A Systemic Intervention for Human-Artificial Intelligence Co-Design of Lesson Plans: Integrating Pedagogical Theories into Prompt Engineering
TL;DR
These findings demonstrate that pedagogically grounded, multi-stage prompts are designed to reconfigure teacher-artificial intelligence interaction from passive output consumption toward structured collaborative design, offering a scalable intervention for integrating LLMs into instructional workflows.
Abstract
Large language models (LLMs) can assist lesson planning, but simple prompts often yield incomplete and misaligned outputs. This study proposes a three-step prompt framework grounded in Bloom’s taxonomy, Adaptive Control of Thought-Rational (ACT-R) theory, Gagné’s nine events, and problem-chain theory, decomposing planning into objective, unit, and activity design. Using three DeepSeek models (R1, V3, 32B) and five prompting strategies, 150 lesson plans were generated on ten computer networking topics. Coverage of Gagné’s nine events and functional quality were evaluated via an LLM judge and human validation. All theory-based strategies significantly outperformed naive prompting, raising Gagné’s event coverage above 90% in the full corpus and from 74.1% to 89.8–93.5% in human ratings. Functional quality scores improved by up to 17.3% (LLM judge) and 53.6% (human raters). Gagné’s five-stage design outperformed ACT-R’s three-stage design under base conditions, while problem-chain guidance benefited ACT-R substantially. Model capability moderated gains: smaller models benefited most in structural completeness, stronger reasoners achieved higher absolute quality. These findings demonstrate that pedagogically grounded, multi-stage prompts are designed to reconfigure teacher-artificial intelligence (AI) interaction from passive output consumption toward structured collaborative design, offering a scalable intervention for integrating LLMs into instructional workflows.