Exploring System Prompts of Leading AIs (ChatGPT, Claude, Gemini, Grok, Perplexity)

GitHub’s system_prompts_leaks project, now with over 40 000 stars, aggregates the system prompts of major AI models such as ChatGPT, Claude, Gemini, Grok and Perplexity. The prompts are organized by vendor (OpenAI, Anthropic, Google, xAI) and span scenarios like chat interaction, coding assistance, financial tools, and multi‑agent collaboration, continuously updated to include the latest releases (GPT‑5.5, Claude Opus 4.7). The repository is released under the MIT license, providing a free, one‑stop learning resource for prompt engineering.

GitHub: https://github.com/asgeirtj/system_prompts_leaks

OpenAI’s design separates personality from utility. Four distinct personas—cynical, geek, zero‑emotion robot, and quiet listener—appear in casual conversations, while formal outputs (emails, code, resumes) are forced into a neutral tone, creating a clear wall between entertainment and professional use.

Anthropic emphasizes engineering rigor. The Claude Code prompt reads like an engineer onboarding manual, specifying Git security policies, sub‑agent scheduling, and permission‑level rules, and explicitly forbids unauthorized destructive actions. Claude for Excel integrates Bloomberg, FactSet, and other financial terminals, mandates that all numeric calculations be expressed as spreadsheet formulas, and supports cross‑software agents for Word and PowerPoint, maximizing suitability for professional finance scenarios.

xAI’s Grok adopts an extreme role‑based strategy, offering six complete roles such as talk‑show host, friend, learning partner, and therapist. The talk‑show role focuses on absurd humor, the learning‑partner role suppresses redundant explanations, and the therapist role follows professional psychological theory; each role is bound by strict safety boundaries, forming a comprehensive AI role‑play platform.

The common insight across these top‑tier prompts is that a System Prompt functions as the full product design, covering six core modules: personality definition, safety protocols, tool invocation, permission management, memory storage, and multi‑agent coordination. This validates the industry view that “everything is prompt”; swapping the prompt layer can transform the same underlying model into entirely different products, making prompt design a decisive factor for user experience.

Beyond merely exposing prompts, the open‑source collection systematizes scattered industry best practices. For ordinary users, understanding the prompt logic enables more efficient use of AI tools; for AI practitioners, the repository offers concrete examples to refine custom agents and adopt leading‑vendor strategies.