Brain parcellation schemes are fundamental to neuroimaging, yet general-purpose atlases may obscure the specific functional architecture relevant to a given cognitive task or clinical condition. This reflects a growing consensus that the “optimal” brain map is context-dependent. Here, we introduce a novel framework that validates this principle by generating task-optimized human brain parcellation maps directly from supervised learning objectives. Our method defines functional parcels by grouping brain regions based on the similarity of their contributions to a classifier’s decision boundary for a specific goal (e.g., cognitive state decoding or clinical group separation). This approach prioritizes a region’s discriminative role over simple signal homogeneity or spatial contiguity. We demonstrate that these objective-driven parcellations reveal a latent functional organization of the brain, an implicit task-relevant architecture defined not by signal homogeneity but by the shared discriminative role of brain regions. On Human Connectome Project data, our parcellations significantly improved cognitive state decoding, and on ADNI data, they enhanced Alzheimer’s Disease classification. Beyond improving accuracy, the resulting parcellations exhibited unique neurobiological properties: they identified spatially coherent, high-resolution maps of task-relevant information that were obscured by standard atlases and showed a trade-off between task-specificity and overall signal homogeneity. These optimized maps generalized across independent datasets, highlighting that they capture robust principles of task-dependent brain organization. This work provides a framework for moving beyond universal atlases, enabling the generation of context-specific brain maps that offer a new window into the functional architecture underlying specific cognitive processes and disease states.