# 9 Rules Lyft Co-Founder Matt Van Horn Uses To Build With Claude Code
**作者**: Tom
**日期**: 2026-06-20T20:09:47.000Z
**来源**: [https://x.com/tomcrawshaw01/status/2082450077709381777](https://x.com/tomcrawshaw01/status/2082450077709381777)
---
He hit number one on GitHub Trending without ever learning to read code. These are the nine rules he follows instead.
Matt Van Horn co-founded the company that became Lyft, then built June, the self-driving oven Weber bought. He also says plainly that he does not read or write code. Then he released Last 30 Days, the first open-source project of his life, and it went to number one on GitHub Trending.
I run Last 30 Days every day, which is how I found him. Here is what he told me in thirty minutes on a livestream, and almost none of it matches what you will read about building with AI this month.
## Rule one: build what you personally need, and ship it the same day
Last 30 Days started as a workaround. He was pasting long prompts into ChatGPT telling it to look only at Reddit, and only at the last thirty days, because he knew OpenAI had paid for the Reddit rights.
He was trying to write better prompts, nothing more.
Then the snow came in badly on a day his daughter was up the mountain. He decided it was not worth skiing, sat down, and shipped it that afternoon. The launch video did a million views.
## Rule two: write a plan for everything, because the model is lazy by default
Matt's reason is less flattering to the models than the usual planning advice. They are inherently lazy, in his words, and they want to solve your problem with the least work possible.
A plan.md fixes that structurally rather than through better prompting. It forces the agent to research first and hands it a task list it cannot skip.
He runs compound engineering for this, specifically ce plan and ce work. He plans holidays the same way, down to where to eat.
## Rule three: never open the plan your agent just wrote
Three months ago Matt would have told you the opposite. Open it in your editor, read it properly, because your agent is about to build exactly that.
He has reversed completely. You should never write code, never read code, and never read plans.
Roughly one time in fifty he should have read it. He thinks the other forty-nine make reading it a bad trade anyway.
What he does instead is interrogate the plan without ever loading it himself. Give me the TLDR. Explain it like I am five. Did you include what I asked for.
The plan is for agents. You wrote it for something else to read.
He states the condition himself. It only holds if you trust the model and the system you picked.
## Rule four: stop arguing about loops, because a goal is a loop
The word loop got noisy fast. Peter Steinberger and Boris Cherny were at the centre of it, and plenty of the people saying they just do loops could not define the term.
Matt's article "WTF is a loop" did 3.4 million views and remains the biggest thing he has published.
> **Matt Van Horn@mvanhorn**: [原文链接](https://x.com/mvanhorn/status/2068426104088748331)
>
His definition takes one sentence. Replace the word loop with goal, because a goal is the loop.
He runs two commands. ce plan builds the plan, then /goal sets the agent working at it without him steering, and he says the Codex version feels more aggressive than the one in Claude Code.
## Rule five: live in Claude Code and send the work to Codex
What Matt optimises is usage, not quality. He pays for the $200 plan on both sides.
His read is that Codex gives him about four times the usage for comparable work, and he says that is what it feels like rather than something he measured.
So he works inside Claude Code and pushes as much real coding into Codex as he can, through an OpenAI plugin almost nobody uses.
He will not say either model is better. He has been marketed to heavily about Fable 5 and cannot prove the difference in his own usage.
## Rule six: give your agent a CLI, not an MCP server or an API
Ask an agent for a baseball score without the right tool and you can watch it struggle. It opens Google, then the team site, then ESPN, hunting for the baseball button and burning tokens the whole way.
His framing for the fix is the clearest thing he said all conversation. A CLI is Google Maps for agents.
He built CLI Printing Press with Trevin to produce them at scale. ESPN has no official API, so an agent sniffed the site, found the undocumented endpoints and mapped the patterns.
They launched with around fifty. There are more than 350 now.
Having a CLI does not mean having a good one. Most companies wrap their API endpoints and call it finished, and that is not the same as designing for an agent.
## Rule seven: let the tool leave notes for itself
When they built the ESPN CLI there was no World Cup data on the site. Nothing in it could have told an agent the competition existed.
