What should go into an agent's system prompt?

30-second answerSay your answer out loud first, then reveal.

Typical structure

Role and objective: "You are a support agent for Acme. Your goal is to resolve order issues end to end."
Context: company policies, the user's account tier, current date, environment details.
Tool guidance: which tools to prefer, in what order, and known caveats ("search_orders is slow; call it once and reuse results").
Constraints and policies: "Never issue refunds above ₹5,000 without approval." "Don't reveal internal notes."
Escalation: when to ask clarifying questions or hand off to a human.
Stopping criteria: what counts as done.
Output format: structure, tone, length.

Explain the why. "Keep replies short because users are on mobile" generalises better than "max 50 words."
Give examples of tricky cases, but diverse ones. A single example tends to get copied verbatim.
Use clear sections (headers or XML tags) so the model can find policies easily.
Keep dynamic data separate (user info, date) from static instructions. This also helps prompt caching (Q28).

Giant lists of ALL-CAPS rules that conflict with each other.
Putting security entirely in the prompt. Prompts are guidance, not enforcement. Real limits belong in code (Q39).

This is what real progress feels like.