Few-shot document-level relation extraction (FSDLRE) aims to identify semantic relations among entities in a document with only limited labeled data. Existing prototype-based meta-learning methods construct class prototypes for matching but face two key limitations: (1) Inadequate None-of-the-Above (NOTA) modeling—by focusing solely on “learning-to-match,” they fail to build robust representations for NOTA cases; (2) Underutilized supervision and reasoning—they insufficiently leverage multi-view supervisory signals and overlook the structured reasoning capabilities of large language models (LLMs), thereby limiting cross-domain adaptability under scarce labels. To address these issues, we propose Chain-of-Thought with Discriminative Multi-view Prototype Tuning (CDMPT), a novel framework that leverages the discriminative power of LLMs to mine supervisory signals from limited annotations through multiple complementary views. By integrating the chain-of-thought mechanism, CDMPT learns to discriminate relation semantics, mitigates the challenges of NOTA modeling, and performs structured prototype construction and discriminative reasoning to accomplish FSDLRE. Extensive experiments demonstrate that our approach yields substantial performance gains, particularly in cross-domain scenarios.