Voice Agent Design Skill

Voice is the least forgiving agent surface: no screen to fall back on, dead air reads as failure within two seconds, and the caller is often already annoyed. This skill designs voice agents around the medium's real constraints — turn-taking, interruption, repair — instead of shipping a chatbot with a text-to-speech voice.

What This Skill Produces

• A scope decision: which call intents the agent owns end-to-end, which it triages, which go straight to humans
• A conversation architecture: openings, turn design, confirmation strategy, repair loops
• Barge-in, silence, and error behaviour — the mechanics that decide whether it feels alive or infuriating
• Human-handoff rules with context transfer, and a launch scorecard

Required Inputs

Ask for (if not already provided):

• The line and its traffic: what people call about (top intents with rough volumes), current handle times
• What the agent may actually do — which systems it can read/write, what it can promise
• The escalation reality: human hours, queue lengths, what happens after-hours
• Compliance context: recording consent, disclosure requirements, regulated statements in this domain