Foundations of Conversation Design

A good conversation starts from an outcome rather than a flow diagram: how should this call end? Designed backwards from that outcome, a call flows naturally. Built forwards on the logic of "ask this, then ask that", it usually feels robotic, because it tries to follow a written sequence instead of adapting to what the customer says.

Three basic principles:

One piece of information at a time. "Could you give me your name, national ID number and date of birth?" is fine on a written form and exhausting on a call. The customer has to hold all three in mind and deliver them in order. Ask for them one at a time, at a natural rhythm.

A clarifying question beats an assumption. When the model meets ambiguous information, for instance a name it did not hear clearly, it should ask a short clarifying question rather than guess and carry on. That adds a second or two to the call and is far safer than proceeding on a wrong record.

What is left unsaid is part of the design too. Unnecessary verification steps, overly formal language, filler such as "understood, now...". All of it lengthens the call. Good conversation design is more about removing than adding.

CBOT's Prompt Manager keeps this from being a one off writing exercise: every persona is held under version control, and changes go live only after they have been tested and evaluated. The conversation becomes something that is continuously improved from real calls rather than written well once and left alone.