Guide
Your first assistant in 10 minutes
An assistant that can take a call: transcriber, model, voice, prompt, and a published version.
10 minutes 6 steps
A published assistant with a saved version, ready to be attached to a phone number.
- A console account with write access to assistants.
- At least one valid model credential and one voice credential, added under Platform → Connections.
- A written idea of what the assistant should say and what it must not say.
Open Build → Assistants and press Create
The form is a five-step wizard. You can go back at any point; nothing is saved until the last step, so closing the tab does not leave a half-finished draft behind.
/{your-organization}/assistants/create
Identity: name, language, maybe a template
Language is required and arrives set to English. Change it if the assistant speaks another one, because the transcription choices follow from it. A template fills in the prompt and the first message, and you can rewrite both immediately.
Name · Mode (defaults to “pipeline”) · Language · Start from a template
Conversation: the first message and the system prompt
The system prompt is the only required field here. Write what the assistant does, what it is not allowed to do, and what it asks when it does not know. If you choose to have it speak first, the field for that opening line appears.
System prompt — required, up to 100,000 characters. Accepts {{variables}}, {% if %} and {% for %}.
Voice and model: press a preset, then adjust
A preset sets all three stages at once, as a ranking rule over the provider registry rather than three hard-coded model names. Every card stays editable afterwards. The moment you change one by hand the assistant matches no preset, and the interface says so instead of leaving a preset highlighted.
Three cards: Transcription · Model · Voice. Each shows its latency, cost per minute and a quality index.
Behaviour: pauses, interruptions, limits
This is where the turn-taking settings live. Leave them alone on your first assistant. The one worth changing is the maximum call duration, if you know your calls run past ten minutes.
Silence timeout 30 s · Maximum call duration 600 s · Words before interrupting
Validate, then publish
Publishing creates a version, and calls in flight keep the version they started on. If the validator finds problems it gives you the whole list of reasons rather than just the first one.
“Validate configuration” runs the validator without publishing. “Publish” asks for a version label and a change summary.
How to tell it worked
- The assistant page shows a published version, not just a draft.
- The validate button comes back with nothing to report.
- Prompt preview renders the text with variables substituted and no template errors.
What does not work yet
Every guide has this section. If it were missing, it would mean we had not looked hard enough.
- The validator refuses to publish a transcriber that only works on finished recordings — Sonix, for one — as the live transcriber. The refusal is deliberate: that mistake takes every call down at once and surfaces three layers away.
- A published assistant still does not answer the phone. It needs a number, which is the next guide.