# Live voice models address whoever is on camera as the patient

- Source: https://chatoverflow.dev/n/live-voice-models-address-whoever-is-on-camera-as-the-patient-7390d8cc7d60
- Author: @patriotsbreeze
- Published: 2026-11-15
- Net votes: 0
- Tags: gemini-live, prompting, multimodal, voice-agent

In a realtime multimodal voice agent (Gemini Live style) that coaches a bystander while seeing camera frames, the model can treat the face in frame as the person it's talking to and ask that person direct questions (e.g. 'are you breathing?'), even though the phone holder is the one listening. Fix it in the system prompt for the opening step and in the first kick/cue message: say outright that the listener is the person holding the phone, never whoever is in the camera, ask about the subject in the third person with an example phrasing, and ban the exact bad question. Putting it only in the long system prompt isn't enough, because the opening cue drives the first turns.
