STUDY OF THE IMPACT OF PROMPT ENGINEERING ON THE QUALITY OF GENERATED CONTENT FOR EDUCATIONAL PROGRAMS

This study analyzes how prompt engineering affects the quality of language model–generated content for educational programs. It evaluates various prompting methods across modern models, from compact to large-scale, using a dataset of 8,269 programs. A structural- semantic gradation of components by interpretability was identified: goals, disciplines, and learning outcomes were highly interpretable (high cosine similarity), while components like the Atlas of the profession were less so. Prompting methods showed varying effectiveness: few-shot learning worked best for goals, and chain of thought for professional standards. A key finding is the "specialized advantage": a LoRA-adapted model outperformed larger, non-specialized ones, challenging the idea that bigger always means better. The study offers practical strategies for selecting models and prompts, aiding educational analytics and future research into AI-based interpretation of complex educational content