<p>In instance-based learning algorithms, storing numerous training examples creates high memory demands, sensitivity to noise, and slower execution. Instance selection addresses these issues by retaining a compact subset that removes redundancy and noise while preserving essential structure. Because the value of a training example depends on its neighborhood, reduced sets should include representatives from safe and ambiguous subregions. This work introduces the Hierarchical Growth Self-generating Prototype (HSGP), an entropy-guided instance-selection algorithm framed through decision-tree principles. HSGP recursively partitions the data by impurity, grows low-uncertainty regions, and synthesizes provisional prototypes. Each prototype serves only as an anchor and is replaced by its nearest real training instances, preserving local density and label fidelity while avoiding artifacts from synthetic points. To prevent over-refinement, HSGP applies a stopping rule based on the dispersion of subset entropies; partitioning halts when uncertainty becomes uniformly low across the active subsets, curbing unnecessary splits and maintaining a favorable balance between reduction and accuracy. The algorithm selects examples from distinct and overlapping class regions, capturing dataset complexity while retaining interpretability. Experiments on twenty-one real-world classification problems show competitive or superior predictive accuracy relative to 1-Nearest Neighbor trained on the full dataset and favorable performance against established instance-selection methods, while delivering substantial reductions in set size. Overall, entropy-guided growth with anchor-and-replace selection offers an effective strategy for building compact training sets that preserve decision-relevant structure.</p>