Leaked AI System Prompt Aggregator Scraper

Aggregates the public GitHub collections of leaked and published AI system prompts — jujumilk3/leaked-system-prompts, x1xhlol/system-prompts-and-models-of-ai-tools, and 0xeb/TheBigPromptLibrary — into one normalized, deduplicated dataset. Returns product name, vendor, full prompt text, version, and leak date for hundreds of prompts spanning ChatGPT, Claude, Cursor, Devin, v0, Perplexity, and dozens of other AI products.

---

Leaked AI System Prompt Aggregator Features

• Unions three separately-maintained GitHub collections into one schema
• Deduplicates across repos by content hash — the same leaked prompt posted in two collections shows up once
• Extracts product, vendor, version, and leak date from each source repo's own filename conventions
• Classifies each record into a coarse category (chat assistant, coding agent, search, image)
• Pure API scraping against GitHub's own tree and raw-content endpoints — no browser, no proxy
• Select which of the three source repos to pull from per run, or leave all three on for full coverage

---

Who Uses Leaked System Prompt Data?

• Prompt engineers — study real, shipped guardrail and tool-use prompts instead of guessing at structure
• AI safety researchers — compare how different vendors constrain the same class of model, or at least the parts they were willing to leave in a system prompt
• Product teams — benchmark their own system prompt against what competitors ship
• Security researchers — track how prompt-injection defenses have evolved across product versions

---

How the Aggregator Works

1. For each selected source repo, the actor fetches GitHub's full file tree in a single API call.
2. Each repo's file list is filtered down to actual prompt files using that repo's own naming conventions — the noise (READMEs, images, unrelated custom-instruction dumps) gets dropped before anything is fetched.
3. Matching files are fetched as raw text and hashed. A file whose content already appeared earlier in the run — the same prompt cross-posted to another collection — is skipped, not saved twice.
4. Each unique prompt is parsed into product, vendor, version, and leak date, then written to the dataset.

---

Input

{
    "maxItems": 15,
    "sources": [
        "jujumilk3/leaked-system-prompts",
        "x1xhlol/system-prompts-and-models-of-ai-tools",
        "0xeb/TheBigPromptLibrary"
    ]
}

Field	Type	Default	Description
maxItems	integer	15	Maximum number of leaked/published system prompt records to emit for this run.
sources	array	all three	Which of the three source GitHub collections to aggregate. Leave all three selected for full union coverage.

To pull only from one collection:

{
    "maxItems": 25,
    "sources": ["x1xhlol/system-prompts-and-models-of-ai-tools"]
}

---

Leaked AI System Prompt Aggregator Output Fields

{
    "product_name": "Claude 2.1",
    "vendor": "Anthropic",
    "prompt_text": "The assistant is Claude, created by Anthropic...",
    "prompt_version": "anthropic-claude_2.1",
    "leak_date": "2024-03-06",
    "source_repo": "jujumilk3/leaked-system-prompts",
    "source_file_path": "anthropic-claude_2.1_20240306.md",
    "source_url": "https://github.com/jujumilk3/leaked-system-prompts/blob/main/anthropic-claude_2.1_20240306.md",
    "char_count": 342,
    "model_family": "",
    "category": "chat_assistant",
    "dedup_hash": "8f08589aa2fe9dc5c571151d7875ab5cee53331a0034e731439a9b1677612739",
    "scraped_at": "2026-07-13T09:39:00.576Z"
}

Field	Type	Description
product_name	string	Name of the AI product/tool the leaked prompt belongs to (e.g. Cursor, Claude Code, Devin)
vendor	string	Company/organization that publishes the product (e.g. Anthropic, OpenAI, Cursor)
prompt_text	string	Full text of the leaked/published system prompt
prompt_version	string	Dated/version label for this prompt where given by the source
leak_date	string	Date this version was leaked/published, YYYY-MM-DD, where derivable from the source filename
source_repo	string	GitHub repo (owner/name) this record was aggregated from
source_file_path	string	Path of the source file within the GitHub repo
source_url	string	GitHub blob URL for the source file
char_count	integer	Character length of prompt_text
model_family	string	Underlying model family where labeled in the source; empty if not labeled
category	string	Coarse product category: chat_assistant, coding_agent, search, image, or other
dedup_hash	string	SHA-256 hex digest of the normalized prompt text, for cross-repo/cross-run dedup
scraped_at	string	ISO-8601 timestamp when this record was scraped

---

FAQ

How do I scrape leaked AI system prompts?

Run this actor with maxItems set to however many records you want. It pulls from three public GitHub collections and hands back a deduplicated dataset — no manual repo-cloning or file-merging required.

What data can I get from these leaked system prompt collections?

Product name, vendor, full prompt text, version label, leak date, and the exact source file each prompt came from — enough to trace any record back to its origin repo and commit history.

Does this actor need proxies?

No. It reads GitHub's tree API and raw.githubusercontent.com, both of which are public and generous with unauthenticated traffic at this actor's scale.

Can I filter by source collection?

Yes. Set sources to any subset of the three supported repos. Leave it at the default to pull from all three.

How much does this actor cost to run?

Standard per-record pricing, no CAPTCHA-solving or residential-proxy surcharge — this is a plain GitHub API read.

---

Need More Features?

Need a fourth source repo added, or different category classification? File an issue or get in touch.

Why Use the Leaked AI System Prompt Aggregator?

• One dataset instead of three repos — no more cloning and hand-merging separately-maintained collections
• Deduplicated by content, not filename — a prompt that got cross-posted to two collections shows up once, with the hash to prove it
• Fully passive — reads only already-public GitHub repositories; adds no new disclosure