Robot foundation models have improved general-purpose manipulation, but adapting them to new tasks still often requires robot demonstrations and additional training. This requirement is particularly restrictive for dexterous hands, whose high-dimensional control makes teleoperation difficult, slow, and expensive. This PhD project investigates how robots can learn new dexterous tool-use tasks at deployment time from one or a few human demonstration videos, with minimal target-task robot data. The project is motivated by a useful asymmetry: for tasks where robot demonstrations are hardest to collect, human videos may be especially informative, capturing functional grasps, finger coordination, and tool-object interactions that provide a rich basis for transfer to anthropomorphic robot hands. Converting these cues into robot actions nevertheless remains difficult: even kinematically similar human and robot hands differ in morphology, dynamics, and scale; critical contacts and finger placements are often occluded; and interaction forces are not directly observable from video. The project will therefore study how task and interaction cues from video can be combined with embodiment-specific priors, contact-aware control, and closed-loop sensory feedback on top of pretrained robot models. The research will focus on contact-rich tool use and evaluate generalization across new tasks, objects, environments, and task compositions. Ultimately, the project aims to make dexterous robots easier for non-expert users to teach without specialized teleoperation or extensive retraining.