State what you will say out loud to your assistant for a real purpose, setting the context at the beginning and one request at a time. Return JSON {"goal":"...","spoken_prompt":"...","followup":"..."}, where spoken_prompt is the spoken phrase itself, followup is how you will clarify if he misunderstands.
A live conversation with an assistant works like a walkie-talkie with a very smart interlocutor: he hears the flow of speech, breaks it down into words and at the same time guesses the meaning from the context. Hence the rules. First, give the context at the beginning, in one phrase: who you are, what the situation is, what the output is needed. A model, like a person, understands a request better when it knows the frame. Second, one request at a time; If you lump three questions into one sentence, the answer will be vague. Third, speak in ordinary words, but a little slower on names, numbers and terms - these are the areas where recognition errors most often occur. Pro trick: allow yourself to interrupt and correct out loud - “stop, I meant something else” - the live mode understands this and picks it up. Second technique: end a complex request with a separate hint to the assistant himself (meta-phrase) “if something is not clear, ask again,” this includes clarifying questions instead of guessing. A common rookie mistake is to remain silent and wait for the perfect answer; conversational AI is strong precisely because of dialogue, not just one command.
Unlock access to submit solutions for instant AI review.