System Prompt

A special set of instructions provided to an LLM at the very beginning of a conversation to dictate its persona, rules, and constraints.

How it works

In a multi-turn conversation, the context sent to an LLM typically begins with a system message, followed by alternating user and assistant turns. The system prompt is injected before any user content and is processed by the model first. It can specify the model's role ('You are a helpful legal research assistant'), constraints ('Never discuss competitor products'), output format ('Always respond in JSON'), and safety rules. Models are trained to give the system prompt elevated authority over user instructions.

Why it matters

The system prompt is the primary mechanism through which product teams customise LLM behaviour without fine-tuning. It is how ChatGPT behaves like ChatGPT rather than a generic model, how Copilot maintains its coding-focused persona, and how enterprise applications enforce business-specific rules. Understanding system prompt design is critical for AI product builders, and the security of system prompts against prompt injection attacks — where user inputs attempt to override system instructions — is a major focus of LLM security research.