Researchers have studied how large language models (LLMs) can transform computer science (CS) instruction at the college level, both from the student and instructor perspectives. At the K-8 level, where culturally responsive teaching is a large focus, instructional materials connected to students’ lives are increasingly used to enhance engagement, and thus learning outcomes. While promising, little is known about using LLMs to augment K-8 CS instruction.

This paper explores the potential of teachers leveraging LLMs to brainstorm instructional Scratch projects using (1) structured Scratch projects from an existing curriculum, and (2) cultural resources presented in their classroom. Specifically, we use GPT-3 to generate projects that match the technical characteristics of an existing project and are related to a proposed theme. We qualitatively analyze 300 project ideas generated by GPT through natural language interactions, and find that 81% of the GPT-generated ideas satisfy our metrics for technical alignment and theme quality. We identified two major weaknesses: code complexity of proposed projects and the presence of potential biases that require further human filtering. We conclude that while not ready as a stand-alone solution, GPT could be used effectively to assist brainstorming ideas for customized instructional materials.