Skip to slide
080 / 221

CH 06 · Prompt Architecture · 02/10

Read as one page

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.).

Finished this idea?

Your learning progress stays in this browser.