Black-box modeling for system identification and control can be constrained by the geometric scale of a plant. Those constrains arise from a technical or economical issue in obtaining input-output data sets for training purposes. A solution is sequential adaptation, which must be achieved in real time in the case of control systems, because a prescribed level of performance is required while external excitations are occurring. A major challenge is the proper selection of inputs that would lead to a quick and optimal adaptation. This paper proposes a new method for selection of inputs in black-box modeling. The method, termed the Self-Organizing Inputs (SOI) algorithm, consists of organizing the input space sequentially and in real-time. This self-organization leads to an enhanced representation of the system dynamics, which improves adaptation convergence and performance. The algorithm is designed based on time series analysis, and on the theory that dynamic systems can be topologically reconstructed using a single observation delayed in time, provided the choice of time delay and embedding dimension is appropriate. The SOI feature sequentially identifies the proper time delay and dimension, and adapts the required dynamic inputs for the black-box model. That allows a representation to self-organize, self-adapt, and identify or control by utilizing the essential dynamics of the plant captured via limited state measurement.