Asked for the USA score, it hit walls until it found the endpoint, then wrote itself a note recording where the data lived.
The second run is fast because the first run was slow and the tool remembered. Most setups throw that away every session.
## Rule eight: ask a research agent which tool to use before you commit to one
Matt wanted Gmail and Sheets access for his agents. His choice was Google's official CLI, or one Peter Steinberger had built months earlier.
He assumed Google's would win. He asked Last 30 Days, and it told him to use Peter's, because it was built for power users and built for agents.
He is straight about the limits of that answer. He has never looked at Google's official CLI himself, so it is a recommendation he acted on rather than a test he ran.
Peter's CLI is gws, and I use it every day for Drive, Docs, Sheets and Gmail.
## Rule nine: keep one task the models still fail, and re-run it on every release
Matt asked Codex to build the mathematically most perfect font ever made for the terminal.
Ninety minutes later it came back with four options, three its own and one a real font, and asked him to spot the difference. Three were trash. He chose the good one, and the good one was the real font.
He binned all three. Then he decided the failure was worth keeping, and said he would ask every new model for that same font on the day it launches.
A task today's models fail is a measuring stick you own. Launch posts tell you what a lab wants you to believe. He reckons he needs the 2027 models before that font exists.
## What he thinks you are getting wrong
I asked him what is overhyped right now. He said AI is still underhyped, and what surprises him is how many people are not using it fully.
His advice for anyone who feels behind is not technical. Go and set it up, let it be a disaster, and keep asking why until it clicks.
Not one of these nine rules requires you to learn to code. Most of them require you to stop doing something you currently believe makes you careful.
Matt is @mvanhorn on X and on GitHub. The full thirty minutes livestream is here.
If you want this running in your business and not just on your laptop, I run a free AI audit. We find the one process worth handing to an agent first, then map what it takes to install it. Book one here.
## 相关链接
- [Tom](https://x.com/tomcrawshaw01)
- [@tomcrawshaw01](https://x.com/tomcrawshaw01)
- [14K](https://x.com/tomcrawshaw01/status/2082450077709381777/analytics)
- [Last 30 Days](https://github.com/mvanhorn/last30days-skill)
- [on a livestream](https://www.youtube.com/live/gzavyoPJ--A)
- [plan.md](http://plan.md/)
