System prompt
A system prompt is the instruction set that defines an AI agent's role, priorities, boundaries, and operating rules before a caller's conversation begins. It guides how the agent interprets requests, speaks, uses tools, handles uncertainty, and decides when to transfer or stop.
For an AI phone agent, the system prompt often describes the business, the jobs the agent may perform, the information it must collect, and the conditions attached to an action. It can specify tone and phrasing, but its more important function is operational: make the agent's permitted behavior clear enough to produce consistent call outcomes.
A useful prompt separates durable instructions from changing facts. Rules such as “confirm the date before booking” belong in the prompt. Current hours, inventory, or account details are better retrieved from an authoritative source when needed. Embedding frequently changing information in a long prompt makes updates harder and increases the chance that stale text will influence a call.
Specific instructions generally work better than broad wishes. “Be helpful” does not explain what to do when no appointment is available. A stronger instruction defines the next steps: offer valid alternatives, ask whether another day works, and provide a human handoff when the caller cannot use them. Examples can clarify edge cases, but too many overlapping examples can obscure the underlying rule.
The system prompt is not a security boundary by itself. Applications should validate tool inputs, restrict what actions are available, and enforce authorization outside the model. A sentence telling an agent not to expose sensitive data does not replace access controls or careful data handling.
Prompt design should account for spoken interaction. Responses need to be easy to follow without a screen, questions should usually request one manageable piece of information, and confirmation should be proportional to risk. The prompt should also cover interruptions, corrections, silence, unsupported requests, and requests for a person.
Treat the prompt as versioned operating logic, not finished marketing copy. Test it with complete scenarios, inspect both the conversation and any resulting actions, and revise unclear rules. A good prompt is concise enough to navigate, specific enough to evaluate, and paired with system-level controls for anything that must always be enforced.