Prompting behaviours and feedback uptake in GenAI-supported writing: longitudinal changes in relation to writing self-regulation and associations with later writing performance

Learner interaction with generative artificial intelligence (GenAI) as part of writing development has drawn increasing attention. However, less is known about how such interaction changes over time and varies with learner characteristics. Adopting a process-oriented perspective, this study examined longitudinal changes in prompting behaviours and feedback uptake, their relations with writing self-regulation, and associations with later writing performance. Data came from 56 undergraduates in a ten-week GenAI-supported academic writing programme with a one-month follow-up. Learners’ draft pairs and GenAI chat logs were coded, transformed into log-ratio contrasts where appropriate, and analysed using linear mixed-effects models and baseline-controlled hierarchical regression. Results showed that prompting frequency increased from T1 to T2 and declined at T3, while meaning-making, regulation-oriented, and evaluative prompting increased relative to direct acquisition prompting from T1 to later time points. The increase in meaning-making relative to direct acquisition was stronger at lower levels of self-regulation, whereas increases in regulation-oriented prompting and in rejected relative to fully implemented uptake were stronger at higher levels. T1-to-T2 increases in meaning-making and regulation-oriented prompting relative to direct acquisition, and in rejected relative to fully implemented uptake, were positively associated with T3 initial-draft performance after baseline performance and self-regulation were controlled. The findings show that GenAI interaction patterns vary with writing self-regulation and that specific prompting and uptake changes are associated with later writing performance. They underscore the value of examining learner differences in human–GenAI interaction from a longitudinal, process-oriented perspective and of providing support responsive to different interaction patterns.