In this paper, a design method of nonlinear controllers is proposed. While physical or chemical analysis is a reliable approach for the construction of nonlinear models, it needs many experiences and might result in very complicated models. The Black-box model approach is adopted in this paper. Identification and controller design of nonlinear processes are very troublesome and complicated. Therefore, a direct design method of nonlinear control, which means not to identify process models, is proposed. The controller is designed using the idea of VRFT (virtual reference feedback tuning). In order to deal with the nonlinearity, input-output data in various operating conditions must be obtained. However, it is very difficult to get sufficient quantities of data for identification of the nonlinearity. Therefore, the model reliability should be evaluated. If the operating condition was not included in the learning data for the nonlinear model identification, the reliability of the model should be judged low. In such a case, it is proposed to suppress the magnitude of the manipulation change. Moreover, in such a case, the online data are stored and are used to re-design the nonlinear model offline. The effectiveness of the proposed controller was shown in a simulation of a tank system.