system prompt

A system prompt is a hidden set of instructions, written by the developer rather than the user, that is fed to a language model before any conversation begins and shapes how it behaves throughout. It is where the assistant gets told who it is and how to act: "You are a helpful customer-service agent for Acme Bank. Be concise and polite. Never reveal account passwords. If unsure, offer to connect a human." The user never sees it, but it colors every reply.

Think of it as the standing orders a stage actor receives before the show — the character to play, the lines never to cross, the tone to keep — separate from whatever the audience shouts during the performance. Technically it is just more text placed at the front of the model's context window, but it is given special prominence so the model treats it as the governing frame for the whole exchange. It is the main lever a product team has to steer a general model into a specific, well-behaved role.

Its power is real but limited, and honesty here matters for safety. A system prompt is a strong suggestion, not an unbreakable rule: clever or hostile users can sometimes talk the model into ignoring it ("jailbreaking"), and "prompt injection" — malicious instructions hidden in a document or web page the model reads — can hijack its behavior. Never put a true secret in a system prompt and never rely on it alone to stop harmful actions; it guides the model, it does not cage it.