Slop grenades“垃圾手榴弹”（Slop grenades）

A slop grenade is AI output you send to someone without reading it first.“垃圾手榴弹”就是你未经阅读就发送给别人的 AI 输出内容。

Tobi Lütke, the CEO of Shopify, coined the term on The Knowledge Project podcast with Shane Parrish. The episode came out on September 15, 2026.Shopify 首席执行官 Tobi Lütke 在 Shane Parrish 的播客《The Knowledge Project》中创造了这个词。该集节目于 2026 年 9 月 15 日发布。

About sixteen minutes in, Shane asked if AI had made anything worse inside Shopify. Tobi said:在大约十六分钟时，Shane 问 AI 是否让 Shopify 内部的某些事情变得更糟了。Tobi 说：

The failure case of lazy work is not lack of output. It’s actually over-output. Internally we have come to call these things that people are lobbing “slop grenades” at each other, which I think is a really fun term that we should push into industry.懒惰工作的失败案例并非产出不足，实际上是产出过剩。在内部，我们开始把人们互相投掷的这些东西称为“垃圾手榴弹”，我觉得这是一个非常有趣的术语，我们应该在行业内推广。

What it looks like它是什么样子的

At Shopify, up to about half of all pull requests now start as conversations in the company’s shared chat with River, its internal AI agent.在 Shopify，现在多达一半的拉取请求都是从公司共享聊天工具中与内部 AI 代理 River 的对话开始的。

You ask River for a change. It creates a pull request, you approve it without reading it, and now a colleague must find everything that is wrong.你让 River 进行修改，它创建一个拉取请求，你未经阅读就批准了它，现在同事必须找出其中所有错误的地方。

The same happens with email. You have one point to make, but an LLM turns it into a long message. The other person then asks another LLM to summarize it.电子邮件也是如此。你只有一个观点要表达，但大语言模型（LLM）却把它变成了一篇长文。对方随后又要求另一个大语言模型对其进行总结。

Tobi called this decompression followed by recompression. We inflate one thought, then use another model to shrink it again.Tobi 将此称为解压后再压缩。我们膨胀了一个想法，然后使用另一个模型将其再次缩小。

The work moves to someone else工作被转移给了别人

You generate something in seconds. Someone else spends twenty minutes checking it.你在几秒钟内生成了内容，而别人却要花二十分钟去检查它。

Earlier in the podcast, Tobi said machines cannot take responsibility. Humans can use machines to make better decisions, but the decision still belongs to a person.在播客的前半部分，Tobi 说机器无法承担责任。人类可以使用机器做出更好的决策，但决策权仍属于个人。

If my name is on a message or pull request, I should know what it contains. Sending raw AI output makes the receiver responsible for checking my work.如果我的名字出现在消息或拉取请求上，我就应该知道它包含什么。发送原始的 AI 输出会让接收者负责检查我的工作。

I use coding agents most of the day, so I can also throw slop grenades at myself. If I accept an agent’s report without checking the result, I might discover the problem a week later, when I no longer remember why the code changed.我大部分时间都在使用编码代理，所以我也会向自己投掷“垃圾手榴弹”。如果我不检查结果就接受了代理的报告，我可能会在一周后才发现问题，那时我已经不记得为什么代码会发生更改了。

Read it before you send it发送前请阅读

The fix is boring. Read the output.解决方法很无聊：阅读输出内容。

If someone asked a direct question, answer it first. A chat reply should look like a chat reply. If you need a document, write a short answer and link to it.如果有人问了直接的问题，先回答它。聊天回复应该看起来像聊天回复。如果你需要一份文档，写一个简短的回答并附上链接。

For code, run the result and inspect the diff. Asking a colleague to review a pull request does not replace your own review.对于代码，运行结果并检查差异（diff）。让同事审查拉取请求并不能取代你自己的审查。

I also ask agents to report in a fixed shape: files changed, tests run, decisions, uncertainties, and review points. This is in my AGENTS.md. It gives me something concrete to check.我还要求代理以固定的格式报告：已更改的文件、已运行的测试、决策、不确定因素和审查要点。这些都在我的 AGENTS.md 中。这为我提供了一些具体的可检查内容。

If I use an LLM to help with a message, I ask it to make my point clearer and shorter. I do not ask it to make the message look more substantial.如果我使用大语言模型来辅助撰写消息，我会要求它让我的观点更清晰、更简短。我不会要求它让消息看起来更充实。

I wrote more about cleaning up AI output in why I use anti-slop skills.我曾在《为什么我使用反垃圾（anti-slop）技能》一文中更多地写到了关于清理 AI 输出的内容。

Want me to talk about your product? You can sponsor this site.想让我谈谈你的产品吗？你可以赞助本网站。