Designing traceable, prompt-scaffolded generative AI feedback to support EFL grammar accuracy and writing complexity: evidence from male Saudi undergraduates

Generative artificial intelligence (GenAI) is increasingly employed as a source of on-demand, dialogic feedback. This study examined a traceable, prompt-scaffolded GenAI feedback agent embedded in Blackboard Ultra to support English-as-a-foreign-language (EFL) grammar learning and transfer to writing. Adopting a sequential explanatory mixed-methods classroom quasi-experimental design, participants completed parallel-form grammar tests and timed narrative writing tasks at pre- and post-test. The experimental group engaged in a structured, rewrite-first dialogue guided by versioned prompt templates that sequence diagnosis, concise rule explanation, guided reformulation, micro-practice, and optional verification. The control group received the same standard instruction and parallel revision and diary routines without agent support. All agent interactions were logged through template identifiers, timestamps, and revision-cycle markers, enabling trace-level linkage between dialogue episodes, revision behavior, and learning outcomes, thereby supporting reproducibility. Mixed-design analyses showed that the prompt-scaffolded group achieved significantly larger gains in explicit grammatical knowledge and indices of syntactic elaboration and morphological complexity, with convergent increases in compression-based complexity. Qualitative findings further underscored the explanatory value of the protocol, while also highlighting the continuing necessity of verification routines and instructor oversight. The study contributes process-linked evidence and a minimally specified, audit-ready design pattern for human–AI collaboration in LMS-embedded interactive learning environments.