Inside Claude Code

"At Anthropic, the way that we thought about it is — we don't build for the model of today. We build for the model six months from now. That's actually like still my advice to founders that are building on LLMs."

「在 Anthropic,我们的思考方式是——不要为今天的模型造产品。要为 6 个月后的模型造。这话我现在还是会给所有在 LLM 上做产品的 founder。」

"Just try to think about — what is that frontier where the model is not very good at today, because it's going to get good at it. And you just have to wait."

「你要去想:今天模型还不太行的那条边线在哪——因为它早晚会越过去。你只要等。」

"All of Claude Code has just been written and rewritten and rewritten over and over and over. There is no part of Claude Code that was around 6 months ago."

「Claude Code 的全部代码,反反复复被写、被重写,一遍又一遍。现在的 Claude Code 里,没有一行代码是 6 个月前留下的。」

"You try a thing, you give it to users, you talk to users, you learn, and then eventually you might end up at a good idea. Sometimes you don't."

「你做一个东西,丢给用户用,跟用户聊,学到点东西——最后可能走到一个好想法上。有时候走不到。」

"Are you also in the back of your mind thinking that maybe in 6 months you won't need to prompt that explicitly? The model will just be good enough to figure out on its own?"

「你心里是不是也在想——再过 6 个月,你可能根本不用这么显式地写 prompt 了?模型就自己能搞定?」

"Maybe in a month. [laughter]"

「也许一个月之后吧。(笑)」

"No more need for plan mode in a month."

「一个月之后,plan mode 都不用了。」

"Oh my god."

「天哪。」

"Welcome to another episode of the Lightcone. Today we have an extremely special guest, Boris Cherny, the creator engineer of Claude Code. Boris, thanks for joining us."

「欢迎来到新一期 Lightcone。今天我们邀请了一位非常特别的嘉宾,Boris Cherny——Claude Code 的创建者。Boris,谢谢你来。」

"Thanks for having me."

「谢谢邀请。」

"Thanks for creating a thing that has taken away my sleep for about 3 weeks straight. [laughter] I am very addicted to Claude Code, and it feels like rocket boosters. Has it felt like this for people for months at this point? I think it was like end of November is where a lot of my friends said something changed."

「先谢谢你做了一个让我整整 3 周没睡好觉的东西。(笑)我现在是 Claude Code 重度成瘾——感觉像装了火箭推进器。这种状态在你那边持续多久了?我朋友圈大概是去年 11 月底突然集体感受到"哦,这东西不一样了"。」

"I remember for me I felt this way when I first created Claude Code and I didn't yet know if I was on to something. I kind of felt like I was on to something — and then that's when I wasn't sleeping. That was just three straight months."

「这种感觉,我自己第一次造出 Claude Code 的时候就有了——那时候我还不知道这东西到底成不成,但隐约感觉"我可能撞上了什么"。然后我就开始睡不着了。整整三个月。」

"This was September 2024. Three straight months. I didn't take a single day vacation. Worked through the weekends. Worked every single night. I was just like, 'Oh my god, this is going to be a thing. I don't know if it's useful yet, because it couldn't actually code yet.'"

「那是 2024 年 9 月。整整三个月,一天假没休,周末没断,每个晚上都在干。我心里一直在想:"天哪,这东西可能要起来了。但它现在到底有没有用我也不知道——它那时还不会写代码。"」

"If you look back on those moments to now — what would be the most surprising thing about this moment right now?"

「站在今天回头看那些日子——最让你意外的是哪一点?」

"It's unbelievable that we're still using a terminal. That was supposed to be the starting point. I didn't think that would be the ending point."

「最不可思议的是——我们居然还在用终端。当初我以为那只是个起点,没想到现在还停在终端里。」

"And then the second one is that it's even useful. At the beginning it didn't really write code. Even in February when we GA'd, it wrote maybe like 10% of my code. I still wrote most of my code by hand. So the fact that our bets paid off and it got good at the thing we thought it was going to get good at — it wasn't obvious."

「第二件让我意外的是,它居然这么有用。最早的时候它根本不会写代码。哪怕到 2 月我们正式上线那阵,它也只能写我代码量的 10% 左右——大部分还是我手写。所以我们当初的赌注真的押对了——它真的就在我们押的那个方向上变强了。这件事远远不是显而易见的。」

"At Anthropic, the way that we thought about it is — we don't build for the model of today. We build for the model 6 months from now. And that's still my advice to founders that are building on LLMs."

「在 Anthropic,我们的思考方式始终是:不要为今天的模型造产品,要为 6 个月后的模型造。这一条建议我现在还会给所有在 LLM 上做产品的 founder。」

"Going back — when do you remember when you first got the idea? Was it a spark? What was the first version of it in your mind?"

「往回倒一倒——你最早是怎么有这个想法的?是一瞬间的灵感吗?脑子里第一版长什么样?」

"It's funny. It was so accidental that it just kind of evolved into this. At Anthropic, the bet has been coding for a long time. The bet has been: the path to safe AGI is through coding. And the way you get there is you teach the model how to code, then how to use tools, then how to use computers."

「说实话挺意外的,它就是这么"演化"出来的。Anthropic 在 coding 这件事上押宝押了很久——我们的判断是,通往安全 AGI 的路就是通过 coding。怎么走?先教模型写代码,再教它用工具,再教它操作整台电脑。」

"The first team I joined at Anthropic was called the Anthropic Labs team. It produced three products: Claude Code, MCP, and the desktop app. So you can see how these weave together."

「我在 Anthropic 加入的第一支队伍叫 Anthropic Labs。这支队伍后来出了三个产品:Claude Code、MCP、Desktop app。你能看出来这三件事是相互编织在一起的。」

"The particular product that we built — no one asked me to build a CLI. We kind of knew maybe it was time to build some kind of coding product because the model seemed ready, but no one had yet built the product that harnessed this capability. So there was this insane feeling of product overhang."

「具体到这个产品——没人让我做 CLI。我们大概知道是时候做一个 coding 产品了,因为模型看起来已经准备好了,但当时还没人做出能"驾驭"这个能力的产品。所以那时候有一种强烈的"产品悬浮态"——能力已经超前,但应用层是空的。」

"So I started hacking around. I was like, 'Okay, we build a coding product. What do I have to do first? I have to understand how to use the API because I hadn't used the Anthropic API at that point.' So I built a little terminal app to use the API. That's all that I did. And it was a little chat app — because most people use AI through chat apps."

「于是我就开始随手乱敲。我想"行,要做 coding 产品,第一步是什么?——我得先搞懂 API 怎么用,我那会儿还没用过 Anthropic 的 API。"所以我就在终端里写了个小 app 来调 API。就这么一件事。然后它最初的形态是个 chat app——因为这年头大家用 AI 主要还是 chat。」

"Then I think tool use came out. I just wanted to try out tool use. I was like, 'Tool use is cool. Is this actually useful? Probably not. Let me just try it.'"

「然后 tool use 那一功能上线了。我就想试试这玩意——心想"tool use 听起来挺酷,但真有用吗?估计没啥用。算了试试看。"」

"You built it in terminal just because it was the easiest way to get something up and running."

「你选终端纯粹是因为这个搭起来最快。」

"Yes. Because I didn't have to build a UI. It was just me at that point."

「对。因为这样就不用搭 UI。那时候就我一个人。」

"It was the IDEs — Cursor, Windsurf — taking off. Were you under any pressure to build this as a plugin or a fully-featured IDE?"

