I model individuals working on a long-term task who, due to projection bias, underestimate changes in marginal disutility. At the start of each day, when current marginal disutility is low, such individuals overestimate how much they will work, and as they grow tired from working, they plan to work less and less. Despite these fluctuating plans, when they face decreasing returns to effort, they work optimally, yet if they commit in advance, they overcommit. When facing increasing returns, they may repeatedly start and quit, working when rested only to give up as they grow tired. In all-or-nothing tasks, starting-and-quitting can lead to wasted effort on never-to-be-completed tasks, or to completing them by working inefficiently little early on.