Current day operational ensemble weather prediction systems generally rely upon perturbed atmospheric initial states, thereby neglecting the eventual effect on the atmospheric evolution that uncertainties in initial soil temperature and moisture fields could bring about during the summer months. The purpose of this study is to examine the role of the soil states in medium-range weather predictability. A limited area weather prediction model is used with the atmosphere/ land-surface system in coupled or uncoupled mode. It covers Europe and part of the north Atlantic, and is driven by prescribed sea-surface temperatures over the sea, and by atmospheric reanalyses at its lateral boundaries. A series of  3 member ensembles of summer simulations are used to assess the predictability of a reference simulation assumed to be perfect. In a first step, two ensembles are simulated: the first with the atmosphere coupled to the land-surface model, the second in the uncoupled mode with perfect soil conditions prescribed every 6 hours. Subsequent experiments are combinations thereof, in which the uncoupled and coupled modes alternate in the course of a simulation. The results show that there are "stable" and "unstable" periods in the weather evolution under consideration. During the stable periods, the predictability (measured in terms of ensemble spread at 500 hPa) of the coupled and uncoupled dynamical systems is almost identical; prescribing the perfect soil conditions has a negligible impact upon the atmospheric predictability. In contrast, the predictability during an unstable phase is found to be remarkably improved in the uncoupled ensembles. This effect results from guiding the atmospheric phase-space trajectory along its perfect evolution. It persists even when switching back from the uncoupled to the coupled mode prior to the onset of the unstable phase, a result that underlines the importance of soil moisture and temperature in data assimilation systems.