The development of conversational large language models has brought a drastic change in educational settings, both in school and at home. While the Curriculum Development Centre of Nepal publishes detailed guides for teachers to create and use teaching materials in their classrooms, for subjects like Nepali these remain largely unused: theoretical topics are taught by rote learning and reproduced in examinations, making it a subject in which students score lowest and feel least confident. With the advent of LLMs, there is a real opportunity to turn these unused guidelines into large amounts of diverse teaching and evaluation material without adding to teachers’ burden. In this work, we use conversational LLMs to generate curriculum-grounded questions for the Grade 7 and 10 Nepali subject and evaluate their usability in a classroom setting. Results show that Nepali objectives require more than one Bloom’s taxonomy level to define them: forcing a single level per objective drops teacher agreement to near chance 𝜅 = 0.13, while allowing multiple levels recovers agreement on at least one level for 84.6% of objectives. We also observe that LLMs consistently dumb down the taxonomy level of objectives, plateauing at roughly 0.50 Macro F1, whereas a supervised classifier over frozen multilingual encoders reaches 0.612. Finally, teachers evaluated the generated questions through a modified eight-item rubric: grounding questions in the relevant taxonomy levels raised teacher endorsement from 12.5% to 88%. The question generation tool made from this approach is currently deployed in a government school in Nepal.