A help-desk assistant returns answers of several paragraphs, but the UI has room for about three sentences. A developer sets max_tokens to 60 to force short replies. Now many answers end mid-sentence and the responses report stop_reason "max_tokens". What is the right way to get concise, complete answers?