We built a natural language processing (NLP) system implementing a "virtual writing conference" for elementary-school children, with German as the target language. Currently, state-of-the-art computer support for writing tasks is restricted to multiple-choice questions or quizzes because automatic parsing of the often ambiguous and fragmentary texts produced by students presents insurmountable problems. Here, we follow a different course by deploying natural language generation technology to evaluate and improve the grammatical quality of learner output. Based on an abstract representation of the story under construction, all paraphrases of simple and combined clauses are generated fully and automatically. From this source, the system produces exercises enabling the learners to improve their sentences. We describe how a pedagogical agent supports the writing process of elementary-school children by providing appropriate feedback. (Contains 3 figures, 1 table and 11 notes.)