Since high dropout rates in online learning platforms were reported, various factors affecting learner retention have been identified, with learners' perceptions of their experiences playing a crucial role in shaping their persistence. For instance, Kittur et al. highlight how success expectations are shaped by perceived system fit and course difficulty. Recent advances in generative Artificial Intelligence (GenAI) present new possibilities for GenAI-mediated learning. AI-generated instructional messages are often perceived as clearer than human-written content, but their impact on learners' perceptions of skill-building experiences remains underexplored. This study examines GenAI-mediated learning in a self-directed context, focusing on communication skills. We compare three messaging styles - Affective, Cognitive, and Action-Oriented - to investigate their influence on learners' perceptions of the learning process. We applied this approach to ten instructional units, using GenAI to generate 30 learning items. Three evaluators assessed them for desirability and appropriateness through numerical ratings and open-ended feedback. The 180 excerpts were analyzed using reflexive thematic analysis, revealing four overarching themes: Prerequisite Common Ground, Intrinsic Value, User Responses, and Expressed Preferences. We discuss these insights to inform the design of GenAI-mediated, self-directed skill-building, with the goal of enhancing engagement, persistence, and learning outcomes.