letting the LLM decide what action to take

splitting a big task into several smaller subtasks with JSON for the LLM

outputting JSON (instead of text) as a response to a prompt