- [compound engineering](https://github.com/EveryInc/compound-engineering)
- [Jun 21](https://x.com/mvanhorn/status/2068426104088748331)
- [449K](https://x.com/mvanhorn/status/2068426104088748331/analytics)
- [an OpenAI plugin](https://github.com/openai/codex)
- [CLI Printing Press](https://github.com/mvanhorn/cli-printing-press)
- [@mvanhorn on X](https://x.com/mvanhorn)
- [on GitHub](https://github.com/mvanhorn)
- [here](https://www.youtube.com/live/gzavyoPJ--A)
- [Book one here](https://theaiarchitects.com/x/audit/matt-van-horn)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [8:56 PM · Jul 29, 2026](https://x.com/tomcrawshaw01/status/2082450077709381777)
- [14.9K Views](https://x.com/tomcrawshaw01/status/2082450077709381777/analytics)
- [View quotes](https://x.com/tomcrawshaw01/status/2082450077709381777/quotes)
---
*导出时间: 2026/7/30 10:47:51*
---
## 中文翻译
# Lyft 联合创始人 Matt Van Horn 使用 Claude Code 构建项目的九条法则
**作者**: Tom
**日期**: 2026-06-20T20:09:47.000Z
**来源**: [https://x.com/tomcrawshaw01/status/2082450077709381777](https://x.com/tomcrawshaw01/status/2082450077709381777)
---
他从未学过读代码，却在 GitHub 趋势榜上登顶。取而代之的是他遵循的这九条法则。
Matt Van Horn 联合创办了后来成为 Lyft 的公司，随后创立了 June——这款自动驾驶烤箱最终被 Weber 收购。他也直言自己不读也不写代码。然而，他发布了人生中第一个开源项目 Last 30 Days，并且它冲上了 GitHub 趋势榜的第一名。
我每天都运行 Last 30 Days，这也是我发现他的原因。以下是他在三十分钟的直播中告诉我的内容，几乎没有任何一条与你本月读到的关于 AI 构建的文章相符。
## 法则一：构建你自己需要的东西，并在当天发布
Last 30 Days 起初只是一个变通方案。他向 ChatGPT 粘贴冗长的提示词，告诉它只看 Reddit，且只看最近三十天，因为他知道 OpenAI 支付了 Reddit 的数据版权费用。
他当时只是想写出更好的提示词，仅此而已。
后来有一天大雪封山，他的女儿在山上。他决定不值得去滑雪，于是坐下来，并在那个下午发布了这个项目。发布视频获得了一百万次观看。
## 法则二：为所有事情写一个计划，因为模型默认是懒惰的
Matt 给出的理由比通常的计划建议对模型的评价要低得多。用他的话说，它们天生就是懒惰的，总是希望用最少的工作来解决你的问题。
plan.md 在结构上解决了这个问题，而不是通过更好的提示词。它迫使代理首先进行研究，并交给它一个无法跳过的任务列表。
他为此运行“复合工程”，具体是 `ce plan` 和 `ce work`。他用同样的方式规划假期，细致到去哪里吃饭。
## 法则三：永远不要打开你的代理刚刚写好的计划
三个月前，Matt 会告诉你相反的做法。在你的编辑器中打开它，仔细阅读，因为你的代理即将严格据此进行构建。
现在他完全反过来了。你应该永远不写代码，永远不读代码，也永远不读计划。
大概五十次里只有一次他本应该读一下。但他认为，为了那一次而牺牲另外四十九次，无论如何都是一笔糟糕的交易。
他现在做的是在从不亲自加载计划的情况下对计划进行质询。给我摘要。像对五岁孩子一样解释给我听。你包含我要求的内容了吗？
计划是给代理看的。你是为了让别的东西去读才写的。
他自己陈述了这个前提条件：这只有在你信任你选择的模型和系统时才成立。
## 法则四：停止争论什么是“循环”，因为目标就是一个循环
“循环”这个词很快就变得嘈杂混乱了。Peter Steinberger 和 Boris Cherny 处于争论的中心，许多声称自己只做“循环”的人甚至无法定义这个术语。
Matt 的文章《WTF is a loop》（循环到底是什么）获得了 340 万次观看，至今仍是他发表过的传播最广的内容。
> **Matt Van Horn@mvanhorn**: [原文链接](https://x.com/mvanhorn/status/2068426104088748331)
>
他的定义只需要一句话。把“循环”这个词替换成“目标”，因为目标就是循环。
他运行两条命令。`ce plan` 构建计划，然后 `/goal` 让代理开始工作，无需他手动操控。他说 Codex 版本比 Claude Code 中的版本感觉更激进一些。
## 法则五：在 Claude Code 中工作，将工作发送给 Codex
Matt 优化的是使用量，而不是质量。他在两边都订阅了 200 美元的套餐。
他的看法是，对于类似的工作量，Codex 给他的使用量大约是四倍，他说这是一种感觉而非他测量的结果。
所以他是在 Claude Code 内部工作，并通过几乎没人使用的 OpenAI 插件，将尽可能多的实际编码工作推给 Codex。
