Finding effective grasp strategies constitutes one of the main challenges in robotic manipulation, especially when dealing with soft, underactuated, and non-anthropomorphic hands. This work presents a Learning from Demonstration approach to extract grasp primitives using a novel reconfigurable soft hand, the Soft ScoopGripper (SSG). Starting from human demonstrations, we derived Gaussian models through which we were able to devise different grasping strategies, exploiting the SSG features. As the grasping strategies are tightly related to the characteristics of the object to be grasped, we tested two different ways of modeling objects in the training dataset and we comparatively evaluated the resulting primitives. Experimental grasping trials on unknown test objects confirmed the effectiveness of the learned primitives and showed how assuming different levels of knowledge about the object representation in the training phase influences the grasp success.