Adjoint models provide a powerful method for estimating the sensitivity of error in short-range numerical weather forecasts to initial conditions. A recent and important extension of the adjoint sensitivity technique uses the adjoint of a data assimilation procedure to provide a direct sensitivity of forecast error (or other forecast costfunction) to the actual observations. This talk will describe applications of the observation sensitivity method to several cases which include very large 72-hr forecast errors, including the January 2000 East Coast Blizzard. Observation sensitivity will be used to estimate the relative importance of data from existing in-situ and satellite instruments, including radiosondes, surface observations, wind data derived from geostationary satellites, and other observing systems. In addition, it will be shown that the observation sensitivity method can be used to design more optimal observing networks, in which new or hypothetical networks of in-situ and satellite data are designed to fill in critical gaps in the existing global observing network.