I designed how voice should handle Nest devices, and how that control stays safe, clear, and accountable. The work covered the action grammar, confirmations and failures, disambiguation, identity and permissions, and the instrumentation underneath. The rule matched the app: do the thing asked, or say plainly why not.

People don’t speak in API verbs, so the system shouldn’t expect them to. We mapped natural phrases to capabilities—“set it to seventy”, “turn on the fan”, “lock the front door”—with room context as the quiet default and named devices as the override. When the target was uncertain, the follow-up stayed short: “Which thermostat, Living Room or Bedroom?” If someone stepped away mid-question, we cancelled rather than guess, so nothing surprising happened a beat later.

Safety needed its own shape. Locking allowed a single turn with a brief confirmation. Unlocking asked for explicit confirmation and a higher bar for identity. If identity was unknown or permissions didn’t allow the request, Assistant declined in one line and offered the next step. Confirmation windows were tight enough that actions did not fire late.

Language stayed aligned with the Nest model, which meant voice, app, and notifications carried the same meaning. Success used one line and an earcon—“Locked Front Door.” Failure told the truth and pointed to recovery: “Can’t lock — door is obstructed. Try again, or check the bolt.” Offline did not pretend: “Device is offline. Check Connect.” So if something went wrong, the fix was already in view.

Identity set the boundaries. With voice profiles, actions were attributed to a household member; without them, the speaker was treated as a guest with limited scope. Every execution wrote to device history in the same one-line format—“Locked by [Name] at 9:28 AM”—so it was clear who did what, when.

With data scientists I defined the events and the success metrics—first-attempt success, confirmation completion, time to complete—and the guardrails that kept us honest: denial rates on safety actions, false-positive identity, and support contacts tied to voice flows. I wrote the initial prompts and confirmations, partnered with a copywriter to tighten cadence, and prototyped edge cases with engineering so spoken feedback matched what you would see in the app a moment later. The integration shipped and kept its shape, so voice became another steady surface when hands were full.