Although traditional data-to-text Natural Language Generation (NLG) technology is useful for disseminating insights from data science projects as textual narratives to wider audiences, traditional NLG applications rely on application knowledge that is not transferable to new data science projects without significant effort. This limitation arises because application knowledge is not acquired and organized around key concepts of data science, such as questions that motivate investigations, algorithms that answer these questions, and sensemaking processes that apply this knowledge at all stages. This paper introduces a “rapid prototyping followed by iterative refinement” methodology and its corresponding development framework. It leverages the strengths of cognitive sensemaking within a question-driven data science process. In the context of human-machine collaboration in data science, our approach facilitates a greater share of sensemaking responsibility to the machine side, enhancing the collaboration between humans and machines. The paper also proposes an information consistency principle to ensure alignment among input data, model results, user requirements, and generated reports. The effectiveness of this framework has been demonstrated through experiments, confirming that the quality of the reports generated matches that of reports crafted by experts.