# Compound Engineering - 4/28/2026
**作者**: Trevin Chow
**日期**: 2026-04-28T18:48:46.000Z
**来源**: [https://x.com/trevin/status/2049199159073550790](https://x.com/trevin/status/2049199159073550790)
---
We shipped v3.3.0. This release does 2 things: it makes the agent show its thinking before acting, and it interrupts you less while doing real work.
## Show me what you think I asked for, before you spend 10 minutes on it
> "You keep using that word. I do not think it means what you think it means."— Inigo Montoya, The Princess Bride
In an earlier version of /ce-plan, the skill could happily spend 5+ minutes producing a polished implementation plan built on one or two incorrect assumptions, and you'd have no idea. By the time the doc lands, those wrong assumptions are baked into a dozen places: painful to fix, and easy to miss entirely.
The marquee change lives in /ce-brainstorm and /ce-plan. Both skills now pause before doing real work and play back what they heard you ask for. Tell /ce-plan to "add OAuth login to the mobile app," and the skill comes back with three short lists:
1. what you literally said (Stated)
2. what it's quietly assuming you also want (Inferred: Google and GitHub providers, a callback handler, user-table changes)
3. what it's deliberately leaving alone (Out-of-scope: existing email login, two-factor).
You scan the bullets, correct the misreads in chat, and the skill keeps going. The cost of catching scope drift drops from "rerun the whole skill" to "edit a bullet."
The same shift shows up in /ce-ideate, where it used to chase the wrong interpretation on minimal prompts. The skill now figures out when to ask scoping questions up front, and tags every surviving idea with its grounding: direct evidence in your code, external prior art, or reasoned argument. The goal is to push the agent to be specifically wrong or specifically right. Less hedging, no speculation without a basis. ce-ideate is still one of my most used skills these days across my projects and this change has made it even better.
## Stop creating ceremony where the agent's judgment is good enough
A lot of recent work has been making the skills quieter, and making them better when they do need to interrupt you.
/ce-doc-review was relentless about asking for your input, even when there was only one obvious choice. It's now much better at applying fixes automatically and grouping related issues together, so the attention it does ask for goes toward decisions that genuinely need it.
/ce-code-review used to defer most fixable findings to your ticket tracker — also overboard in most cases. It now fixes more of them automatically, so you don't carry weight forward unnecessarily.
The goal of this release: make the agent cheaper to correct and faster to trust.
https://github.com/EveryInc/compound-engineering-plugin
## 相关链接
- [Trevin Chow](https://x.com/trevin)
- [@trevin](https://x.com/trevin)
- [5.7K](https://x.com/trevin/status/2049199159073550790/analytics)
- [v3.3.0](https://github.com/EveryInc/compound-engineering-plugin)
- [https://github.com/EveryInc/compound-engineering-plugin](https://github.com/EveryInc/compound-engineering-plugin)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [2:48 AM · Apr 29, 2026](https://x.com/trevin/status/2049199159073550790)
- [5,747 Views](https://x.com/trevin/status/2049199159073550790/analytics)
- [View quotes](https://x.com/trevin/status/2049199159073550790/quotes)
---
*导出时间: 2026/4/29 09:31:59*
---
## 中文翻译
# Compound Engineering - 2026年4月28日
**作者**: Trevin Chow
**日期**: 2026-04-28T18:48:46.000Z
**来源**: [https://x.com/trevin/status/2049199159073550790](https://x.com/trevin/status/2049199159073550790)
---
我们发布了 v3.3.0 版本。这个版本做了两件事：让智能体在行动前先展示其思考过程，并在它处理实际工作时减少对你的打扰。
## 在你花十分钟去做之前，先给我看看你对我要求的是什么
> “你一直在用那个词。我不认为它表达的是你心目中的意思。” —— 英奇·蒙托亚， 《公主新娘》
在较早版本的 `/ce-plan` 中，这项技能可能会愉快地花费 5 分钟甚至更长时间，基于一两个错误的假设生成一份看似完美的实施方案计划，而你却一无所知。等到文档生成时，那些错误的假设已经嵌入到十几个地方：既痛苦难修，又容易被彻底忽略。
最重要的改动体现在 `/ce-brainstorm` 和 `/ce-plan` 中。这两个技能现在会在开始实际工作前暂停，并复述它们听到你要求的内容。让 `/ce-plan` “给移动应用添加 OAuth 登录”，该技能会反馈三个简短的列表：
1. 你字面上所说的内容（所述内容）
2. 它默认假设你也想要的内容（推断内容：Google 和 GitHub 提供商、回调处理器、用户表变更）
3. 它刻意不涉及的内容（范围之外：现有的邮箱登录、双重认证）。
你浏览这些要点，在聊天中纠正误解之处，然后该技能继续执行。发现范围偏差的成本从“重新运行整个技能”降低到了“编辑一个要点”。
同样的转变也出现在 `/ce-ideate` 中，以前它面对极简提示时常常会跑偏。该技能现在能判断何时需要先提出界定范围的问题，并且会给每个保留的想法贴上其依据标签：你代码中的直接证据、外部先例，或推理论证。目标是推动智能体做到要么具体地错，要么具体地对。少一些含糊其辞，不做毫无根据的推测。这些天来，ce-ideate 依然是我的项目中使用频率最高的技能之一，而这次改动让它变得更好用了。
## 停止在智能体的判断力足够好的地方搞繁文缛节
最近的大量工作都在让这些技能变得更“安静”，并在它们确实需要打扰你时表现得更好。
`/ce-doc-review` 曾经不厌其烦地征求你的意见，即使只有一个显而易见的选择也是如此。它现在更擅长自动应用修复并将相关问题归类，因此它所请求的关注点真正用于那些确实需要人工决策的地方。
`/ce-code-review` 以前会将大部分可修复的发现提交到你的工单追踪系统——这在大多数情况下也有些过头了。它现在会自动修复其中的更多问题，这样你就不必背负不必要的负担。
本版本的目标：让智能体更易于纠错，并更快地值得信赖。
https://github.com/EveryInc/compound-engineering-plugin
## 相关链接
- [Trevin Chow](https://x.com/trevin)
- [@trevin](https://x.com/trevin)
- [5.7K](https://x.com/trevin/status/2049199159073550790/analytics)
- [v3.3.0](https://github.com/EveryInc/compound-engineering-plugin)
- [https://github.com/EveryInc/compound-engineering-plugin](https://github.com/EveryInc/compound-engineering-plugin)
- [升级到 Premium](https://x.com/i/premium_sign_up)
- [2026年4月29日 上午2:48](https://x.com/trevin/status/2049199159073550790)
- [5,747 次观看](https://x.com/trevin/status/2049199159073550790/analytics)
- [查看引用](https://x.com/trevin/status/2049199159073550790/quotes)
---
*导出时间: 2026/4/29 09:31:59*