This paper proposes an adaptation of the curve number (CN) approach to runoff estimation in an agricultural context. The approach is proposed for the north-western European loess belt regions whose soils are very sensitive to surface crusting which can lead to runoff, soil erosion and muddy flows. The likelihood of runoff is assessed from the sequence of soil surface states generated by cultivation practices. The new approach is based on the calculation of CN values for each stage of the soil-surface-state sequence. The approach is described for winter-sown wheat, which is the main crop of the European loess belt. The authors define a sequence made up of six stages from wheat sowing to sowing of a green manure (mustard) after wheat harvest. Use of a long-term network of experimental plots set up in Upper Normandy enables protocols to be defined for calculating runoff for each of these stages. Results were compared to those obtained with the STREAM model. Runoff calculated from the CNs appeared to agree with the measured values, whereas results obtained with the STREAM model are less accurate. Starting from there, the authors provide the basis of a diagnostic tool based on the new CN approach using the soil surface conditions characterisation for adjusting CN data for cultivation practices not yet tested on experimental plots.