Large language models provide a flexible way to generate natural-language explanations for recommender systems, but more fluent explanations do not necessarily improve user understanding. This work compares four prompting strategies across three LLMs, three recommender families, and two application domains. Explanations are evaluated for persuasiveness, transparency, satisfaction, and accuracy, together with conventional NLP metrics. LLM-generated explanations generally receive stronger quality judgments than rule-based templates, but different prompts optimize different qualities. Traditional text-similarity metrics can move in the opposite direction from explanation-quality judgments, suggesting that prompt design is a substantive explanation-design choice rather than merely an implementation detail.