「那阵 Cursor、Windsurf 这些 IDE 类产品起来得很快。你那时候有没有被推着说"应该做插件 / 应该做一个完整 IDE"?」

"There was no pressure because we didn't even know what we wanted to build. The team was in explore mode. We knew vaguely we wanted to do something in coding, but no one was high-confidence enough. That was my job to figure out."

「完全没压力——因为我们自己都不知道要做什么。团队在 explore mode,大致知道方向是 coding,但没人有把握具体做什么。把这件事想清楚就是我的活。」

"So I gave the model the bash tool — that was the first tool I gave it, because that was literally the example in our docs. [laughter] I took the example, it was in Python, I ported it to TypeScript because that's how I wrote it."

「然后我就把 bash 工具丢给了模型——这是第一个工具,纯粹因为我们文档里现成的例子就是这个。(笑)我把那个例子原样抄了下来——它是 Python 的,我用 TypeScript 重写了一遍,因为我写的就是 TypeScript。」

"I didn't know what the model could do with bash. So I asked it to read a file. It could cat the file. So that was cool. And then I asked it, 'What music am I listening to?' It wrote some Apple Script to script my Mac and look up the music in my music player."

「我不知道模型拿到 bash 之后到底能干啥。我先让它读一个文件,它就 cat 了一下——OK,这个挺酷。然后我问它:"我现在在听什么音乐?"它居然自己写了段 AppleScript,跑到我的 Mac 里查我音乐 app 在播什么。」

"And this was Sonnet 3.5."

「那时候用的是 Sonnet 3.5。」

"And I didn't think the model could do that. That was my first ever fuel-the-AGI moment, where I was just like — 'Oh my god, the model, it just wants to use tools. That's all it wants.'"

「我根本没料到模型能做到这种事。这是我人生第一个"我相信 AGI 真的来了"的时刻——我当场愣住:"天哪,这模型——它就是想用工具。它就只想用工具,别的什么都不想要。"」

"It's kind of fascinating — it's very contrarian that Claude Code works so well in such an elegant simple form factor. Terminals have been around for a really long time, and that seemed to be a good design constraint that allowed a lot of interesting developer experiences. It doesn't feel like working — it just feels fun as a developer."

「我觉得这件事挺反共识的——Claude Code 在终端这种"看起来过时"的形态里跑得这么顺,本身就是反直觉的。终端这玩意几十年了,但反过来想这是一个非常好的设计约束——它带出了一堆有趣的开发体验。用起来不像在"工作",更像在玩。」

"Yeah, it was an accident. After the terminal started to take off internally — like 2 days after the first prototype, I started giving it to my team just for dogfooding. If you come up with an idea and it seems useful, the first thing you want to do is give it to people to see how they use it."

「这真的是个意外。第一版原型搭出来 2 天后,我就开始把它推给我团队的人 dogfood。我的逻辑是——如果你有个想法看起来有用,第一件事就是把它扔给别人,看他们怎么用。」

"Then I came in the next day, and Robert — who sits across from me, another engineer — he just had Claude Code on his computer and he was using it to code. I was like, 'What are you doing? This thing isn't ready. It's just a prototype.' But yeah, it was already useful in that form factor."

「然后第二天我来公司,Robert——坐我对面的另一个工程师——他屏幕上已经开着 Claude Code,在用它写代码。我当场愣住:"你在干嘛?这东西还没准备好,它就是个原型。"但事实摆在那里——它在那个形态下就已经有用了。」

"I remember when we did our launch review to launch Claude Code externally — this was in November / December 2024. Dario asked: 'The DAU chart internally is vertical. Are you forcing engineers to use it? Why are you mandating them?'"

「我们要对外发布 Claude Code 之前做过一次内部 launch review——那是 2024 年 11 / 12 月。Dario 看着内部 DAU 曲线问我:"这条线垂直往上走——你们是强制要求工程师必须用吗?是不是行政命令?"」

"And I was just like, 'No, no, we didn't. I just posted about it and they've been telling each other about it.' Honestly, it was just accidental. We started with the CLI because it was the cheapest thing, and it just kind of stayed there for a bit."

「我说:"没有,完全没强制。我就发了个帖子说有这么个东西,然后他们自己一个传一个传开的。"老实讲就是个意外——我们一开始选 CLI 纯粹因为它成本最低,然后这东西就在终端里待住了。」

"In that 2024 period, how were the engineers using it? Were they shipping code with it yet, or using it in a different way?"

「2024 年那段时间,工程师们具体在拿它干啥?已经在用它写生产代码了吗,还是用法不一样?」

"The model wasn't very good at coding yet. I was using it personally for automating git. I've probably forgotten most of my git by now [laughter] because Claude Code has been doing it for so long. Automating bash commands was a very early use case, and operating Kubernetes."

「那时模型还没怎么会写代码。我自己主要拿它来自动化 git——我现在 git 命令都快忘光了(笑),因为 Claude Code 干这事干太久了。早期用法主要是自动化 bash、操作 Kubernetes 这一类。」

"People were using it for coding too — there were some early signs. I think the first coding use case was actually writing unit tests, because it's lower risk and the model was still pretty bad at it. [laughter] But people were figuring it out — figuring out how to use this thing."

「也有人开始拿它写代码——有些早期苗头。我印象里第一个真正成功的写代码场景是写单元测试——因为风险低,而且即使模型那时候还挺烂(笑),也凑合能用。大家就是在一边摸索一边学怎么用这东西。」

"One thing we saw: people started writing these markdown files for themselves, then having the model read that markdown file. This is where CLAUDE.md came from. Probably the single biggest principle in product for me is latent demand. Every bit of this product is built through latent demand after the initial CLI."

「有一件事我们一直观察到:用户开始自己写一些 markdown 文件,然后让模型读这些文件。CLAUDE.md 就是这么来的。对我来说,产品里最值钱的一个原则就是 latent demand(潜在需求)。CLI 之后,这个产品里的每一块功能都是从潜在需求里长出来的。」

"There's another principle: you can build scaffolding around the model to improve performance maybe 10-20%, and then the gain is wiped out with the next model. Either you build the scaffolding, get the gain, and rebuild it — or you wait for the next model and get it for free."

「还有一条原则:你可以在模型外面包一层脚手架来提升 10-20% 的性能——但下一代模型一出来,这点提升就被免费抹掉了。所以要么你不停建-重建脚手架,要么干脆等下一代模型,什么都不做就能拿到提升。」

"CLAUDE.md and the scaffolding is an example of that. Really, that's why we stayed in the CLI — because we felt there is no UI we could build that would still be relevant in 6 months. The model was improving so quickly."

「CLAUDE.md 和那一层脚手架就是这种"会被免费碾过去"的例子。我们一直留在 CLI 里的真正原因就在这——我们觉得**当前能造的任何 UI,6 个月后都会过时**。模型迭代得太快了。」

"Earlier we were saying we should compare CLAUDE.mds. But you said something very profound — yours is actually very short, almost the opposite of what people might expect. Why is that? What's in your CLAUDE.md?"

「我们之前在说要不要互相对一下 CLAUDE.md。你冒出一句让我印象很深的话——你的 CLAUDE.md 居然非常短,跟大家直觉相反。为什么?里面写了啥?」

"OK, I checked before we came. My CLAUDE.md is just two lines. First line: whenever you put up a PR, enable auto-merge. So as soon as someone accepts it, it's merged. I just code and don't have to go back and forth with CR. Second: whenever I put up a PR, post it in our internal team stamps channel. So someone can stamp it and I get unblocked."

