A little hint:
Look at the prompt trick I used here with my LLM timer script to make the LLM use my timer script instead of the default HASS Timer intents.
Works reliable (at least with the OpenAi models that I tested so far) for me.