In a constructive learning setting, a robot builds up beliefs about the world by interacting —interacting with the world, and with other agents. Asking questions is key in such a setting. It provides a mechanism for interactively exploring possibilities, to extend and explain the robot's beliefs. The paper focuses on how to linguistically phrase questions in dialogue. How well the point of a question gets across depends on how it is put. It needs to be effective in making transparent the agent's intentions and beliefs behind raising the question, and in helping to scaffold the dialogue such that the desired answers can be obtained. The paper proposes an algorithm for deciding what to include in formulating a question. Its formulation is based on the idea of considering transparency and scaffolding as referential aspects of a question.