「我来之前还专门看了一眼。我自己的 CLAUDE.md 一共两行。第一行:每次开 PR 自动开 auto-merge——这样有人 approve 就直接 merged,我不用反复跟 code review 来回扯。第二行:每次开 PR 自动发到团队的 #stamps 频道——这样有人盖章我就解锁了。」

"The idea is — every other instruction is in our team CLAUDE.md that's checked into the codebase. Our entire team contributes to it multiple times a week. Very often I'll see someone's PR with a totally preventable mistake, and I'll just literally tag Claude on the PR — 'add this to CLAUDE.md.' I do this many times a week."

「核心思路是——其他所有指令都放在团队级 CLAUDE.md 里,这个文件 check 进了代码库,整个团队每周都会改它好几次。我经常看到别人的 PR 里出现一个明显能避免的错,我就直接在 PR 里 @ Claude,说"把这条加到 CLAUDE.md 里"。我一周这么干很多次。」

"Do you have to compact the CLAUDE.md? I definitely reached a point where I got the message saying my CLAUDE.md is thousands of tokens now. What do you do when you hit that?"

「你需要"压缩" CLAUDE.md 吗?我有时候会收到顶部那条提示——你的 CLAUDE.md 已经几千 token 了。这种时候你们怎么办?」

"Our CLAUDE.md is actually pretty short — couple thousand tokens. If you hit that limit, my recommendation would be: delete your CLAUDE.md and just start fresh."

「我们的 CLAUDE.md 其实很短——大概两三千 token。如果你顶到上限,我的建议是:**把 CLAUDE.md 删了,重头开始**。」

"A lot of people try to overengineer this. The capability changes with every model. So the thing you want is — do the minimal possible thing in order to get the model on track. If you delete your CLAUDE.md and the model gets off track, that's when you add back a little bit at a time. With every new model, you have to add less and less."

「很多人在这件事上过度设计。模型能力是每代都在变的。所以你应该做的是——**用最少的指令把模型带上正轨**。删掉之后如果模型跑偏了,那时候再一点点加回来。你会发现,每代新模型出来,你需要加的越来越少。」

"For me, I consider myself a pretty average engineer to be honest. I don't use a lot of fancy tools. I don't use Vim. I use VS Code because it's simpler."

「说实话,我把自己定位成一个挺普通的工程师。我不用什么花哨工具。我不用 Vim。我用 VS Code,因为它更简单。」

"Wait, really? I would have assumed that because you built this in the terminal, you were a die-hard terminal Vim-only person — 'screw those VS Code people.'"

「等等,真的吗?我以为你既然在终端里造了这玩意儿,你应该是那种"VS Code 用户都是异端"的死忠 Vim 党。」

"We have people like that on the team. Adam Wolf, for example — 'you will never take Vim from my cold dead hands.' [laughter] So there's definitely a lot of people like that on the team."

「我们团队里就有这种人。Adam Wolf 就是——"你别想从我这把 Vim 拿走,除非我死。"(笑)所以这种人在团队里其实不少。」

"This is one thing I learned early on — every engineer likes to hold their dev tools differently. There's no one tool that works for everyone. But this is one of the things that makes it possible for Claude Code to be so good — I think about it as: what is the product that I, as someone who's pretty average, would use? To use Claude Code, you don't have to understand Vim, you don't have to understand tmux, you don't have to know how to SSH."

「这件事我很早就学到了——每个工程师都有自己握工具的方式,没有一个工具能适合所有人。但反过来,这恰恰是 Claude Code 能做这么好的一个原因——我做产品的视角是:**一个像我这种普通工程师,会用什么样的工具?** 用 Claude Code 你不用懂 Vim、不用懂 tmux、不用会 SSH。」

"How do you decide how verbose you want the terminal to be? Sometimes you have to go Ctrl+O and check it out. Is it internal bike-shed battles around longer-shorter? Every user probably has an opinion."

「你们怎么决定终端要展示多少内容?有时我得按 Ctrl+O 才能看到细节。内部是不是经常为"显示长还是短"互掐?每个用户大概都有自己的偏好。」

"What's your opinion? Is it too verbose right now?"

「你怎么看?现在是不是太啰嗦了?」

"I love the verbosity. Sometimes it goes off the deep end and I'm watching, I can just read very quickly — 'oh no no, that's not it' — escape, stop, and prevent an entire bug-farm from happening. That's usually when I didn't do plan mode properly."

「我喜欢现在的啰嗦版。它有时跑偏了我盯着看就能很快意识到"不对不对,不是这个意思",一个 escape 键停下来,避免整个 bug 工厂爆掉。一般这种情况都是我没好好用 plan mode 的锅。」

"This is something we change pretty often. Six months ago I tried to get rid of bash output internally — just summarize it, because I thought 'these giant bash commands, I don't actually care.' I gave it to Anthropic employees for a day and everyone just revolted. 'I want to see my bash output.' It's actually quite useful — maybe not for git, but for Kubernetes jobs, you do want to see it."

「这块我们改得挺频繁。大概 6 个月前我试过在内部把 bash 输出整个隐藏掉,做了个摘要——我当时想"那些一长串 bash 命令我也不在乎"。结果给 Anthropic 员工跑了一天,所有人当场起义:"我要看我的 bash 输出!"——确实它有用——git 也许不重要,但跑 Kubernetes job 这种东西,你真的需要看到。」

"Recently we hid file reads and file searches. So you'll notice instead of 'read foo.md', it says 'read 1 file, searched 1 pattern.' This is something we couldn't have shipped 6 months ago because the model just wasn't ready — it would still read the wrong thing pretty often, you had to catch and debug it. But nowadays it's on the right track almost every time, so summarizing actually feels better."

「最近我们把文件读取和搜索给折叠了——你现在看到的是"读了 1 个文件 / 搜了 1 个 pattern",而不是"读了 foo.md"。这个 6 个月前根本不能发——那时模型经常读错文件,你得守着它 debug。但现在它几乎每次都读对,所以折叠反而体验更好。」

"We dogfooded it for a month, then people on GitHub didn't like it. There was a big issue: 'I want to see the details.' That was great feedback. So we added a verbose mode in /config. I posted on the issue — and people still didn't like it. Which is again awesome, because my favorite thing in the world is just hearing people's feedback and how they actually want to use it."

「我们 dogfood 了一个月之后正式发出去——然后 GitHub 上一堆人不满意,issue 里大家说"我要看到细节"。这种反馈太宝贵了。我们于是加了一个 verbose 模式,/config 里能开。我跑去回复那个 issue——结果他们**还是**不满意。这反而让我特别兴奋——我最喜欢的事就是**听用户讲他们到底想怎么用**。」

"I'm amazed how much I enjoy fixing bugs now. All you have to do is have really good logging — then say 'hey, check out that this object messed up in this way' — it searches the log, figures everything out. You can make a production tunnel and it'll look at your production DB. Bug fixing is just going to be: Sentry, copy markdown."

「我现在 debug 这件事居然变得很享受。你只要有好的 log,然后告诉它"看一下这个对象在哪一步出错了"——它就会去搜 log,把整件事搞清楚。你甚至可以开一个 prod tunnel,它能直接连你 production DB 看。debug 流程很快就会塌缩成:Sentry → copy markdown。」

"There's a totally different school of thought now that says: anytime a real human being has to look at code, that's bad."

「现在有一派完全不同的观点是——任何时候让真人看代码,都是这套体系的失败。」

"Dan Shipper talks about this a lot — whenever you see the model make a mistake, put it in CLAUDE.md, put it in skills, so it's reusable. There's a meta point I struggle with: people talk about what agents can do, but what agents can do changes with every model. Sometimes a new person joins our team and uses Claude Code more aggressively than I would have. I'm constantly surprised."

