Uncovering Hidden System Prompts of Major AI Models

Project Overview

system_prompts_leaks is a public GitHub repository that collects the system prompts (pre‑set instructions) used by major conversational and coding models such as OpenAI GPT‑5.4, GPT‑5.3, Codex; Anthropic Claude Opus 4.6, Sonnet 4.6; Google Gemini series; xAI Grok; Perplexity, etc.

Why the repository matters

Transparency – the prompts expose the “factory settings” that determine model behavior, turning the black‑box perception of AI into a viewable artifact.

Reference for prompt engineering – developers can study concrete prompt wording from leading providers.

Research baseline – the collection enables side‑by‑side comparison of safety strategies, creativity emphasis, and alignment philosophies across vendors.

Content structure

Each model version has its own Markdown file; separate files exist for different modes (e.g., “thinking mode”, “code mode”, “no‑tool mode”).

ChatGPT personalities (friendly, professional, quirky) are documented as distinct prompt files.

Additional files contain system instructions for auxiliary tools (web search, deep research, Python execution, image generation) and policy files (image‑safety, automated‑context handling).

Example comparison

Opening the Claude Opus 4.6 file shows Anthropic’s wording that directs the model to “stay humble, helpful, and avoid harmful content.” A parallel ChatGPT file emphasizes “usefulness” while embedding safety constraints, illustrating how vendors balance utility versus risk.

Getting started

Search for asgeirtj/system_prompts_leaks on GitHub or locate it on the trending page.

Read the repository README, which lists vendors and model identifiers with links to the corresponding Markdown files.

Select a file (e.g., Claude_Opus_4.6.md) to view the exact system prompt text.

Advanced use: benchmark a custom model against these prompts or reuse the prompt structures to shape product‑level AI interactions.

Intended audience

AI application developers seeking concrete prompt‑engineering examples.

Product managers and researchers who need comparative data on safety boundaries and UX design.

Technical enthusiasts learning how a single sentence can steer model behavior.

Security and ethics researchers analyzing bias, safety, privacy clauses, and potential prompt‑injection vectors.

Community impact

Within weeks the repository accumulated over 37 000 stars, with daily star growth exceeding 2 000, indicating strong demand for model‑level transparency. The visibility also sparked discussions about intellectual‑property rights, system security, and prompt‑injection attacks.

Visualization