本集简介
双语字幕
仅展示文本字幕，不包含中文音频；想边听边看，请使用 Bayt 播客 App。
今天我们请来了Anthropic的朋友Thariq，一起在演播室里聊聊。
We're here in the studio with our friend, Thariq, from Anthropic.
我觉得，总的来说，Claude Code，我，我……
And I guess, generally, the Claude Code, I I...
现在有很多东西的界限都在模糊，而你自从加入Anthropic以来，一直对一切都了如指掌。
There's there's so much, sort of merging of boundaries, and you've been so on top of everything since you joined Anthropic.
你早早地就接触了Claude Code，而且还有……
You have been early to Claude Code itself, but then also...
你在其他播客里也讲过那个故事。
And you've told that story in other podcasts.
你还一直在谈论像智能体这样的东西。
And you've also been talking about seeing like an agent.
最近，你做了那个顶级的AIE、世界博览会演讲、Fable的实地指南，显然你们推出了Fable，所以这是不是……
Most recently, you did the top AIE, World Fair Talk, Field Guide to Fable, which obviously you guys launched Fable, so does it...
那有点作弊了。
That's cheating.
而且最近还推出了Claude Tag，我们还会聊到《Pacing the Frontier》。
And mostly, most recently also launching Claude Tag, and we're also gonna be talking about Pacing the Frontier.
Anthropic最近事情真的很多。
There's a lot going on in Anthropic.
我想最核心的问题是，在这么多事情同时发生的时候，在Anthropic工作是什么感觉？
I guess top of the question is, what's it like being Anthropic when there's so much going on?
我觉得这就像……
I think that it is like...
有时候真的会感觉有点应接不暇。
I think you can get whiplash sometimes.
我觉得，怎么说呢……
I I think, like, going...
我加入Anthropic的时候，就是因为Claude Code。
When I joined Anthropic, I joined because of Claude Code.
那时候Claude Code刚出来，我就觉得，这东西太棒了。
Like like, Claude Code had just come out, and I was like, this is so good.
而对我来说，Opus 4 就像是……
And Opus four to me was like like, just...
我简直无法想象它有多好。
I could not imagine, like, how good it was.
你懂我的意思吗？
You know what I mean?
那对我来说真的是一个重要的时刻。
And that was like a real moment for me.
但我当时基本上是在试图说服我的创业朋友们使用 agent decoding。
But I was, like, trying to convince, like, my startup friends, basically, to to use agent decoding.
他们就说，哦，不。
And they're like, oh, no.
他们说，我们的工程师觉得它还不够好什么的。
Like, our engineers don't think it's good enough or something.
我当时就觉得，这太离谱了。
And I was like, that's insane.
现在你往前看，你知道，十二个月不到。
And now you, like, fast forward, you know, twelve months, less.
然后，就像，是的，大家默认的编程方式就是这样。
And, like, it's just, like, yeah, the default way that everyone codes.
我觉得，从推销它，到现在，你知道，教人们怎么...
And I think that, like, just having to go from, like, selling it to, like, now, you know, teaching people how to be...
你知道，充分利用它，提高效率之类的。
You know, make the most use of it and and be more efficient and things like that.
这就像是一个巨大的，你知道，巨大的变化。
It's just like a like a big, you know, like, big change.
是的，我觉得，作为人类，要跟上所有事情真的很难。
And, yeah, I think, like, it's just hard to stay on top of everything as a human.
你知道，我觉得事情发生得太快了。
You know, like, I think things happen so so fast.
这两个代理又在搞这个。
These two more agents at it.
我的意思是，是的，实际上那种智能体相关的东西比人工处理要扩展得好得多，人工处理就是，哦，现在同时有三件事在发生，而且都是紧急情况，你怎么应对呢？
I mean, yeah, like, that's actually like the agentic stuff scales much better than the, like, human stuff where it's like, oh, like, there are three things happening right now and, like, they're all emergencies and, like, how do you, like, you know, respond to it?
你是怎么分配你的时间的？
What do you split your time on?
你做了很多技术写作和工程方面的工作。
You do a lot of technical writing, engineering work.
所以我觉得，当我加入Claude Code团队时，我基本上是想教人们如何使用Claude Code。
So I I think that, like, when I joined the Claude Code team, I wanted to teach people how to use Claude Code, basically.
我觉得这已经变成了某种……
And I think that, like, that has been something that, like...
我原本以为我可能会花一点时间在这上面，或者，你知道，做……
I thought, like, maybe I would spend a little bit of time on it or, like, I'd, you know, like, do...