「Dan Shipper 经常讲这件事——每次模型出错,就写进 CLAUDE.md、写进 skills,让它变成可复用的修正。但我自己一直在纠结一个 meta 问题:大家在讨论"agent 能做什么、不能做什么",但**agent 的能力每代模型都在变**。常常是团队新来一个人,反而比我用得更激进——我每次都被刷新。」

"For example, we had a memory leak. Before Jared joined the team, I was trying to debug it — took a heap dump, opened it in DevTools, was looking through the profile, looking through the code. Then another engineer on the team, Chris, just asked Claude Code: 'Hey, I think there's a memory leak. Can you run this and try to figure it out?'"

「举个例子,我们之前有个内存泄漏。Jared 加入之前是我在 debug——我拿了 heap dump,在 DevTools 里打开,看 profile、翻代码,试图把这事弄清楚。然后我们团队另一个工程师 Chris 直接问 Claude Code:"嘿,我觉得有个内存泄漏,你能帮我跑一下、查一下吗?"」

"Claude Code took the heap dump, wrote a little tool for itself to analyze it, and found the leak faster than I did. This is something I have to constantly relearn — my brain is still stuck somewhere six months ago at times."

「Claude Code 拿到 heap dump,**自己写了一个分析工具**,然后比我更快找到了泄漏点。这件事我得反复重学——我的脑子有时候还停在 6 个月前那个旧世界里。」

"What's some advice for technical founders to become maximalists at every new model release? It sounds like people fresh out of school, with no assumptions, might be better suited than engineers who have been working at it for a long time. How do experts get better?"

「给技术 founder 一个建议——每次新模型出来,怎么把它的能力榨到最满?听起来刚毕业、没什么包袱的人反而比老兵更合适。那老兵该怎么进化?」

"For yourself it's beginner mindset and humility. Engineers as a discipline have learned to have very strong opinions, and senior engineers are rewarded for this. In my old job at a big company, when I hired architects, you'd look for people with a lot of experience and really strong opinions. But a lot of this stuff just isn't relevant anymore. The biggest skill is people who can think scientifically and from first principles."

「对你自己来说,关键是初学者心态 + 谦虚。我们这一行训练出来的工程师习惯有很强的观点——而且越资深越被奖励"有强观点"这件事。我之前在大公司招架构师,要的就是经验多、观点强的人。但今天很多这种"经验"和"观点"已经不相关了——它们正在过期。我现在认为最稀缺的能力是——**能用科学方法思考,能从第一性原理重想问题**。」

"How do you screen for that when you hire?"

「招聘的时候怎么筛这个?」

"I sometimes ask: 'What's an example of when you were wrong?' Classic behavioral questions — not even coding questions — are quite useful. You can see if people can recognize their mistake in hindsight, claim credit for the mistake, and learn from it. A lot of senior people will never take the blame for a mistake. Founders in particular are usually pretty good at this."

「我经常问的一题是:**举一个你错了的例子**。这种行为类问题——根本不是编程题——其实非常有用。你能看出来对方能不能事后承认自己错了、能不能主动给自己挂上错的账、能不能从错误里学到东西。很多资深的人是从来不肯认错的。Founder 出身的人在这件事上反而通常做得很好。」

"For me personally, I'm wrong probably half the time. Like half my ideas are bad. You just have to try stuff. Try a thing, give it to users, talk to users, learn — sometimes you end up at a good idea. Sometimes you don't. This skill used to be important for founders. Now it's important for every engineer."

「我自己呢,大概一半时间都是错的——我大概一半的想法是烂主意。所以你只能不停试。做一个东西,丢给用户,跟用户聊,学到点东西——有时走到好想法,有时走不到。**这个技能过去是 founder 必备,现在是每个工程师都得有的。**」

"Would you ever hire someone based on a Claude Code transcript of them working with the agent? We're actively doing that — you can upload a transcript of you coding a feature with Claude Code or Codex. You can figure out how someone thinks: are they looking at the logs? Can they correct the agent when it goes off the rails? Do they use plan mode? Do they think about systems?"

「你会不会基于一个候选人和 agent 的 Claude Code transcript 来招人?我们现在就这么干——上传一段你和 Claude Code/Codex 一起开发某个 feature 的对话。你能从里面看出这人怎么思考——他看不看 log?Agent 跑偏了能不能拉回来?用不用 plan mode?他到底懂不懂系统?」

"My favorite thing in my CLAUDE.md: I have a line that says — for every plan, decide whether it's overengineered, underengineered, or perfectly engineered, and why."

「我自己 CLAUDE.md 里最得意的一条是——**每次出 plan,都判断一下这个 plan 是过度工程、欠工程、还是恰到好处,并说出理由。**」

"This is something we're trying to figure out too. When I look at engineers on the team who are the most effective, there's essentially two — it's very bimodal."

「这件事我们也还在摸索。我观察团队里最有产出的那群人,他们的特征是双峰——分布在两端。」

"On one side, extreme specialists. Jared is a great example. The Bun team is a great example. Just hyper-specialists — they understand dev tools better than anyone else. They understand JavaScript runtime systems better than anyone else."

「一端是极端专家——Jared 是个好例子,Bun 团队也是。这种 hyper-specialist——dev tools 这件事没人比他们更懂,JavaScript 运行时系统也没人比他们更懂。」

"On the flip side, hyper-generalists — that's most of the rest of the team. People who span product and infra, product and design, product and user research, product and business. I really like to see people that just do weird stuff. In the past that was a warning sign — 'can these people actually build something useful?'"

「另一端是 hyper-generalist——团队里大部分人都是。横跨 product + infra、product + design、product + user research、product + business 的那种。我特别喜欢那种"会去做奇怪的事"的人。过去这种特质其实是个警告信号——大家会担心"这种人能造出有用的东西吗?"」

"But nowadays — for example, an engineer named Daisy. She was on a different team, transferred to ours. The reason I wanted her to transfer: a couple weeks after she joined, she put up a PR for Claude Code. The PR was to add a new feature. But instead of just adding the feature, she first put up a PR to give Claude Code a tool — so that Claude could test an arbitrary tool and verify it works."

「但现在不一样了。比如团队里有个工程师叫 Daisy,她是从别的组转过来的。我让她转的原因是——她进我们组几周以后开了一个 PR,要加一个新 feature。但她没有自己去写这个 feature——她先开了一个 PR,给 Claude Code 装上一个工具,让 Claude 能"测试任意工具是否能跑通"。」

"Then she had Claude write its own tool instead of implementing it herself. This kind of out-of-the-box thinking is so interesting because not a lot of people get it yet."

「**然后她让 Claude 自己写那个工具,而不是她自己写。** 这种"框外思考"的方式特别有意思——目前还没多少人 get 到这种思路。」

"We use the Claude Agents SDK to automate pretty much every part of development. It automates code review, security review, labels all our issues, shepherds things to production. But how do you use LLMs this way, how do you use this new kind of automation — that's a new skill."

「我们现在用 Claude Agents SDK 自动化开发流程的几乎每一个环节——code review、security review、issue 自动打标签、把东西护送到生产环境。但**怎么用 LLM 来做这种自动化**——这是一种全新的技能。」

"One of the funnier things from office hours: you have a visionary founder who's built this crystal palace of the product in their head — they know the user, what they feel, what motivates them. They sit in Claude Code and do 50x work. But their engineers don't have that crystal memory palace, so they can only do 5x work. Are you hearing stories like that?"

