This paper proposes a strategy to derive an adjoint-based optimization code from a numerical simulation code. The strategy involves utilization of Automatic Differentiation (AD), and in fact was a motivating application for several studies and improvements of AD. The strategy gives best results when interfaced with modern optimization methods, such as "one-shot" or "multi-level". This paper presents a set of application examples where this strategy was used, and gives some experimental results. We also discuss some theoretical questions related to the computation of gradients and the optimization process in general.