Robotic therapy enables mass practice of complex hand movements after stroke, but current devices generally enforce patients to reproduce prescribed kinematic patterns using rigid actuators, without considering individuals' unique impairment characteristics, thereby reducing their efficacy. In this paper, we tested the feasibility of a novel, theory-based "biomimetic" approach to restoring mechanics of complex hand tasks with subject-specific assistance patterns. Twelve chronic stroke survivors performed two simulated functional tasks: hand open and simulated pinch task (distal pad press). Assistance was provided by non-restraining actuators (exotendons) that counteracted 'subject-specific' impairments, identified during unassisted task performance. There was no constraint of movement to predefined patterns. Assistance patterns required to complete tasks were significantly different across subjects, reflecting high variability in impairment and required assistance patterns. For hand open, range of motion and interjoint coordination were significantly improved for severely impaired patients, while movement quality was enhanced (reduction in jerk) for those less impaired. For simulated pinch, subject-specific assistance restored task mechanics before injury, as patients were able to direct fingertip force toward the direction normal to surface; angular deviation reduced from 16.8°±10.4° to 3.7°±2.6°. Notably, electromyography data confirmed that subjects maintained an effort level under assistance comparable to unassisted conditions. The proposed method could lead to a novel paradigm for hand rehabilitation that restores complex task mechanics with a subject-specific assistance reflecting individual impairment characteristics while promoting subjects' participation.