From Generation to Adaptation: Comparing AI-Assisted Strategies in High School Programming Education

TL;DR

This study compares two pedagogical strategies using language model code generation in high school programming, showing that adapting existing code with scaffolding is more effective than generating from scratch.

Contribution

It introduces a dual-scaffolding model combining technical and pedagogical support for effective AI-assisted programming education.

Findings

MFU-based approach achieved 100% MVP completion

From-scratch approach achieved only 20% MVP completion

Effective integration depends on instructional design, not just AI capabilities

Abstract

This exploratory case study investigated two contrasting pedagogical approaches for LCA-assisted programming with five novice high school students preparing for a WeChat Mini Program competition. In Phase 1, students used LCAs to generate code from abstract specifications (From-Scratch approach), achieving only 20% MVP completion. In Phase 2, students adapted existing Minimal Functional Units (MFUs), small, functional code examples, using LCAs, achieving 100% MVP completion. Analysis revealed that the MFU-based approach succeeded by aligning with LCA strengths in pattern modification rather than de novo generation, while providing cognitive scaffolds that enabled students to navigate complex development tasks. The study introduces a dual-scaffolding model combining technical support (MFUs) with pedagogical guidance (structured prompting strategies), demonstrating that effective LCA integration depends less on AI capabilities than on instructional design. These findings offer practical guidance for educators seeking to transform AI tools from sources of frustration into productive learning partners in programming education.