A Home Assistant user running local Qwen3 1.7B reports errors such as a light-status question triggering a bedtime routine and a room-specific command affecting the whole house. The user reduced exposed entities and grouped multi-device actions into tested scripts. Temperature ranges and room boundaries are enforced by execution tools rather than prompts, with sensitive controls requiring separate confirmation or remaining inaccessible. This is one user’s experience, but separating language interpretation from execution authority is a reusable design principle.