他不会说哪个模型更好。关于 Fable 5 他看了很多营销宣传，但在他自己的使用中无法证明其中的区别。
## 法则六：给你的代理一个 CLI，而不是 MCP 服务器或 API
如果没有合适的工具就问代理棒球比分，你可以看着它艰难挣扎。它打开 Google，然后是球队网站，接着是 ESPN，一路寻找棒球按钮，同时还在不断消耗 Token。
他对解决方案的表述是整个对话中最清晰的一点。CLI 就是代理的谷歌地图。
他和 Trevin 一起构建了 CLI Printing Press 来大规模生产这些工具。ESPN 没有官方 API，所以一个代理嗅探了网站，找到了未记录的端点并映射了模式。
他们发布时大约有五十个。现在已经有超过 350 个了。
拥有一个 CLI 并不意味着拥有一个好的 CLI。大多数公司只是把它们的 API 端点封装一下就宣布完工了，这与专门为代理设计不是一回事。
## 法则七：让工具为它自己做笔记
当他们构建 ESPN CLI 时，网站上还没有世界杯的数据。其中的任何信息都无法告诉代理这项比赛的存在。
当被问及美国的比分时，它碰壁直到找到那个端点，然后它给自己写了一个笔记，记录了数据所在的位置。
第二次运行很快，因为第一次运行很慢，而且工具记住了。大多数设置在每个会话中都会丢弃这些信息。
## 法则八：在承诺使用某个工具之前，先问一个研究代理该用哪个工具
Matt 想让他的代理访问 Gmail 和 Sheets。他的选择是谷歌官方的 CLI，或者是 Peter Steinberger 几个月前构建的一个。
他本来以为谷歌的会胜出。他问了 Last 30 Days，它告诉他使用 Peter 的，因为它是为高级用户构建的，也是为代理构建的。
他坦率地说明了这个答案的局限性。他从未亲自看过谷歌的官方 CLI，所以这是一个他照做的建议，而不是他运行的测试结果。
Peter 的 CLI 叫做 gws，我每天都用它来处理 Drive、Docs、Sheets 和 Gmail。
## 法则九：保留一个模型目前仍然失败的任务，并在每次发布时重新运行
Matt 让 Codex 构建有史以来数学上最适合终端的完美字体。
九十分钟后它回来了四个选项，三个是它自己生成的，一个是真正的字体，并让他找出区别。三个是垃圾。他选了那个好的，而那个好的就是真正的字体。
他把这三个都扔了。然后他决定保留这个失败案例，并说他会在每个新模型发布的那天要求它生成同一个字体。
一个今天的模型会失败的任务，是你自己拥有的衡量标准。发布文章告诉你实验室希望你相信什么。他估计在 2027 年的模型出现之前，那个字体是不可能存在的。
## 他认为你哪里搞错了
我问他现在什么被过度炒作。他说 AI 仍然被低估了，让他惊讶的是有多少人没有充分利用它。
他给任何感到落后的人的建议不是技术性的。去把它设置好，让它搞砸，然后不断问为什么，直到你豁然开朗。
这九条法则中没有一条要求你学习写代码。它们中的大多数都要求你停止做一些你目前认为代表了“谨慎”的事情。
Matt 在 X 和 GitHub 上的账号是 @mvanhorn。完整的三十分钟直播视频在这里。
如果你希望这能在你的企业里运行，而不仅仅是在你的笔记本上，我提供免费的 AI 审计。我们首先找到一个值得交给代理的流程，然后规划部署它所需的步骤。在这里预订。
## 相关链接
- [Tom](https://x.com/tomcrawshaw01)
- [@tomcrawshaw01](https://x.com/tomcrawshaw01)
- [14K](https://x.com/tomcrawshaw01/status/2082450077709381777/analytics)
- [Last 30 Days](https://github.com/mvanhorn/last30days-skill)
- [on a livestream](https://www.youtube.com/live/gzavyoPJ--A)
- [plan.md](http://plan.md/)
- [compound engineering](https://github.com/EveryInc/compound-engineering)
- [Jun 21](https://x.com/mvanhorn/status/2068426104088748331)
- [449K](https://x.com/mvanhorn/status/2068426104088748331/analytics)
- [an OpenAI plugin](https://github.com/openai/codex)
- [CLI Printing Press](https://github.com/mvanhorn/cli-printing-press)
- [@mvanhorn on X](https://x.com/mvanhorn)
- [on GitHub](https://github.com/mvanhorn)
- [here](https://www.youtube.com/live/gzavyoPJ--A)
- [Book one here](https://theaiarchitects.com/x/audit/matt-van-horn)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [8:56 PM · Jul 29, 2026](https://x.com/tomcrawshaw01/status/2082450077709381777)
- [14.9K Views](https://x.com/tomcrawshaw01/status/2082450077709381777/analytics)
- [View quotes](https://x.com/tomcrawshaw01/status/2082450077709381777/quotes)
---
*导出时间: 2026/7/30 10:47:51*