System prompt

A workable structure covers role and scope, the specific task, style constraints, escalation rules, and explicit boundaries. Style constraints matter unusually much in voice: without an instruction to keep responses to a sentence or two, models produce paragraphs that take an uncomfortably long time to hear, and the caller starts talking over the agent.

Formatting instructions need to be voice-aware because the output is synthesised, not rendered. Markdown, bullet characters and emoji either get read aloud or silently dropped. Numbers, dates and currency should be written the way they are to be spoken. Long enumerations should be broken into two or three options at a time with a check-in, because a caller cannot rewind.

Prompts should state what the agent must not do as explicitly as what it should. Do not quote prices that are not in the provided data. Do not confirm anything a function has not returned. Transfer to a human when the caller asks, when they become distressed, or when the request falls outside scope. These negative instructions are what keep a plausible-sounding model from inventing commitments.

Treat the prompt as versioned configuration rather than a text box. It changes agent behaviour globally and immediately, so changes deserve review, a record of who made them and why, and the ability to roll back — the same discipline applied to any other production configuration.