The paper outlines a multi-step approach for recommender system design, where a Large Language Model is guided by a knowledge-based system, exploiting a user model. The work focuses on energy communities, with the aim to engage potential members by leveraging their values, literacy, and available resources to generate personalized descriptions of the benefits they could achieve if they join the EC. The approach involves the fine-tuning of an LLM, by integrating features of both items and users. The dataset used is based on domain-specific documents, developed thanks to the first phase of the project, which involved studying the domain through EC case studies in different EU countries. We initially compared different LLM models through an evaluation done with domain experts, who assessed the clarity, relevance, tone, and alignment with the goal of the system. Based on the results, we selected the best performing model, LLaMA3-8B-Instruct, that was employed in the preliminary user evaluation that is presented in this paper.