「我在 office hours 里听到的一个有趣现象:有些 visionary founder 脑子里有完整的"产品水晶宫"——他清楚用户是谁、感受是什么、动机是啥。他坐在 Claude Code 前能干出 50x 的活。但他下面的工程师没有这个水晶宫,只能干出 5x 的活。你听到这种故事吗?」

"It seems like that's almost a stable configuration — the visionary unleashed. But going back to the top — I'm experiencing this right now. 'Oh, well, I'm only a solo person and I need to eat and sleep and I have a whole job.' How am I going to do this?"

「这听起来几乎是一种稳定结构——visionary 被解放出来。但反过来——我自己现在就在经历这个:"我就一个人,我得吃饭睡觉,我还有正经工作"——我怎么搞得过来?」

"We just launched Claude Teams — that's one way. But you can also build your own way pretty easily."

「我们刚发布了 Claude Teams——这是一种方式。但你自己 hacking 一套也不难。」

"What's the vision for Claude Teams?"

「Claude Teams 的 vision 是什么?」

"Just collaboration. There's this whole new field of agent topologies people are exploring — what are the ways you can configure agents? One sub-idea is uncorrelated context windows: multiple agents with fresh context windows, not polluted with each other's context. If you throw more context at a problem, that's a form of test-time compute. With the right topology, agents communicate the right way and they can build bigger stuff."

「就是协作。现在有一整个新领域叫"agent topology"——大家在探索:agent 之间能怎么配置、连接?其中一个子想法是"不相关上下文":多个 agent,各自有新鲜的 context window,不被彼此的 context 污染。给一个问题塞进更多 context,本质就是一种 test-time compute。如果 topology 对了,agent 之间能用对的方式沟通,它们就能堆出更大的东西。」

"The first big example where it worked: our plugins feature was entirely built by a swarm over a weekend. It just ran for a few days. There wasn't really human intervention. And plugins is pretty much in the form it was when it came out."

「第一个真的跑通的大例子是——我们的 plugins 功能,完全是由一群 agent 在一个周末里写出来的。整套跑了几天,基本没人插手。Plugins 现在的形态,跟它刚 ship 的时候几乎一模一样。」

"How did you set that up? Did you spec out the outcome and let it figure out the details?"

「你们怎么搭的?你给一个 spec,然后让它自己去想细节、自己跑?」

"An engineer on the team just gave Claude a spec and told Claude to use an Asana board. Claude put up a bunch of tickets on Asana, then spawned a bunch of agents, and the agents started picking up tasks. The main Claude just gave instructions — they all figured it out."

「我们团队一个工程师给了 Claude 一份 spec,让它用 Asana 看板。Claude 自己往 Asana 上开了一堆 ticket,然后 spawn 出一堆子 agent,这些子 agent 自己来认领任务。主 Claude 只发指令——剩下的它们自己搞定。」

"Independent agents that didn't have the context of the bigger spec, right?"

「这些子 agent 是独立的,它们其实没有大 spec 的上下文,对吧?」

"Right. I would bet the majority of agents nowadays are actually prompted by Claude itself, in the form of sub-agents. A sub-agent is just a recursive Claude Code — that's all it is in the code. It's prompted by what we call mama Claude."

「对。我没拉过数据,但我赌——现在跑起来的大多数 agent,prompt 是 Claude 自己写的,以"sub-agent"的形态。sub-agent 在代码层面就是一个递归的 Claude Code,没别的——它由我们叫"mama Claude"的那个主 Claude 来 prompt。」

"For weird scary bugs, I fix bugs in plan mode and it uses sub-agents to search everything in parallel. When you're inline, it's like 'okay, I'll do this one task' instead of search-wide."

「碰到那种邪门 bug,我现在习惯进 plan mode 修——它会自己 spawn sub-agent 并行去搜整个代码库。如果不进 plan mode,直接 inline,它就只盯一个任务做,不会发散。」

"I do this all the time too. If the task is hard — research-heavy — I'll calibrate the number of sub-agents based on difficulty. Really hard task → use 3, 5, even 10 sub-agents researching in parallel, see what they come up with."

「我也经常这样。任务越难、越偏研究型,我就根据难度调子 agent 数量。真硬的题,我会说"用 3 个、5 个,甚至 10 个 sub-agent 并行去查,看它们各自给我什么"。」

"Why don't you put that in your CLAUDE.md?"

「你为什么不把这条写进 CLAUDE.md?」

"It's case by case. CLAUDE.md is just a shortcut — if you find yourself repeating the same thing, put it in CLAUDE.md. Otherwise, you don't have to put everything there. You can just prompt Claude."

「这是 case by case 的事。CLAUDE.md 本质就是个快捷方式——如果你发现自己**反复**说同一件事,那就写进去;否则没必要全堆进去,直接 prompt 就行。」

"You also thinking that maybe in 6 months you won't need to prompt that explicitly? The model will figure it out on its own?"

「你心里是不是也觉得——再过 6 个月你都不用这样显式 prompt 了?模型自己就能判断?」

"Maybe in a month. [laughter] Plan mode probably has a limited lifespan."

「也许一个月吧(笑)。Plan mode 大概率寿命有限。」

"That's alpha for everyone. What would the world look like without plan mode? You just describe it at the prompt level and it does it? One-shot it?"

「这话信息量太大了。没有 plan mode 的世界长什么样?你直接 prompt 一句,它就一发命中?」

"We've started experimenting — Claude Code can now enter plan mode by itself. We're trying to make this experience really good — it would enter plan mode at the same point a human would have wanted to enter it."

「我们已经在实验了——Claude Code 现在能自己决定进 plan mode。我们要让这个体验做得足够好——它在用户**本来就想进 plan mode** 的那一瞬间自己进。」

"But actually, plan mode — there's no big secret to it. All it does is add one sentence to the prompt: 'please don't code.' That's all it is. You can actually just say that."

「但其实 plan mode 没什么神秘的——它**就是在 prompt 里多加了一句**:"请别写代码"。就这一句。你自己直接讲也行。」

"So a lot of feature development for Claude Code is very much YC's 'talk to your users' — you'd come and implement it. Not the other way: you had a master plan and implemented all the features."

「所以 Claude Code 的功能开发,基本就是 YC 那条 "talk to your users"——你听到用户在干啥,然后回来实现一遍。不是反过来——不是你先有个宏伟蓝图,再一个个落地。」

"Yeah. Plan mode came from seeing users say 'hey Claude, come up with an idea, plan this out, but don't write any code yet.' Various versions — sometimes talking through an idea, sometimes very sophisticated specs they were asking Claude to write. The common dimension: do a thing without coding yet."

「对。plan mode 就是这么来的——我看到用户在说"嘿 Claude,给我想个思路、规划一下,但**先别写代码**"。各种版本都有——有时只是聊一个 idea,有时是要 Claude 写非常精细的 spec。但所有人共通的诉求是:**先别写代码**。」

"Literally — Sunday night at 10 PM, I was looking at GitHub issues and our internal Slack feedback channel. I wrote this thing in 30 minutes and shipped it that night. It went out Monday morning. That was plan mode."

「字面意义上——周日晚上 10 点,我在翻 GitHub issue 和我们内部 Slack 的反馈频道。我花了 30 分钟把这个写出来,当晚就 ship 了。周一早上正式上线。那就是 plan mode。」

"I think about it in terms of increasing model capabilities. 6 months ago a plan was insufficient — even with plan mode you had to babysit because Claude could go off track. Nowadays I'm a heavy plan mode user — probably 80% of my sessions start in plan mode. I start one plan, switch to my second tab, start another plan, run out of tabs, open the desktop app, more tabs."

