Designing for voice

Getting a call working is a week. Getting one people use twice is the rest of the work, and almost none of it is about the API. Speech has no scrollback. A caller can interrupt you mid-word. The screen is right there and it holds detail the ear cannot. These pages are the arguments, written from running production voice agents rather than from first principles.

The two channels

Voice points. The screen holds. Speech is linear, transient and interruptible — it is good at direction, acknowledgement and one number at a time. The screen is durable and scannable — it is good at lists, totals, forms and anything the caller would otherwise have to remember. A number the caller must hold in their head belongs on the screen.

The mechanism for the second channel is actions; this section is about what to send and when.

The pages

Voice points, the screen holds	What belongs in the ear and what belongs on the page.
The turn budget	How long a unit may be, and what the caller does when it is longer.
Interruption and heard truth	Barge-in, and why history holds what was heard.
Misunderstanding and reversal	Recovering when the agent got it wrong, without starting over.
Parallel workstreams	Work that outlives the turn that started it.
Prompt design for voice	A prompt written for an ear, not a text box.
Tool design for voice	Tools whose latency the caller can hear.

When you are done here

You have an application worth putting in front of people. Keeping it working is Operate.