A Bayesian methodology is described for designing experiments or surveys that will optimally complement all previously available information. This methodology uses strong prior information to linearize the problem, and to guide the design toward maximally reducing forecast uncertainties in the interpretation of the future experiment. The prior information could possibly be correlated among model parameters or the observation noise. With no prior information this approach reduces to the fast recursive implementation of the D-optimality criterion. Synthetic geophysical tomography examples are used to illustrate the benefits of this approach. The dependence of the design on the model representation is also discussed.