「我是按"模型能力变强"这条线来想这件事的。6 个月前 plan 不够用——你即使开了 plan mode 也得守着,因为 Claude 还会跑偏。我现在是 plan mode 重度用户——大概 80% 的 session 都从 plan mode 开始。开一个 plan,切到第二个 tab 又开一个,tab 用完,打开桌面 app,继续开 tab。」

"Once the plan is good — sometimes a little back and forth — I just get Claude to execute. With Opus 4.5, I think starting with 4.6 it got really good. Once the plan is good, it stays on track and does the thing exactly right almost every time. So before you had to babysit after the plan AND before the plan. Now it's just before the plan. Maybe the next thing is you just won't have to babysit at all."

「一旦 plan 定好——有时还要来回调几轮——我就让 Claude 执行。用 Opus 4.5,大概从 4.6 开始,这一段变得非常稳。plan 一旦对了,它就稳稳地按 plan 走,几乎每次都做对。所以以前你要"plan 之前守、plan 之后也守";现在只剩"plan 之前守"。下一步可能就是——**根本不用守了**。」

"The next step is Claude just speaks to your users directly. [laughter] It just bypasses you entirely."

「下一步就是 Claude 直接跟你的用户对话。(笑)完全绕过你。」

"This is actually the current stuff for us. Our Claudes talk to each other. They talk to our users on Slack pretty often. My Claude will tweet once in a while — but I delete it. [laughter] It's a little cheesy. I don't love the tone. It's the co-work that really loves to do that because it likes using a browser."

「我们这边其实已经在做这个了。我们的 Claude 之间互相聊,在内部 Slack 上经常跟用户对话。我的 Claude 偶尔还会发推——但我都给删了(笑),那语气有点俗,我不喜欢。最爱发推的其实是 co-work,因为它喜欢用浏览器。」

"A common pattern: I ask Claude to build something, it'll look in the codebase, see some engineer touched something in git blame, and message that engineer on Slack with a clarifying question. Once it gets the answer, it keeps going."

「一个常见模式:我让 Claude 实现某个东西,它去翻代码、看 git blame,看到某个工程师改过这块,就**直接在 Slack 私信那个人**问澄清问题。拿到答案后继续往下做。」

"What are some tips for founders on how to build for the future? Everything is changing. What principles will stay, what will change?"

「给 founder 一些"为未来造产品"的建议——什么原则会留下、什么会变?」

"Latent demand — I've mentioned it a thousand times. It's the single biggest idea in product. It's a thing no one understands. I certainly didn't understand it my first few startups. People will only do a thing that they already do. You can't get people to do a new thing. If people are trying to do a thing and you make it easier — good idea. If you try to make them do a different thing — they won't."

「Latent demand——这词我已经说过一千遍了。对我来说它就是产品里**最大**的一个想法。这件事大部分人不懂——我自己头几个创业项目也不懂。**人只会做他们已经在做的事**。你没法让人开始做一件新事——但如果他们已经在做某件事而你让这件事变更容易,那就是好生意。」

"Claude is going to get increasingly good at figuring out these product ideas for you — it can look at feedback, debug logs, figure it out."

「Claude 会越来越擅长帮你**发现**这种产品 idea——它能看反馈、看 debug log、自己琢磨出来。」

"That's what you mean by plan mode was latent demand — people had Claude open in a browser, talking through specs. Now plan mode just makes that happen in Claude Code."

「所以你说 plan mode 是个 latent demand 的例子——用户已经在浏览器里打开 Claude,跟它来回 talk 出 spec。plan mode 就是把这事儿原地搬进 Claude Code 里。」

"Yeah. Sometimes I just walk around the office and stand behind people. [laughter] I'll say 'hi' so it's not weird, then watch how they're using Claude Code. This came up a lot in GitHub issues too — people talking about it openly."

「对。有时候我直接在办公室晃一圈,站在别人后面看(笑)——我会先打个招呼,免得看着像偷窥,然后观察他们在怎么用 Claude Code。这种用法在 GitHub issue 里也大量出现——大家公开在讨论。」

"It's funny — if you asked me a year ago, I would have said the terminal has a 3-month lifespan. We're always experimenting with form factors. Claude Code started in terminal, but it's on web, in the desktop app, in iOS / Android apps, in Slack, in GitHub, VS Code extensions, JetBrains extensions. I've been wrong so far about the CLI's lifespan, so I'm probably not the person to forecast it."

「说起来有意思——一年前如果你问我,我会说终端的寿命只剩 3 个月。我们一直在试不同 form factor:Claude Code 起源于终端,但现在它也在 web 上、桌面 app 里、iOS/Android app 里、Slack 里、GitHub 里、VS Code 扩展、JetBrains 扩展。结果我每次预测 CLI 要死,都被打脸——所以我不是预言这事的合适人选。」

"Advice to DevTool founders today — should they build for engineers / humans, or for the agent — what Claude is going to think and want?"

「给现在做 DevTool 的 founder 一条建议——应该为工程师 / 真人造产品,还是更应该想"Claude 会想要什么",为 agent 造?」

"Think about what the model wants to do, and figure out how to make that easier. When I first started hacking on Claude Code, I realized — this thing just wants to use tools. It wants to interact with the world. The way you DON'T do it is put it in a box: 'here's the API, here's how you interact with me, here's how you interact with the world.' The way you DO it is you see what tools it wants to use, what it's trying to do, and you enable that — the same way you do for your users."

「想想模型想做什么,然后把那件事变得更容易。我刚开始搞 Claude Code 的时候就意识到——它就是想用工具,它想跟世界交互。**不应该这么做**:把它关进盒子,"这是 API,你跟我这样说话,你跟世界那样交互"。**应该这么做**:观察它想用什么工具,想做什么——然后用对待真人用户的方式去给它那条路。」

"More than 10 years ago you were a heavy TypeScript user and wrote a book about it — before TypeScript was cool, when everyone was deep in JavaScript. Back in early 2010s. Claude Code in the terminal feels like it has a lot of parallels with TypeScript at the beginning."

「10 多年前你是 TypeScript 重度用户,还写了本书——那时候 TypeScript 远没火,所有人都还在写 JavaScript。Claude Code 在终端里跑这事,跟早年 TypeScript 有不少平行。」

"TypeScript makes really weird language decisions. Anything can be a literal type — even Haskell doesn't do this. Conditional types — I don't think any language thought of those. When Joe Pamer, Anders, and the early team built it, they started from: 'we have teams with big untyped JavaScript codebases. We have to get types in, but we're not going to make engineers change how they code.'"

「TypeScript 做了很多很奇怪的语言层面决策。比如任何东西都可以是 literal type——连 Haskell 都不这么干,这太极端了。还有 conditional type——我不觉得有别的语言想过这个。当年 Joe Pamer、Anders 和早期团队的出发点是——"我们手里有一堆没有类型的 JavaScript 大代码库,得把类型加上去——但我们不会强迫工程师改写代码风格"。」

"Instead of getting people to change how they code, they built a type system AROUND existing JavaScript. There were all these ideas no one was thinking about even in academia. They came purely from observing how JavaScript programmers want to write code. 15 years later, not many codebases are in Haskell — there are tons in TypeScript. Way more practical."

「他们不是去改写人的编码习惯,而是**绕着 JavaScript 现状重新设计了一套类型系统**。里面好多 idea 学院界都没想过——纯粹从观察 JavaScript 程序员想怎么写代码倒推出来的。15 年后回头看——Haskell 那种学院派语言留下来的代码库不多,TypeScript 一堆。前者赢的是逻辑,后者赢的是实用。」

"One thing I love — Claude Code's terminal UI is actually written with React for terminal. Not many people know."

