This paper presents a method for mass-generating isomorphic physics problems using ChatGPT. Using prompt chaining and tools, it precisely controls structural variations, such as numerical values and spatial relationships, while supporting diverse contextual variations in the problem body. It addresses key limitations of existing LLM-based methods by leveraging a Python code interpreter to support automatic solution verification and simple diagram generation. By generating two isomorphic problem banks and comparing them to simple prompt-based methods, we demonstrate that prompt chaining produces significantly higher quality and more consistent results. This study demonstrates an efficient problem generation method accessible to even the average instructor and opens new possibilities for personalized adaptive testing and automated content development.
