed accuracy or production-ready documentation, since captures are unofficial and may contain errors. Before relying on any file, verify it against official sources, and check the repository's recent commit history to confirm the version you need is present. The real value is in the raw text and the update cadence; treat each file as a snapshot, not a specification.

Frequently asked questions

What does "prompt leakage" mean?

In this context it means obtaining a chatbot's hidden operating instructions, word for word, without the vendor publishing them. system_prompts_leaks exists because contributors extracted that text from Anthropic, OpenAI, Google, xAI and a dozen other vendors.

Is it true that Claude Code's entire system prompt has been leaked?

The repository's README lists several captures of Claude Code's prompt as current, including builds for Sonnet 5.5, Opus 5.5, and three separate Fable 5.1 variants for the web, desktop and headless clients, each stored as its own file such as Anthropic/claude-code/claude-code-sonnet-5.5.md.

What are system prompts?

They are the instructions a vendor loads before a conversation starts, separate from anything the user types. This repository treats web, desktop and headless builds of the same model as different system prompts when a vendor ships them as different files, and keeps injected mid-conversation reminders in a separate document, anthropic_reminders.md, rather than folding them into the same file.

Official sources

  1. Official README
  2. Project repository
  3. README
  4. Releases
Add this badge to your README

If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.

Add this badge to your README

markdown
[![Hysen Labs](https://hysenlabs.com/badge/asgeirtj-system-prompts-leaks.svg)](https://hysenlabs.com/projects/asgeirtj-system-prompts-leaks)

Star history chart for your README

markdown
[![Star history](https://hysenlabs.com/chart/asgeirtj-system-prompts-leaks.svg)](https://hysenlabs.com/projects/asgeirtj-system-prompts-leaks)

Was this useful?