AI planners typically use a set of generalized operator descriptions as a representation of the possible domain actions. These operators are then selected and instantiated to accomplish particular task goals. However, in complex domains it may be difficult to enumerate all of the operators, many of which may contain redundant information. Adapting the planner to related domains or to changes in the domain properties may require extensive modifications to update and maintain the operator descriptions. A more maintainable representation is required for complex domains. This paper proposes a technique called Operator Construction which can be applied to physical planning domains. In this approach, instead of representing the possible domain actions as a set of generalized operators, knowledge about domain objects that generate actions is used to generate operators. The result is better suited for complex domains because Operator Construction offers the advantage of ease of maintenance, and a more direct link between domain properties and possible actions. This teclmique has been used in a complex manufacturing planning domain.