A large-scale dataset is essential for learning good features in 3D shape understanding, but there are only a few datasets that can satisfy deep learning training. One of the major reasons is that current tools for annotating per-point semantic labels using polygons or scribbles are tedious and inefficient. To facilitate segmentation annotations in 3D shapes, we propose an effective annotation tool, named iSeg for 3D shape. It can obtain a satisfied segmentation result with minimal human clicks (< 10). Under our observation, most objects can be considered as the composition of finite primitive shapes, and we train iSeg3D model on our built primitive-composed shape data to learn the geometric prior knowledge in a self-supervised manner. Given human interactions, the learned knowledge can be used to segment parts on arbitrary shapes, in which positive clicks help associate the primitives into the semantic parts and negative clicks can avoid over-segmentation. Besides, We also provide an online human-in-loop fine-tuning module that enables the model perform better segmentation with less clicks. Experiments demonstrate the effectiveness of iSeg3D on PartNet shape segmentation. Data and codes will be made publicly available.