Conversation Design

Conversation design is the work of deciding how an agent talks: how it opens, how it asks for a phone number, what it says when it does not know, how it confirms a booking, when it stops talking and listens. It is not prompt writing alone, though the prompt is where most of it ends up. It covers the flows, the wording, the error paths and the handoff, and it is the main reason two agents on the same model with the same knowledge feel completely different to a customer.

How it works. Design starts from the top intents and the data each needs. For each, the designer writes the happy path (greeting, question, confirmation, close) and the exception paths: the caller gives a date in the past, the customer asks two things at once, the system is down, the person wants a human. Wording rules set tone and length: short turns on voice, one question at a time, confirmations that repeat the key fact ("Thursday the 12th at 10:30, is that right?"). The design becomes a system prompt, a set of few-shot examples, tool descriptions and canned phrases for errors. It is then tested on transcripts of real calls and chats, revised, and measured after launch on completion rate and where conversations break.

Why it matters for booking and support rates. Most failed conversations fail on design, not on the model: a question that asks for three things at once, a confirmation that is skipped, a long paragraph read aloud on a call, an agent that argues instead of handing off. Each of these costs a booking or a resolution. Good design also controls cost, because shorter, clearer turns mean fewer replies per conversation and fewer minutes per call.

Example. A spa books facials through a voice agent. The first design asked "what service would you like, what day works, and do you have a preferred therapist?" and callers answered one of the three. The redesign asks one question at a time, offers the two nearest slots rather than an open date, confirms the booking in a single sentence, and ends with the cancellation policy in under ten words. Calls get shorter, the booking completion rate rises, and the number of calls that end with "let me transfer you" falls because the common cases now have a clear path.

Conversation design vs system prompt. The system prompt is the artifact: the instructions the model reads on every turn. Conversation design is the thinking that produces it, plus the flows, examples and tests that live outside the prompt. A prompt without a design is a guess; a design without a good prompt is a document.

Common pitfalls. Writing for chat and deploying on voice without shortening. No error path for silence or background noise. Personality before clarity. Open questions where a choice would do. Skipping confirmation on anything with a date or a number. Not reading transcripts after launch.

Botcadence chat agents are grounded in uploaded documents and go live on a website in about 5 minutes, with the design revised from real transcripts in conversation analytics.