The current models of skill in video games make one of two impositions on players: either to provide an estimate of their own skill, or complete several games before they can be properly assessed. However, in order to experience the most enjoyment and greatest sense of immersion, players need to play against the right difficulty. In order to assign the appropriate difficulty, the player's skill must first be captured accurately and quickly, before the player gets frustrated or bored. Rather than relying on game results that need to be averaged over several games, this thesis proposes predicting a player's skill from their behaviour within the first game. In order to do this, we explore methods for measuring skill in both a multiplayer and single-player game and methods for extracting appropriate information from the player's behaviour. The resulting predictions can then be used to automatically assign an appropriate difficulty to the player.In a multiplayer environment, we first demonstrate that a player's final rank canbe predicted within the first 30 seconds of a game with a correlation of over 0.8.This process is transferred to a single-player first-person shooter, where our modelis shown to assign difficulties comparable to a player's own assessment of theirskill within the first 30 seconds of a campaign. We argue that these methods forcapturing skill in a first-person shooter are transferable to other genres, and havethe potential to improve difficulty selection systems.