The choice of password composition policy to enforce on a password-protected system represents a critical security decision, and has been shown to significantly affect the vulnerability of user-chosen passwords to guessing attacks. In practice, however, this choice is not usually rigorous or justifiable, with a tendency for system administrators to choose password composition policies based on intuition alone. In this work, we propose a methodology that draws on password probability distributions constructed from large sets of real-world password data which have been filtered according to various password composition policies. Password probabilities are then redistributed to simulate different user password reselection behaviours in order to automatically determine the password composition policy that will induce the distribution of user-chosen passwords with the greatest uniformity, a metric which we show to be a useful proxy to measure the overall resistance to password guessing attacks. Further, we show that by fitting power-law equations to the password probability distributions we generate, we can justify our choice of password composition policy without any direct access to user password data. Finally, we present Skeptic, a software toolkit that implements this methodology, including a DSL to enable system administrators with no background in password security to compare and rank password composition policies. Drawing on 205,176,321 passwords across 3 datasets, we lend validity to our approach by demonstrating that the results we obtain align closely with findings from a previous empirical study into password composition policy effectiveness.