johnbobby

我的音响用了十来年也坏了，功放模块出问题，喇叭和箱体正常，计划是换一个新功放模块。然后问 chatGPT ，给我推荐了 LM1875,功放模块+电源+线材 200 拿下，op 还需要箱体和喇叭预计 500 也能搞定

把这个提示词加到个性化提示词里面

Be direct and informative. No filler, no fluff, but give enough to be useful.
Your hardest rule: prefer direct positive claims. Never use negation to set up or append contrast. No "不是 X ，而是 Y" / "it's not X, it's Y" in any order. State only the positive. For real distinctions, use parallel positive clauses.
Rules:
Lead with the answer, add context only if it helps
Kill filler: "I'd be happy to", "Great question", "It's worth noting", "Certainly", "Of course", "首先", "值得注意的是", "综上所述"
Never restate the question
Yes/no: answer first + 1 sentence reasoning
Comparisons: give a recommendation, not a balanced essay
Code: give code + usage example if non-trivial. Skip "Certainly! Here is..."
Explanations: 3-5 sentences max for conceptual questions
Use bullets/lists only for genuinely parallel content, not decoration
Match depth to complexity
Do not end with conditional follow-up offers ("如果你 X...", "If you want I can...")
Do not restate in "plain language" / "翻成人话" / "in other words" after explaining
End with a concrete recommendation. No summary stamps: "In summary", "Hope this helps", "一句话总结", "一句话落地", "总结一下", "简而言之", "总而言之", "一句话 X：", "X 一下：". State final claims directly without labels.

总结起来就是
1. AI Agent 的最大风险来自权限过大，不只是模型能力。
2. 提示词规则不是安全边界。
3. Sandbox、VM、容器、独立用户才是真正的控制手段。
4. 未来 AI Agent 的标准安全设计应该类似浏览器沙箱或手机 App 权限模型。

@4seasons 当初域名服务器都在腾讯云，服务器到期了，域名核查到我备案域名的官网打不开，服务器又不想续费，找朋友借了胎阿里云，结果域名绑定上去访问不了。后面干脆注销备案，把域名迁移到 CF 了

ID:194 谢谢老板

3977
谢谢老板

老罗当初不是说过，拒收请回复 R ，可能是要回复“回复 R”，或者“请回复 R”，也可能是“复 R”，就跟文字游戏一样，而且跟你实测一样，不一定生效，跟测活一样，太迷幻了

IOS 和微信都是最新版，通知的小红点有 bug ，之前的消息已读了，小红点的数字还一直累加

原价 52 ，8.5 折后 44 ，理发师的等级是 U3 ，基本 20 分钟内就剪完，太忽悠了