System prompt
A system prompt is the instruction text that tells an AI agent its role, tone, limits and rules before a conversation begins.
A system prompt is the standing instruction an agent built on a large language model follows in every conversation. The caller or the person writing in never sees it, yet it decides who the agent is, on whose behalf it speaks, what it helps with and where it stops. For a voice agent, the system prompt covers how to greet callers, how to address them, what information to collect, when to use tools and when to hand over to a person. Decisions made in conversation design usually reach the agent through this text.
How it works
On each turn, the model reads the system prompt, the conversation so far and any relevant passages retrieved from the knowledge base, and produces its reply from all of them. The clearer and more consistent the prompt, the more predictable the agent's behavior. Contradictory rules, overly long prompts and heavy use of "always" and "never" can lead the model to skip some instructions.
Why it matters
A good system prompt keeps the agent within scope and reduces guessing on topics it doesn't know (hallucination). A practical structure: role and goal, tone, step-by-step flow, information to collect, things not to do, handover conditions and example replies. Information that changes often, such as prices, opening hours and promotions, belongs in the knowledge base rather than the prompt. Every change to the prompt should be re-run against the same test scenarios before it goes live.
In Speakline
In Speakline, you write each voice agent's instructions in the agent settings and can start from ready-made templates such as receptionist, sales and support. In campaigns, you can add the purpose of the call or replace the instructions entirely, so the same agent can serve different goals. See Voice agents for details.