What belongs in a system prompt

A system prompt for an agent is doing a different job from one for a single call. A single call needs a task. An agent needs a standing brief it can consult on turn twelve, when the original request is far behind it.

Four things earn their place. Who it is and what it is for, in a sentence. What it may and may not do, stated as boundaries rather than hopes. How to behave when things are unclear — ask, assume and say so, or stop. And the shape of a good answer, so the last turn does not have to be guessed.

You are the support triage agent for Meridian Tools. You investigate a
ticket and prepare a reply for a human agent to send.

You may: read orders, shipments, invoices and the public docs.
You may not: change any record, contact a customer, or issue a refund.
When the answer needs one of those, say so and stop.

Ground every factual claim in something a tool returned. If the systems
disagree, report both rather than choosing. If you cannot find something
after two attempts with different search terms, say what you tried.

Finish with: the customer question in one line, what you found with the
order or shipment id it came from, and a suggested reply.

Notice what is absent. No exhaustive list of edge cases, no lengthy persona, no restatement of what each tool does — the tool descriptions carry that. Long system prompts written as a growing pile of "and also, never..." rules tend to work less well than short ones, because a rule buried at line ninety competes with everything above it.