HackerNews中文版

我一直在尝试使用 NotebookLM，不仅用于总结，还用于对我的自动化工作流程进行“红队”测试。我向它输入了我构建的 n8n 研究代理的视频演示，并使用自定义指令迫使“音频概述”主持人扮演特定角色：“愤世嫉俗的科技评论员”。我的目标是看看该模型是否能够仅从视频 feed 中识别出特定的节点逻辑和架构选择（例如手动处理服务器发送事件和强制 JSON 输出），而不仅仅是总结项目的“要点”。我使用的提示是：“扮演两个愤世嫉俗的科技评论员……只关注最令人印象深刻的单个功能……注意‘服务器发送事件’（SSE）配置……承认 4 层架构是合法的工程设计。” 结果：该模型成功地识别出我在 n8n 中实现的特定超时处理逻辑，而标准总结通常会忽略这一点。如果你提示负面评价而不是赞同，这似乎是一种对架构流程进行“合成第二意见”的可行方法。输出视频（4 分钟）：https://youtu.be/oof9JB3OFO4 还有其他人成功地在 NotebookLM 中强制使用特定的技术角色吗？

查看原文

I've been experimenting with using NotebookLM not just for summarization, but for "Red Teaming" my own automation workflows.I fed it a video demo of an n8n research agent I built and used Custom Instructions to force the Audio Overview hosts into a specific persona: "Jaded, cynical tech reviewers."My goal was to see if the model could pick up on specific node logic and architectural choices (like manual Server-Sent Events handling and JSON output forcing) from the video feed alone, rather than just summarizing the "gist" of the project.The Prompt I used: "Act as two jaded, cynical tech reviewers... Focus ONLY on the single most impressive feature... Notice the 'Server-Sent Events' (SSE) configuration... Admit that the 4-Layer Architecture is legitimate engineering."The Result: The model successfully identified the specific timeout handling logic I implemented in n8n, which a standard summarization usually misses. It seems like a viable way to get a "synthetic second opinion" on architectural flows if you prompt for hostility rather than agreement.Video of the output (4 min): https://youtu.be/oof9JB3OFO4Has anyone else had success forcing specific technical personas in NotebookLM?

实验：使用 NotebookLM 作为愤世嫉俗的代码审查员（通过自定义提示）