A self-serve Terri is only as good as the instructions you give it. Write the escalation rule first, then the routine answers, then test from an outside line.
Write the cases that must reach a person immediately — emergencies, distress, regulated questions, anything where a wrong answer creates liability. Test that path from an outside line before the agent takes a real customer call.
Add hours, services, prices you are willing to state, who you do not serve, and the exact wording for common questions. Keep it in the system prompt — that is the standing instruction field on the agent identity.
Put caller-specific background on contacts, not in the prompt. Import CSV or JSON so a known number arrives with a name and notes. Leave agentMemoryEnabled off unless you also set callerMemoryMode to opt_in on that identity.
Set handoffPhoneNumbers to the verified numbers Terri should conference in. Narrow allowedOutboundPrefixes to the destinations you actually want called. Leave recordingMode at metadata unless you have a reason and a consent rule for audio.
Call the number yourself. Ask something the instructions cover, something they do not, and something that should transfer. Read the transcript. Fix the instruction that failed. Repeat until the three cases are clean.
A tenant can have a default identity plus additional identities bound to specific numbers. Use a second identity when two lines should sound or escalate differently — a main line and an after-hours line, for example.
Leave voice blank to use the platform voice. Set it only when a specific OpenAI Realtime voice is required for that identity.
Instructions that try to cover every edge case become a novel nobody can audit. Prefer short rules you can test. When the transcript shows a miss, add one line, not a paragraph.
Set the agent up yourself, or have us design the inbound agent a call center will run. Neither path is a fallback.