摘要：

目的 构建面向学术质量控制的智能审校辅助平台，探索大语言模型（large language model，LLM）的氛围编程（vibe coding）模式在医学科技期刊编辑中的应用可行性。 方法 采用氛围编程模式，由不具备专业编程背景的编辑人员以自然语言提示驱动AI代码生成，结合Coze扣子编程平台与Doubao Seed 1.8大模型，开发了涵盖文字、术语、逻辑、统计分析、数据及参考文献6大维度的医学论文多维度审核平台及参考文献真实性核查工具RefCheck Pro。以10篇医学论文为样本，平行对比专业人工审校与AI审核在覆盖广度、问题类型及互补性方面的表现。 结果 在文字、术语、逻辑、统计分析及数据维度，人工审校共1151条分类标注（1235条独立批注），AI审核生成583条独立批注。双方均准确识别拼写、标点等常见问题（共197项）；人工审校在语言表达细节（如口语化、图表题名不匹配）上优势明显，AI在统计分析、公式符号缺失、单位换算遗漏等技术细节上表现更优。在参考文献维度，对10篇样本论文的404篇参考文献进行审核发现，“LLM格式审核+RefCheck Pro真实性核验”模式审核与人工审校结果高度一致。在审校论文的时效性方面，人工审校每篇耗时约2个工作日，AI审核仅需3~4 min。 结论 氛围编程模式可显著降低非技术背景编辑人员构建AI工具的门槛，实现从“概念”到“可用系统”的快速迭代，或为科技期刊编辑智能化工具的开发提供可参考的实践路径。同时，人机协同审稿在效能上优于单一方式，“AI初筛-人工精审”的双层质量控制体系在统计规范性审查和参考文献真伪辨识方面价值尤为突出。

关键词: 氛围编程, 大语言模型, 质量控制, 智能审校, 人机协同

Abstract:

Purposes To develop an intelligent proofreading assistance platform for academic quality control and to explore the feasibility of applying large language model （LLM） vibe coding mode in the editorial workflow of medical scientific journals. Methods Using the vibe coding mode， editors without professional programming background drove AI code generation through natural language prompts. A multidimensional review platform for medical manuscripts，overing text， terminology， logic， statistical analysis， data， and references，was developed alongside RefCheck Pro， a tool for verifying reference authenticity， by integrating the Coze programming platform and Doubao Seed 1.8 LLM. A total of 10 medical manuscripts was used as test samples to compare professional editors’ manual review and AI review in terms of coverage breadth， types of issues detected， and complementarity. Findings In the dimensions of text， terminology， logic， statistical analysis， and data， manual review yielded 1151 categorized annotations （1235 independent comments）， while AI review produced 583 independent comments. Both approaches accurately identified common issues such as spelling errors and punctuation standardization （a total of 197 items）. Manual review demonstrated significant advantages in detecting subtle linguistic issues （e.g.， colloquial expressions， mismatched chart and figure titles）. AI review performed excellently in statistical analysis and independently identified technical details easily overlooked by humans， such as missing formula symbols and omitted unit conversion labels. In the reference dimension， the review of 404 references in 10 sample papers showed that the combined “LLM-based formatting check+RefCheck Pro authenticity verification” approach demonstrated high concordance with manual review. In terms of the timeliness of paper review， manual review took about 2 working days per paper， while AI review only took 3~4 minutes. Conclusions The vibe coding mode markedly lowers the technical barrier for editorial staff to construct professional AI tools， enabling rapid iteration from concept design to deployable system， and can provide practical reference for the development of intelligent tools for sci-tech journal editing. Meanwhile， human-AI collaborative reviewing demonstrates superior efficiency compared to either approach alone， and the two-layer quality control system of “AI preliminary screening-manual detailed review” is particularly valuable in statistical standardization review and reference authenticity identification.

Key words: Vibe coding, Large language model, Quality control, Intelligent proofreading, Human-machine collaboration