Most existing multi-view clustering methods assume complete data and overlook feature-level missingness. To address this, we propose TLMVC (Two-level Incomplete Multi-View Clustering), which integrates a feature repair mechanism that captures global data structure using a small set of representative samples. This approach reduces clustering time and enhances robustness to missing features. A weighted fusion strategy, guided by view completeness, adaptively adjusts for varying completeness across views, improving the accuracy of the spectral similarity matrix. Spectral clustering is then applied to the repaired data. Experiments on six benchmark datasets under different missing rates demonstrate the effectiveness and superior performance of TLMVC.