The paper studies a data driven design approach of HMM topology in a hybrid Neuro-Markovian system for on-line cursive handwriting recognition. Artificial neural networks (ANNs) are used as primitive models at state level and hidden Markov models (HMMs) are used at character level. Primitives are shared among all characters in the alphabet and an individual handwriting is characterized by a primitive sequence. The typical prototypes of a letter are reflected in HMM's topology. Firstly, we build a prototype analyser that creates a primitive prototype for each training example. Secondly, a number of the most typical prototypes are selected for each letter through a special clustering method. At last, letter models are built by using the selected prototypes as Markov chain's topology. The concepted system is evaluated on the wildly used UNIPEN database and the advantages are clearly approved with very encouraging results.