This study focuses on the “Connotative Quality Control of Admission Records in Medical Records by Large Language Models” task at the China Health Information Processing Conference (CHIP) 2025, under the non-fine-tuning track. The task involves predicting connotative quality control defects using 300 unannotated admission records each for validation and testing. We employ the Qwen3-30B-A3B-Thinking-2507 model without fine-tuning and propose a dual-enhancement framework integrating precision in-context learning and Chain-of-Thought (CoT) reasoning. The framework first applies semi-structured segmentation to extract rule-relevant content from medical records, reducing irrelevant information. It then introduces a stepwise judgment mechanism based on CoT to decompose abstract quality control logic into executable reasoning steps. Results are output in structured JSON format to ensure interpretability. Experimental results show that our method achieved 2nd place on the test set (final ranking) with a comprehensive score of 0.6644. This work demonstrates the feasibility of applying non-fine-tuned LLMs enhanced by precision in-context learning and CoT reasoning to medical record connotative quality control, offering a practical and referential approach for intelligent medical record quality assurance.