Optimal team composition has been the focus of exhaustive analysis, academic and otherwise. Yet, much of this analysis has ignored possible dynamic effects: e.g., anticipating that team formation is based on prior performance will affect prior performance. We test this hypothesis in a lab experiment with two stages of a real effort task. Participants first work individually without monetary incentives and are then assigned to teams of two where compensation is based on team performance. Our results are consistent with a simple investment-cum-matching model: pairing the worst performing individuals with the best yields 20% lower first stage effort than random matching. Pairing the best with the best, however, yields 5% higher first stage effort than random matching. In line with the theory the latter result is more pronounced when the task has less scope for learning-by-doing. Moreover, pairing the best with the best achieves the same effort response as having explicit monetary incentives in the first stage.