System Prompts & Personality

System Prompts & Personality

The system prompt is the single highest-leverage setting on your agent. It determines what the agent asks, what it refuses to do, and what it captures. This guide covers writing one that holds up on real calls.

Where to Edit It

Dashboard → AI Agent Settings → System Prompt. Industry templates pre-fill this with a working starting point; editing from a template is almost always better than writing from scratch.

Structure That Works

Effective prompts follow a consistent shape:

1. Identity — who the agent is and which business it represents
2. Objective — the one thing the call should achieve
3. Required information — the explicit list of fields to collect
4. Boundaries — what the agent must never do
5. Escalation — when to transfer to a human

Be Explicit About Required Fields

The most common failure is an empty field in the lead record. Models do not reliably infer what you need. Compare:

Vague: "Collect the caller's details."

Explicit: "Collect, in this order: full name, callback number, service address including postcode, the nature of the problem, and whether this is an emergency. Do not end the call until you have the callback number."

Set Hard Boundaries

State what the agent must not do, in plain language:

• "Never quote a price. Say a team member will confirm pricing."
• "Never promise a specific arrival time. Offer a window and confirm it will be verified."
• "Never give medical, legal, or financial advice."
• "If the caller says the words emergency, gas, flooding, or fire, transfer immediately."

Handle the Awkward Cases

Real calls are messier than test calls. Add explicit handling for:

• Existing customers — different questions than new enquiries
• Sales and spam calls — end politely and do not create a lead
• Wrong numbers — confirm and close quickly
• Callers who will not answer questions — capture the number and escalate

Iterate From Transcripts, Not From Imagination

The best prompt improvements come from reading real transcripts in Call Logs. When a call goes wrong, find the exact turn where it went wrong and add one instruction addressing it. Change one thing at a time.