Conversation design
Conversations are state machines wearing small talk: openings that set expectations, clarifying questions that earn their turn, and endings that don't trap the user.
A support conversation has a shape, and designing it beats improvising it. The four moments that decide most of the experience:
- The opening sets expectations in one breath: who (an AI assistant), what it's good at (the in-scope list, compressed: 'orders, returns, products, and store info'), and the exit ('or type agent anytime for a person'). Openings that oversell ('Ask me anything!') buy the first turn and bankrupt the fifth.
- Clarifying questions — the highest-skill move in conversation design. Ask when the answer genuinely forks ('Is this for an online order or a store purchase?'), never to stall, and never more than one at a time. The anti-pattern customers hate most: the interrogation (three questions before any value). Rule: give something before asking anything — 'Our return window is 30 days. To check yours specifically, what's the order number?'
- The confirmation beat before anything that acts: 'I'll email a return label for order HL-1042 to dana@… — go ahead?' Cheap to add, and the difference between an agent that acts with customers and one that acts on them.
- Endings that release: resolve → confirm resolution → offer the exit ('Anything else? If something's off, reply here or type agent'). Never loop a customer who said 'no thanks' back into a menu. And the conversation's whole transcript should read coherently — Module 4 depends on it, because the transcript is what the human inherits.
Designing for the mood, not just the message
Support traffic arrives in three moods, and the same information lands differently in each: task mode (just answer fast — most traffic), confusion mode (they don't know what to ask; the agent's job is diagnosis: 'It sounds like the order shows delivered but nothing arrived — is that right?'), and distress mode (angry, urgent, or vulnerable — shorten sentences, drop the persona's warmth ornaments, lead with what happens next, and lower the handoff threshold; Module 4 formalizes this). You can't detect mood perfectly; you can design responses that degrade gracefully when you guess wrong — which mostly means: never cheerful boilerplate under a furious message.
Before touching a system prompt, hand-write 8-10 complete ideal conversations — one per in-scope topic plus the ugly ones (angry customer, out-of-scope request, handoff). These become your north star for prompting, your few-shot material, and (Module 5) the seed of your eval suite. Teams that skip this write prompts that argue with themselves; the golden conversations settle the arguments on paper, where edits are free.