What makes a receptionist book instead of just answer?

A narrow spec: define the caller intents you will handle, the knowledge you will cite, the calendar you will write to, and the transfer rule when you won’t. VoiceForge encodes that as a Zod-validated Agent Spec JSON — goals, flow, tools, compliance — so the booking path is reviewable before it dials.

Can I see a prompt template?

Yes — the guide includes a full greeting, fallback, booking and no-show SMS template. The key is the binding: knowledge is agent-scoped, tools are permissioned, and the spec version you test is the version you publish.

How do I handle interruptions and silence?

Both directions matter: barge-in (caller talks over agent) needs voice-activity cancellation; premature end-of-turn (caller pauses mid-thought) needs silence timing. Test both in a browser call before telephony.

Builder’s guide

Build a receptionist that books — not one that just answers

A receptionist that books is a narrow, testable spec: greeting, intent graph, calendar write, fallback to human and SMS follow-up — all versioned as an Agent Spec, not scattered prompts.

Start building See how it works

Start with the spec, not the prompt

VoiceForge generates a Zod-validated Agent Spec JSON from a natural-language brief: goals, voice persona, call flow, tools, compliance policy and analytics. That spec is the contract; prompts are derived from it, not the source of truth.

Greeting, turn-taking and fallback

Write a 7-second greeting that states who you are, what you can do and how to reach a human. Handle barge-in (cancel downstream work) and premature end-of-turn (wait for the caller finding an order number). Define fallback: “I’ll connect you” — then transfer with full context.

Calendar intent that doesn’t double-book

Check live availability before offering a time. Validate the write, enforce idempotency and log the booking outcome. VoiceForge validates tool calls, checks permissions per client and records every write.

Prompt + script template

Greeting → intent capture → knowledge answer → calendar check → confirmation SMS. Keep knowledge agent-scoped, cite only approved sources and disclose AI handling where required.

Test in the browser, then publish

Run the complete call: transcript, event stream, tool activity and outcome. Test interruption, silence, double-question, transfer-during-load and the “caller refuses every option” path before telephony is attached.

Frequently asked questions

What makes a receptionist book instead of just answer?: A narrow spec: define the caller intents you will handle, the knowledge you will cite, the calendar you will write to, and the transfer rule when you won’t. VoiceForge encodes that as a Zod-validated Agent Spec JSON — goals, flow, tools, compliance — so the booking path is reviewable before it dials.
Can I see a prompt template?: Yes — the guide includes a full greeting, fallback, booking and no-show SMS template. The key is the binding: knowledge is agent-scoped, tools are permissioned, and the spec version you test is the version you publish.
How do I handle interruptions and silence?: Both directions matter: barge-in (caller talks over agent) needs voice-activity cancellation; premature end-of-turn (caller pauses mid-thought) needs silence timing. Test both in a browser call before telephony.