Conversation design · Explore this field ↗ · Design · 2 min read
Turn-taking is the hidden interface
Fast answers matter less than knowing when to listen, wait, interrupt, and stop.
A turn has phases
A conversational turn includes capture, interpretation, planning, response, and acknowledgement. Users need cues for each phase. In text, typing indicators and progressive results can help, but false animation erodes trust. In voice, end-of-speech detection, barge-in, and audio cancellation are core controls. In an embodied character, gaze and posture add another timing channel that can agree with—or contradict—the words.
Design for interruption
Interruption is not an edge case. A user may correct a number, stop an irrelevant explanation, or react to the first half of an answer. Cancel speech immediately, preserve what was heard, and decide whether the new utterance edits the current task or opens another. For consequential actions, interruption should pause execution until intent is clear.
Short turns create room for control
Break complex responses into a decision, its reason, and a next step. Offer detail on request. Do not ask several unrelated questions in one bubble; users answer only part and the model guesses the rest. For voice, read back critical values in groups and provide a visual or message-channel copy when possible.
Failure mode
Latency is often disguised with personality. Jokes and filler do not fix a blocked workflow. Measure time to first useful signal, time to completed task, interruption success, and abandoned turns. A graceful ‘I’m still checking’ state is better than pretending the system has finished.
Test with deliberate disorder
A scripted demo rarely exposes timing failures, so test people who pause mid-sentence, restart a thought, correct a number, answer before the prompt finishes, and change topic while a tool is running. In voice, include background speech and headphones disconnecting. In text, send three short messages rapidly. Review whether the system preserves the user’s intended edit, cancels obsolete output, and prevents an interrupted confirmation from executing. These cases should become automated conversation scenarios where possible and recurring moderated tests where timing perception matters.
Primary reading
Sources and limits
These links support the architecture, policy, or product behavior discussed above. Vendor documentation describes vendor features; it is not independent proof of performance. Current details should be rechecked before a production decision.