In the present socio-economic context in which the social functions of agriculture are being questioned, the transformation of production practices has become a subject of discussion between farmers, development agents and researchers. The diversity of agricultural practices may prove to be an asset for the development of a multifunctional agriculture, as it may show the complementary nature of farms as well as the innovations they carry out. It is therefore important to be able to represent this diversity in order to enrich such discussions and help researchers and development agents reformulate issues in partnership. However there is no methodology for classifying practices which thus remain confined to monographic studies. Grounded in an engineering research setting, this paper suggests a method for formalising typologies of practices, according to two theoretical choices. Following Rosch’s cognitive theory on prototypical categorisation, our method aims at formalising types as poles seen as average patterns for categories instead of squares with clear-cut boundaries. In a constructivist and situated perspective, the method is supported by a collective setting and such knowledge engineering tools as repertory-grids: it aims at collectively defining types of practice combinations linked to the expertise of participants to the project and to the problem they intend to tackle. The method consists of four stages, mainly iterative, which are illustrated along the way by examples from our engineering setting: stage A): construction of the setting, which involves choosing the zone, building a database identifying people involved in agricultural activity, choosing a population sample for the interview stage (B), and clarifying each participant’s expectations and role; stage B): semi-structured interviews with farmers, followed by the writing up of a summary sheet to form the basis for collective work in stage C; stage C): the first step in the formalisation of the typology is to characterize the criteria, and then to build prototypes using a multivariate analysis and numerous group discussions on preliminary formalisations. In accordance with the prototypical theory, types are firstly defined by more typical practices, and structural data are examined afterwards to deepen types’ description. The resulting typologies then reveal the strategic choices of farmers and their “management styles”; stage D): even though the problem is redefined in a separate step, this aspect frequently emerges during the preceding stages which allow researchers and development agents to review their future actions. Apart from the raw result of this categorisation process, the method has proven its usefulness in that it enlarges the participants’ representations of farm diversity. As an iterative learning loop, it has also provoked discussions on extension practices. This method may seem cumbersome as it requires the acquisition of a lot of data on practices and is therefore limited to research-development partnerships. Nevertheless, it has proven a useful tool for “exploratory partnerships”, by enhancing mutual understanding and stimulating collective exploration of the problem. In that sense, a typology is not a goal in itself, but a means of collectively redefining the problem.