A set of 30-day integrations made with the operational ECMWF model over a period of three and a half years (from 1985 to 1988) is studied. The impact of model reformulations during this period on the climate drift of the model is assessed, and the level of extended-range forecast skill achieved by the operational model is studied. Diagnostics of zonal average 30-day mean wind and temperature error show a systematic reduction in the extratropics over three years. It is argued that these improvements are consistent with the changes in model formulation. Eddy fluxes of heat and momentum, and levels of eddy kinetic energy are similarly improved. We also show that the ability to simulate blocking activity in the extended range has improved substantially over the three winters. From maps showing the ratio of systematic to total error it is concluded that, by the final winter, the middle-latitude extended-range error is almost entirely associated with random errors. This is supported by the growth of the dispersion, or spread, between forecasts initialized 24 hours apart. In particular, for the extended winter 1987/88, it is shown that the asymptotic level of (internal) spread is comparable with the asymptotic level of (external) skill. By contrast, in the tropics, the total error is dominated by the systematic error for all years. Despite improvements to the model, extended-range forecast skill is modest. Skill has been measured in a spatially-filtered three-dimensional phase space, spanned by rotated EOF (empirical orthogonal function) coefficients of 500 mb height, which contain the principal weather regimes in the atmosphere. On average the skill scores are consistent with those of other major centres studying dynamical extended-range prediction. In particular there is evidence of skill to day 15 and possibly to day 20 in middle latitudes. Beyond day 20 skill is marginal. We demonstrate that in the winter sample of operational-model forecasts, skill is correlated with the observed value of the first rotated EOF, which can be described as a global version of the Pacific/North American mode. By contrast, the ability to predict monthly mean fields in the tropics depends primarily on a knowledge of the boundary forcing, including sea surface temperature. It is shown that some aspects of the interannual variations in the summer monsoon circulation can be predicted, in particular the tropical easterly jet anomalies, African rainfall, and, to a much lesser extent, Indian monsoon rainfall.