A chief hindrance to the practical value of AI scheduling and planning tools stems from the difficulty in adequately encod-ing domain knowledge. Using imperfect domain knowledge, fully automated systems that abstract away the 'scruffy' real world tend to produce fragile schedules that omit important constraints and optimize artificial metrics. As a result, these systems are ultimately often rejected by the user. We de-scribe the design of a user-centric scheduling system, Pisces, that assists the user in exploring the rich space of schedules in complex, real-world domains with multifaceted objectives. Pisces retains the strength of humans in understanding sched-ule quality and nuances of domain constraints, while leverag-ing the power and flexibility of constraint-based scheduling algorithms. The system helps the user to iteratively craft a so-lution by expressing both high-level guidance and low-level specific constraints and preferences.