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
Design Method
- Scope by intent, ruthlessly. From the intent list, the agent owns only intents that are (a) high-volume, (b) completable with its actual system access, and (c) low-stakes-if-wrong. It triages everything it can identify but not complete. It immediately passes anything emotional, legal, or high-value — a furious caller is a human's job on the first turn, not after three failed bot turns.
- Disclose and set the frame in the first five seconds. The agent says it's an AI (increasingly required by law; always required by trust), what it can do, and how to reach a human ("say 'agent' anytime"). Hiding the escape hatch inflates containment metrics and rage in equal measure.
- Design turns for ears, not eyes. One question per turn · ≤2 sentences before yielding · numbers and options in threes at most ("I can do A, B, or C — which one?") · never read a paragraph. Anything long ("your options are…") gets offered as SMS/email instead of spoken.
- Engineer the mechanics that make it feel alive:
- Barge-in on: the caller can interrupt any utterance; the agent stops mid-sentence and processes.
- Latency masked: acknowledge within ~1s ("let me check that…") whenever a lookup exceeds it; dead air past 2s is where trust dies.
- Confirmation proportional to stakes: implicit for low stakes ("okay, Tuesday…"), explicit read-back for money, addresses, and anything irreversible.
- Repair, not repeat: on a misunderstanding, change strategy — rephrase, offer options, or fall to keypad — never re-ask the same question the same way twice.
- Make the handoff a feature. Triggers: caller asks (always, instantly) · two failed repairs on one slot · negative-emotion cues · any regulated topic. The transfer carries a whisper summary (who, what they want, what's been tried, account pulled up) — the caller never repeats themselves; that single property beats every other quality bar in perceived experience.
- Score what callers feel, not what dashboards flatter. Containment alone is gameable (trap callers and containment "improves"). The scorecard pairs it with: task success as the caller defines it (post-call yes/no), escapes-requested rate, repair rate, silent-transfer rate, and hang-ups mid-flow. Set launch gates on the pairs.
Output Format
Voice Agent Spec: [line/product]
Intent scope
| Intent | Volume | Own / Triage / Pass | Why |
|---|
Opening script: [verbatim — disclosure, capability, escape hatch]
Conversation architecture: [turn rules · confirmation strategy by stakes · the repair ladder (rephrase → options → keypad → human)]
Mechanics: [barge-in behaviour · latency masking thresholds · silence handling]
Handoff: [triggers · whisper-summary fields · after-hours behaviour]
Compliance: [disclosure line · recording consent flow · statements the agent must never make]
Launch scorecard
| Metric | Gate | Why paired |
|---|---|---|
| Containment + caller-scored success | containment alone is gameable | |
| Escape-request rate | measures trapped callers | |
| Repair rate / hang-ups mid-flow | frustration signals |
Quality Checks
- Every owned intent is completable with the agent's actual system access — no "owns refunds" without refund API access
- The opening discloses AI status and the escape hatch, verbatim in the spec
- No designed utterance exceeds two sentences before yielding
- The repair ladder changes strategy at each rung — no repeat-louder step
- Handoff carries the whisper summary; "please hold while I transfer you" to a cold human fails the spec
- The scorecard pairs containment with caller-scored success
Anti-Patterns
- Do not port the chatbot script to voice — text tolerates paragraphs and menus; ears don't
- Do not hide the human escape hatch to protect containment metrics — callers find the exit anyway, angrier
- Do not let the agent bluff on regulated topics (medical, legal, financial advice) — pass or read the approved statement
- Do not re-ask a failed question unchanged — the caller heard you; the strategy failed, not their ears
- Do not launch without the mid-flow hang-up metric — it's where voice agents quietly hemorrhage trust
Example Trigger Phrases
- "Design a voice agent."
- "Automate a phone line."
- "Spec an IVR replacement."
- "Review why callers hate an existing voice bot."
