A practical knowledge-based operator support system is being developed for Japanese pressurized water reactors. This system will be implemented at real power plants in the near future. The difficulty of realizing a practically usable system using normative models based on deep knowledge is discussed. Instead of adopting the normative approach, the system introduces a hierarchically organized model, called a “plant abnormality model,” to its diagnosis part. With its ability to envelope unforeseen events, it avoids the use of imperfect deep knowledge and the scenario dependability that is considered an inherent problem in abnormality models. Existing operational procedures are broken down into functionally independent task units and specified as knowledge sources for operational guidance. Depending on the plant status, relevant task units are dynamically integrated to synthesize operational procedures that are provided as operational guidance. Estimated information on unobservable or predictive plant status is used to enable flexible and timely synthesis of the procedures. An attempt is made to organize the information so that it is better understood by the operators by adopting the hypothesis-and-test scheme as a framework for the inference control mechanisms of the diagnosis system.