1. Convert voice to text.
2. Pre-parse vs voice command library of commands. If there are, do them, pass confirmation and jump to 6.
3. If no valid commands, then pass to LLM.
4. have LLM heavily trained on commands and some API output for them. If none, then other responses
5. have response checked for API outputs, handle them appropriately and send confirmation forward, otherwise pass on output.
6. Convert to voice.

The LLM part obviously also needs all kinds of sanitation on both sides like they do now, but exact commands should preempt the LLM entirely, if you’re insisting on using one.