Abstract. The generation of formal poetry involves both complex creativity usually exercised by a human poet - and strict algorithmic restrictions regarding the metrical structure of the poem - determined by literary tradition. Starting from a generating system that enforces automatically the metrical restrictions, this paper presents a model for the literary style of a user based on four key features for user preferences - word selection, language structures, poem planning, and restrictions on realisation - governing the generation of poetry from input data provided by the user - a prose paraphrase of the intended message, a task specific vocabulary, and a corpus of construction patterns. The system exploits the CBR paradigm as a means to evolve a case base (a vocabulary / construction pattern grouping) that effectively models the style of a specific user as a result of multiple iterations through the CBR cycle.