The growing volume of time series (TS) data across diverse fields presents both opportunities and challenges for efficient and practical classification that can give us interpretable and in-time insight into the data. This paper introduces Time Series Step Tree (TSST), a novel and inherently interpretable method for prompt TS classification. Addressing the growing need for both timeliness and transparency in time-sensitive applications, TSST employs a progressive, step-wise evaluation of TS data with dynamically adjusted observation windows. At the core of TSST is a decision tree constructed using a novel “witness time series” (WITESS-TS) selection process at each node to maximize information gain, ensuring both classification efficacy and interpretability. Experimental results on univariate TS datasets demonstrate TSST’s ability to achieve high accuracy while enabling prompt classification decisions. The inherent interpretability of the method is further highlighted through decision tree visualization and analysis. Future work will focus on extending TSST to multivariate TS and further empirical validation across diverse datasets.