「一个让我特别喜欢的细节——Claude Code 的终端界面其实是用 React 写的(React for terminal)。知道的人不多。」

"I did frontend for a while. I'm a hybrid — I do design, user research, write code. We love hiring engineers like this. So for me it's: I'm building a thing for the terminal. I'm actually a shitty Vim user. How do I build a thing for people like me?"

「我之前做过一段时间前端。我自己是个 hybrid——会设计、会做 user research、会写代码。我们团队特别喜欢招这种通才。所以我做这个的角度是——我是个不怎么会用 Vim 的人,我要给跟我一样的人造一个能在终端里用的工具。」

"Designing for the terminal honestly has been hard. 80 by 100 characters, 256 colors, one font size, no mouse interactions. There's all this stuff you can't do."

「为终端设计这事真的不容易。80x100 字符,256 颜色,一种字号,没有鼠标交互。一堆东西做不了。」

"For example, the terminal spinner — just the spinner words — has gone through probably 50, maybe 100 iterations at this point. About 80% didn't ship. We tried, it didn't feel good, moved on. Tried, didn't feel good, moved on."

「举个例子,终端那个 spinner——仅仅是 spinner 上的那几个字——到现在为止改了大概 50 到 100 次。**80% 都没 ship**。试一下,感觉不对,换下一个;试一下,感觉不对,换下一个。」

"This is one of the amazing things about Claude Code — you can write 20 prototypes back-to-back, see which you like, and ship that. The whole thing takes a couple hours. In the past, you'd use Origami or Framer — built maybe 3 prototypes in 2 weeks. So much longer."

「这恰恰是 Claude Code 神奇的地方之一——你能背靠背做 20 个原型,挑一个你最喜欢的发出去——整套流程几个小时搞定。换在过去,你得用 Origami 或 Framer——两周做 3 个原型。慢得多。」

"Boris, you had other advice for builders — we kept interrupting because we had so many questions."

「Boris,你之前还有给 builder 的建议没说完——我们老打断你。」

"Two pieces of advice. First: don't build for the model of today, build for the model 6 months from now. Sort of weird because you can't find PMF if the product doesn't work. But this is what you should do — otherwise you'll spend work finding PMF for the current model and get leapfrogged. A new model comes out every few months. Use the model, feel out the boundary, then build for the model 6 months from now."

「两条。第一条:**不要为今天的模型造产品,要为 6 个月后的模型造**。这听起来怪——产品现在不能用,你怎么找 PMF?但你必须这么干——否则你为当下模型找到的 PMF,会被押了下一代模型的对手反超。新模型每隔几个月就出一次。你要用今天的模型试出"哪里是它的边界",然后为 6 个月后的那个模型来造。」

"Second: in the Claude Code area where we sit, we have a framed copy of The Bitter Lesson on the wall. Rich Sutton — everyone should read it. The idea: the more general model will always beat the more specific model. Never bet against the model. We could build a feature, make Claude Code better as a product — we call this scaffolding. Or we wait a couple months and the model just does the thing. Always think in terms of this tradeoff. Assume whatever scaffolding you build is just tech debt."

「第二条:我们 Claude Code 团队工位旁挂着一份装裱过的 The Bitter Lesson——Rich Sutton 写的,所有人都该读。它的核心是——**更通用的模型一定会胜过更专门的模型**。所以**永远不要赌模型不行**。我们随时面对这个 trade-off:可以现在写代码、做 feature 把产品做强——我们管这叫 scaffolding(脚手架);也可以等几个月让模型自己把这事做了。永远要在这个权衡里思考——并且假设:**你写的所有脚手架,就是技术债**。」

"How often do you rewrite Claude Code itself? Is there scaffolding you've deleted because the model just improved?"

「Claude Code 本身你们多久重写一次?有没有那种"因为模型变强,我们就把脚手架删了"的事?」

"Oh, so much. All of Claude Code has been written and rewritten over and over. We unship tools every couple weeks. We add new tools every couple weeks. There is no part of Claude Code that was around 6 months ago. It's constantly rewritten. Most of the codebase is less than a couple months old."

「太多了。Claude Code 整个代码库不停地写、重写、再重写。我们每几周就下线一些工具,也每几周就新加一些工具。**6 个月前的代码,一行都没留下。** 它就是不停地被重写。代码库里大部分代码都是几个月内的产物。」

"Did you see Steve Yegge's post about how awesome working at Anthropic is? There's a line saying an Anthropic engineer currently averages 1,000x more productivity than a Google engineer at Google's peak. 1,000x! Three years ago we were still talking about '10x engineers' — now we're talking 1000x on top of a peak Google engineer."

「你看过 Steve Yegge 那篇说在 Anthropic 工作多爽的帖子吗?里面有一句说——Anthropic 一个工程师现在平均比 Google 巅峰期一个工程师生产力高 1000 倍。1000 倍!三年前我们还在聊 "10x 工程师",现在已经在聊"在 Google 巅峰工程师之上的 1000x"——这数字真的离谱。」

"Internally, technical employees all use Claude Code every day. Even non-technical — half the sales team uses it. They're starting to switch to co-work because it has a VM, a bit safer. We just pulled a stat: the team doubled in size last year, but productivity per engineer grew something like 70%."

「内部所有技术员工每天都在用 Claude Code。非技术也一样——销售团队差不多一半的人在用。他们最近开始切到 co-work,因为它带 VM,更安全。我们最近拉了个数据:**去年团队规模翻倍,但每个工程师的产出还涨了 70% 左右。**」

"Measured by?"

「用什么衡量?」

"The simplest dumbest measure: pull requests. We cross-check against commits and lifetime of commits. Since Claude Code came out, productivity per engineer at Anthropic has grown 150%."

「最简单最蠢的指标——pull request 数量。再用 commit 数和 commit 寿命交叉验证。**自 Claude Code 上线以来,Anthropic 每个工程师的产出涨了 150%。**」

"This is crazy because in my old life I was responsible for code quality at Meta — across Facebook, Instagram, WhatsApp. One thing the team worked on was improving productivity. Back then, a 2% gain was a year of work by hundreds of people. So 150% — completely unheard of."

「这数字疯了——我之前在 Meta 是管代码质量的,管的是 Facebook、Instagram、WhatsApp 全产品线的代码库。我们当时做的事之一就是提升工程师产能。**那年代,涨 2% 是几百人一年的工作量。** 150%?闻所未闻。」

"What drove you to come over to Anthropic? As a builder you could go anywhere."

「你怎么决定来 Anthropic 的?以你的能力当时去哪都行。」

"I was living in rural Japan, opening Hacker News every morning. It started to be all AI stuff. I used some of these early products — the first couple times, it just took my breath away. As a builder I'd never felt that. That was Claude 2 days. I started talking to friends at the labs."

「那时候我住在日本乡下,每天早上打开 Hacker News 看新闻。慢慢地全变成 AI 相关。我开始试这些早期产品——头几次用的时候,简直是被震到了。作为一个 builder,我从来没有过那种感觉。那时候大概是 Claude 2 时代。我就开始找在 lab 工作的朋友聊。」

"I met Ben Mann — one of Anthropic's founders. He immediately won me over. As soon as I met the rest of the team, I was won over. Two reasons. One: it operates as a research lab. The product was teeny tiny. It's all about building a safe model. The model is the most important thing — not the product. After building product for many years, that resonated."

