Sedentary behaviour and obesity have become one of the greatest threats to modern society’s health and well-being. The treatment and management of non-communicable diseases associated with them cost governments billions of dollars in healthcare expenditure and lost economic opportunities. To promote physical activity, fitness recommender systems have been proposed. However, most of the existing systems often use ad hoc design principles or a one-size-fits-all approach. To address these shortcomings, we proposed the incorporation of explainable, contextual, and theory-driven frameworks into the design and evaluation of fitness-based recommender systems to make them more effective. In this paper, we integrated classical machine learning, state-of-the-art large language models, and psychology-based design frameworks to generate adaptable fitness recommendations in real time.