This paper describes an approach to treebank development which relies on the manual development of annotation tools. The overall process of tree annotation is described, and a special emphasis is put on the description of the last tool which has been built, i.e. a dependency-based robust chunk parser. The modularization of the parser and the central role of verbal subcategorization is presented. The first experimental results, carried out on a corpus of 645 sentences are reported and discussed. 1.