High-fidelity 3D clothed human reconstruction plays a pivotal role in applications such as virtual reality, game animation, and digital human interaction. To address the limitations of existing methods in representing garment details, ensuring boundary continuity, and achieving rendering efficiency, we propose a novel reconstruction framework named SharpAvatar, which enables high-quality, interactive clothed human reconstruction and real-time rendering. Our method is built upon Gaussian Splatting and incorporates a semantics-guided initialization strategy, where initial Gaussian point distributions for the body and garments are separately constructed based on human semantic segmentation results. We further propose a hierarchical Gaussian modeling mechanism, which models the body and garments as independent Gaussian layers. Different binding and deformation strategies are applied: body Gaussians are deformed via mesh-based pose-driven binding, while garment Gaussians adopt free-floating or soft-binding strategies to enhance modeling flexibility. During training, we introduce a Surface-Edge Optimization (SEO) module, which leverages a Transformer-based architecture to extract high-response features in edge regions of the point cloud. Combined with a normal-consistency constraint, this module significantly improves detail preservation and surface continuity. Finally, a multi-layer Gaussian rendering strategy is employed, enabling fast inference and high-fidelity reconstruction at 1080p resolution. Experiments on the ZJU-MoCap dataset demonstrate that SharpAvatar outperforms existing state-of-the-art methods in rendering speed, geometric accuracy, and garment structural integrity. Notably, it achieves real-time rendering at 350 FPS with a PSNR of 33.03, showcasing superior performance for digital human modeling tasks.