The system prompt is a contract

Think of the system prompt as a contract between you and the model: "behave like this, follow these rules, emit output in these formats." It's the one piece of context present on every call, so it carries the agent's invariant behaviour: the things true regardless of the specific task. Keep task-specific instructions out of it (load those on demand; Chapter 5) so it stays focused on the universal.

A good system prompt typically covers:

• Identity and scope: who the agent is, who it serves, what it does. One or two sentences set the persona.
• Output protocols: any machine-readable formats the model must produce (see below).
• Behavioural rules: how to do the core tasks well, including domain consequences.
• Guardrails: what never to do (fabricate, leak internal identifiers, etc.).