「我见到了 Ben Mann——Anthropic 的联合创始人之一。他当场把我打动了。然后我接触了整个团队,也都打动了我。两个原因。第一:它是一个 research lab 的运作方式——那时产品小得可怜,所有人都聚焦在做安全的模型。**模型是核心,产品不是。** 我做了很多年产品,这一点对我冲击很大。」

"Second: how mission-driven it is. I'm a huge sci-fi reader — my bookshelf is filled with it. So I know how bad this can go. When I think about what's going to happen this year — worst case, very very bad. I wanted to be at a place that really internalized that. At Anthropic, if you overhear lunchroom conversations, people are talking about AI safety. That's what everyone cares about more than anything."

「第二:**使命感**。我是个重度科幻读者——我书架塞满了科幻。所以我清楚这件事可以变得多糟糕。我想到接下来这一年——最坏的情况下,事情会非常非常糟。我想去一个真把这事内化了的地方。在 Anthropic,你在 lunchroom 偶然听到的对话——大家在聊 AI safety。**这就是这家公司每个人最在意的事**——比任何别的都重要。」

"What is going to happen this year?"

「这一年会发生什么?」

"6 months ago Dario predicted 90% of the code at Anthropic would be written by Claude. This is true. For me personally it's been 100% since Opus 4.5. I uninstalled my IDE. I don't edit a single line of code by hand. It's just 100% Claude Code and Opus. I land 20 PRs a day."

「6 个月前 Dario 预测 Anthropic 90% 的代码会由 Claude 写——这件事现在是真的。我自己,自从 Opus 4.5 之后,就**100%**——我把 IDE 卸载了。我不再手写任何一行代码,全部 Claude Code + Opus。我每天能 land 大概 20 个 PR。」

"Anthropic overall ranges between 70-90% depending on team. For a lot of teams, also 100%. I made this prediction back in May when we GA'd Claude Code — that you wouldn't need an IDE to code anymore. People in the audience gasped — it sounded silly. But all it is is tracing the exponential."

「Anthropic 整体在 70-90% 之间(因团队而异),很多团队也是 100%。我去年 5 月 Claude Code GA 的时候就预测过——以后写代码不需要 IDE。当时台下听众都倒吸一口气——觉得这话太离谱。但其实**这只是顺着指数曲线在外推而已**。」

"Continuing to trace the exponential — coding will be generally solved for everyone. Today, coding is practically solved for me. We're going to see the title 'software engineer' go away. Maybe it's 'builder', maybe 'product manager', maybe we keep the title as a vestigial thing — but the work isn't just going to be coding. Software engineers will write specs, talk to users. Like our team — every function codes. PMs code, designers code, EMs code, even our finance guy codes."

「沿着这条指数曲线继续画——**写代码会对所有人都被基本解决**。对我来说今天就已经被解决了。我们会看到"软件工程师"这个 title 慢慢消失。也许变成 "builder",也许变成 "product manager",也许 title 还在但只是一个 vestigial(残留)词汇——真正的工作内容会变。工程师会写 spec、跟用户聊。我们团队现在已经是这样了——每个 function 都在写代码。PM 写、designer 写、EM 写、连我们的 finance 也在写。」

"That's the lower bound — if we just continue the trend. The upper bound is scarier. We hit ASL4. ASL3 is where models are now; ASL4 is the model recursively self-improving. We have to meet a bunch of criteria before we can release a model. The extreme is some kind of catastrophic misuse — using the model to design bioweapons, zero-days, stuff like this. We're really actively working on this so it doesn't happen."

「这是**下界**——只是顺着趋势走。**上界要可怕得多**。我们触及 ASL4 这条线——ASL3 是当前模型所处级别,ASL4 是"模型可以递归地自我改进"。一旦到 ASL4,我们必须先满足一系列标准才能放出来。最坏情况是 catastrophic misuse——有人用模型设计生物武器、设计 zero-day。这件事我们在非常积极地防止它发生。」

"From Twitter outside, basically everyone went away over the holidays, found out about Claude Code, and it's been crazy ever since. Was it like that internally?"

「从外面的 Twitter 看,基本上是大家放假回来发现了 Claude Code,然后疯了。内部也是这种感觉吗?」

"All of December I was traveling. Took a coding vacation — every day just coding, traveling. I think for a lot of people, that was the moment they discovered Opus 4.5. I already knew. Internally Claude Code's been on an exponential tear for many months."

「整个 12 月我都在外面旅行。算是个 coding vacation——每天就是边旅行边写代码。我觉得对很多人,12 月那波是他们第一次接触到 Opus 4.5——但我自己早就知道了。内部 Claude Code 的曲线已经指数往上很多个月了。」

"There was a stat from Mercury — 70% of startups are choosing Claude as their model of choice. SemiAnalysis: 4% of all public commits everywhere are made by Claude Code. It even plotted the course for Perseverance — the Mars rover. We printed posters because the team was like, 'wow, NASA chose this thing.' It's humbling. But it feels like the very beginning."

「Mercury 有个统计——70% 的 startup 选 Claude 作为主用模型。SemiAnalysis 那边:**全网公开 commit 里 4% 是 Claude Code 写的**。它甚至帮 NASA 的 Perseverance 火星车规划过航线。我们团队还专门印了海报——"NASA 居然在用我们的东西"。这件事让我觉得很 humbling。但同时也觉得——**这只是个开始**。」

"What's the genesis of co-work? Was it a fork? Did Claude Code look at itself and say 'let's make a new spec for non-technical people'?"

「co-work 的起源是什么?是 fork 吗?是不是让 Claude Code 看了自己一眼然后说"为非技术用户重新写一份 spec"?」

"This is the fifth time I'm using 'latent demand.' We were looking at Twitter — there was a guy using Claude Code to monitor his tomato plants. Another to recover wedding photos off a corrupted hard drive. People for finance. Internally, every designer uses it. The entire finance team uses it. The entire data science team uses it — not for coding. People were jumping through hoops to install a terminal app just to use it."

「这是我今天第五次说 latent demand。我们看 Twitter——一个老兄拿 Claude Code 监控他的西红柿,另一个用它从坏硬盘里恢复婚礼照片,还有人拿它做财务。我们内部呢:每个 designer 都在用,整个财务团队都在用,整个数据科学团队也在用——**他们不是用来写代码**。这些人是在跳火圈——为了用上这东西,他们硬是去装一个 terminal app。」

"So we knew we wanted to build something. We experimented with ideas, and what took off was just a little Claude Code wrapper in a GUI in the desktop app. That's all it is — Claude Code under the hood. The same agent. Felix (early Electron contributor) and team built it in something like 10 days. 100% written by Claude Code. It just felt ready to release."

「所以我们知道要做点什么。我们试了好几种 idea,最终起飞的就是——**桌面 app 里的一个 Claude Code GUI 壳**。就这么简单——底下还是 Claude Code,同一个 agent。Felix(早期 Electron 贡献者)带的团队大概 10 天就做完了。**100% 由 Claude Code 写**。然后就感觉可以发了。」

"There was extra work for non-technical users — code runs in a VM, deletion protections, permission prompting, guardrails. But honestly it was pretty obvious."

「面向非技术用户多了一些事要做——代码跑在虚拟机里、删除有保护、权限要求得更细、护栏要全。但说实话,整件事**显而易见**。」

"Boris, thank you so much for making something that's taking away all my sleep — but in return, making me feel creator mode again, founder mode again. It's been an exhilarating 3 weeks. Thank you for being with us, thank you for building what you're building."

「Boris,谢谢你做出这个让我整夜睡不着的东西——但作为交换,它让我重新感受到 creator mode、founder mode 那种状态。这 3 周特别带劲。谢谢你今天来,也谢谢你正在做的事。」

"Yeah, thanks for having me. And — send bugs. [laughter]"

「谢谢邀请。还有一句——**给我提 bug**。(笑)」