A considerable effort has been recently directed towards developing separable (mean-field) approximations for quantum molecular dynamics, such as the Time-Dependent Self-Consistent Field (TDSCF) or the Classical Separable Potential (CSP) methods. Unlike numerically exact solutions of the time-dependent Schrodinger equation, the accuracy of separable quantum dynamical simulations crucially depends on the choice of the coordinate frame. Since the approximate methods replace exact interactions between individual degrees of freedom by mean-field couplings the goal is to work with coordinates which separate modes as well as possible. Unfortunately, for larger system no practical way to optimize coordinates for mean-field quantum dynamics exists. Here, we suggest a simple and practical method for estimating the error of separable simulations, which allows to select from a given set the optimal coordinate frame, or to identify modes the couplings between which have to be treated more accurately. In the spirit of the CSP method, the time-dependent error estimate is based on differences between the exact and mean-field Hamiltonians along a swarm of classical trajectories. This makes it possible to very simply determine optimal coordinates for CSP or TDSCF propagation before actually performing any quantum simulation. The present methodology is applied to realistic and experimentally relevant systems, namely to the ultrafast relaxation following electron photodetachment in I-Arn (n=2 and 12) and Cl-H2O clusters. It is shown that the accuracy of separable quantum methods is strongly system and coordinate dependent. Comparison with numerically exact results shows that the suggested error measure correlates well with the actual error of the approximate quantum propagation, the accuracy of which can be consequently improved significantly, practically without additional computational effort. Finally, the feasibility of the proposed method for simulations of large polyatomic systems is demonstrated.