Was just wanting something like this today, running qwen3 q2 model. Spent 3 paragraphs working out how to respond to "hello how are you".

The third options would be a constrained thinking, ie 1-2 sentences max before output.