Most poetry generation research overlooks the challenging task of transforming prose into structured verse, especially for morphologically rich languages like Russian, where complex stress and flexible word order demand precise control over both semantics and poetic form. This paper addresses this underexplored area by proposing a novel approach to Russian prose-to-poetry transformation with explicit rhyme and meter control. We introduce a new dataset of over 180,000 Russian poetic quatrains annotated with rhyme schemes and stress patterns, along with 11,000 parallel prose-poetry pairs. We fine-tune a Qwen-based language model using explicit markup for rhyme and stress, evaluating our models against strong baselines (Gemini and GigaChat) on semantic and formal metrics. Our results demonstrate that rhyme and stress markup significantly improves rhyming and meter accuracy. However, we identify a crucial trade-off: while formal control improves, it often comes at the cost of preserving the original prose’s meaning, overall semantic coherence, and grammatical correctness, proving detrimental to overall text quality. This underscores the inherent difficulty of simultaneously maintaining strict poetic form and preserving original content. To better integrate metrical and rhyme information in generation, we plan to explore character-level models and pretraining on stress-annotated corpora.