What AI Chatbots Are Told Before You Type: A Developer's Guide to the system_prompts

Every time you launch an AI chatbot like ChatGPT, Claude, or Gemini, the model has first read a comprehensive set of instructions known as the system prompt. This hidden set of instructions shapes the assistant's behavior, including its tone, formatting habits, tool preferences, and refusal to perform certain actions. Developers can access these system prompts in full detail through the GitHub repository asgeirtj/system_prompts_leaks.

The concept of a system prompt is straightforward - it sets the rules for a chat request, with the user's input being the actual conversation. For example, in the OpenAI-style format, the system prompt might declare: "You are a support bot for Acme Inc. Answer only questions about Acme products. Keep answers under 100 words." This invisible message guides the entire response.

The repository is a vast collection of these system prompts, organized by company such as Anthropic, OpenAI, Google, and others, with each company's prompts stored in separate folders. The prompts can be thousands of words long, covering various aspects from personality and formatting rules to safety policies and product-specific behaviors.

The collected prompts offer valuable insights for developers. They serve as a masterclass in production prompt engineering, showcasing how large-scale teams structure instructions for millions of users. Key takeaways include structuring prompts with tags or headers, explaining the rationale behind rules, using examples to pin down behavior, and explicitly addressing edge cases.

The prompts also demonstrate how tools and agents are integrated, offering a mental model of modern AI agent design. Lastly, they help explain odd model behaviors, such as an assistant's refusal to use bullet points or insistence on searching for information repeatedly.