展示HN：Kokki – 一种“双核”系统提示，以减少大型语言模型的幻觉现象
我构建了一个系统提示架构，迫使大型语言模型（如 Gemini 3.0 / Claude 4.5）在回应之前进行自我审查。<p>该架构将模型分为两个角色：<p>代理：生成草稿。<p>审计者：对草稿进行逻辑错误和幻觉的批评。<p>代码（摘要）：[<a href="https://gist.github.com/ginsabo/9d99ee98068d904b214be9351f09323f" rel="nofollow">https://gist.github.com/ginsabo/9d99ee98068d904b214be9351f09...</a>]<p>我还写了一本关于该理论的文档书，但上述摘要是完全可用且免费的。我很想听听你对这种“用户端安全”方法的反馈。
查看原文
I built a system prompt architecture that forces LLMs (like Gemini 3.0 / Claude 4.5) to self-audit before responding.<p>It splits the model into two personas:<p>Agent: Generates the draft.<p>Auditor: Critiques the draft for logic errors and hallucinations.<p>The Code (Gist): [<a href="https://gist.github.com/ginsabo/9d99ee98068d904b214be9351f09323f" rel="nofollow">https://gist.github.com/ginsabo/9d99ee98068d904b214be9351f09...</a>]<p>I also wrote a documentation book about the theory, but the Gist above is fully functional and free to use. I'd love to hear your feedback on this "User-Side Safety" approach.