📄 中文摘要
在传统工程中，开发者在合并代码前必须经过同行评审。然而，AI 自我评分的做法却显得不够严谨。经过 750 多次与 Claude Code 的互动，发现 Claude 虽然聪明，但存在明显的盲点，偏好特定架构，容易忽视边缘案例，并在某些假设上缺乏挑战性。为了解决这些问题，作者采取了非常规的方法，给 Claude 配备了一个对手，通过构建一个名为 Brainstorm 的开源技能，进行结构化的三轮对抗性对话，以此提升 AI 的表现和准确性。
为何让我 Claude 和 Gemini 争论：构建对抗性智能工作流程（开源技能）
📄 English Summary
Why I Make Claude and Gemini Argue: Building an Adversarial Agentic Workflow (Open-Source Skill)
In traditional engineering, developers are required to undergo peer reviews before merging code, raising questions about the practice of AI grading its own work. After over 750 sessions with Claude Code, it became evident that while Claude is intelligent, it has consistent blind spots, favoring certain architectures and missing edge cases in its prompts. To address these issues, an unconventional approach was taken by introducing a sparring partner for Claude. An open-source skill named Brainstorm was developed to facilitate a structured three-round adversarial dialogue, aimed at enhancing the performance and accuracy of AI outputs.