We improve seasonal hindcast skill of European summer climate in an ensemble based coupled seasonal prediction system by selecting individual ensemble members based on their respective consistent chain of processes that describe a physical mechanism. This mechanism is associated with the second mode of seasonal climate variability in the North-Atlantic- European sector and is contrary to the summer North Atlantic Oscillation. We initially analyse the mechanism in the ERAInterim reanalysis and then test the influence of the mechanism on European hindcast skill in an initialised coupled seasonal climate model. We show that the mechanism originates in the tropical North Atlantic in spring, where either warm or cold sea surface temperature anomalies (SSTs) are connected with the European climate by an upper-level wave-train. This wavetrain is accompanied by a zonal pressure gradient, that in turn influences the climate over central Europe in the following summer. We analyse the seasonal summer hindcast skill in a mixed resolution hindcast ensemble simulation generated by MPI-ESM, with 30 members starting every year in May. While the mean over the full ensemble shows no seasonal hindcast skill in summer, we achieve significant hindcast skill through forming a new mean over subselected ensemble members. For this selection, we test every ensemble member for the proposed consistent chain of connections between the wave-train, the zonal pressure gradient and their impact on European summer temperatures, and find that the processes that describe the mechanism are not represented in every ensemble member. Due to its influence on European summer climate, we use the condition of the persistent spring SSTs to anticipate the phase of the mechanism in each considered year. We thus use statistical relations to select ensemble members generated by a dynamical prediction system. With this approach, we significantly enhance the seasonal hindcast skill and the reliability of the hindcasts in the North-Atlantic-European sector, especially in the areas where the mechanism is showing a prominent signal. Since we only use knowledge that would be available in a real forecast set-up, this approach can potentially be applied in operational ensemble prediction systems.