我一开始在Agent SDK上花了一些时间，而且我不太确定，你知道，关于工具框架，那个惨痛的教训会怎么发展。
I was spending some time on the agent SDK first, and I I wasn't exactly sure, like, you know, how the bitter lesson would go, you know, when it comes to, like, harnesses.
就像，我觉得有时候我们会想，哦，Claude Code之后会是什么呢？
Like, I I think sometimes we're like, oh, like, what's after Cloud Code?
你懂我的意思吧？
You know what I mean?
所以一开始，我就想教大家怎么用 Cloud Code，让它用起来更简单。
And so initially, I was like, I just wanna teach people how to use Cloud Code and make it easier to use Cloud Code.
我觉得这就像是……
And I think that is just like...
随着工具越来越好用，现在最大的问题就是怎么用好这些智能体。
As the harnesses have gone better and better, that's like the dominant problem now is like how do you use the agents.
这真的是一件很考验技巧的事情。
Like, it's, like, such a high skill expression thing.
所以我做这些，然后也做工程方面的工作。
So I I do that, and then I do engineering work.
我也会做演讲。
I give talks.
但我觉得，我做工程工作的时候，目标就是把用户的反馈带回来，然后还能跟大家分享怎么用 Cloud Code 来做工程。
But I think, like, when I'm doing engineering work, my goal is to take that feedback that we get from users and also, like, then be able to talk about, like, hey, how how to use Cloud Code to do engineering.
所以这里形成了一个很好的循环。
So there's kind of like a a good loop there.
我会……
I'll...
听众朋友们，我们会附上你和Sarah在DevWriters聚会上的那次演讲，我们之前稍微提到过，就是先做工作，然后再谈论它，对吧。
For listeners, we'll attach the talk that you did with Sarah for the DevWriters meetup, which we talked a little bit about, well, first you do the work and then you talk about the Right.
差不多就是这样。
Something like that.
种瓜得瓜，种豆得豆，或者……
Sow and reap or...
收获和……
Reap and...
种瓜得瓜，种豆得豆。
Sow and reap.
差不多就是这样。
Something like that.
那么……
So...
然后稍微预告一下，我们会聊聊这个工具的发展历程。
And then just to preview a little bit, we are gonna talk about the evolution of the harness.
它已经从最初只是一个命令行工具，发展了很多。
It has come a long way from just being a CLI.
我们会聊聊Claude Mods，这个今天开始泄露了，因为你实在保守不住秘密。
We're gonna talk about Claude Mods, which is starting to leak today, because you couldn't keep it secret.
基本上就是这样。
Basically.
这里面内容可多了。
There's there's a lot, there.
我觉得你一开始是加了那个“询问用户问题”的工具，这个工具大家其实又爱又恨。
I I think you started off with, like, adding ask user question tool, which people love and hate actually.
但我个人觉得这个想法挺有创意的。
But I actually thought it was, like, very innovative.
然后现在我自己也有一个版本了。
And then now I have, like, my own version.
你还有你的那个采访我的版本。
You have your your your interview me version.
对，现在每个人都有自己的东西。
And, yeah, everyone just has, their own their own stuff.
而且，这已经不重要了，因为现在你得写那些能生成其他提示词的提示词，还有循环什么的。
And, like, it no longer matters because now you're supposed to write prompts that create other prompts and loops and all these things.
当然。
Sure.
那现在最先进的做法是什么？
So what's the state of the art today?
比如，大家都在……
Like, what what are people...
你现在会建议大家怎么做？
What what what are you, like, telling people to do today?
问用户问题，这是模型第一次在引导提问方面表现得很好。
Ask user question was the first time that the model was good at elicitation.
你知道，我觉得这有点像一种涌现出来的行为，我其实很想看看模型能不能做到这一点。
You know, I think this was, like, kind of an emergent behavior that I, like, you know, wanted to see if the models could do.
我有一些人机交互的背景，所以我在本科和研究生阶段都做过这方面的研究。
I have, like, kind of, like, a human computer interaction background, so I, like, know, did that in in undergrad and grad school.
所以这就像是……
And so this was like...
我觉得这对我来说有点像人机交互，就是试图弄清楚，代理如何与你沟通，并提取出你的需求。
I think it's kind of like human agent interaction to me, like, know, trying to figure out, like, how can the agent communicate with you and extract, you know, the the requirements.
我觉得关于……的一件事是，
I think that, like, one of the things about...
随着Claude Code变得越来越广泛，困难的是每个人都有自己使用它的方式。
Like, that's difficult as Claude Code has gone broader and broader is that everyone has, like, their own way of using it.
而且实际上很难改变默认行为。
And it's actually very hard to, like, change the default behavior.
比如说，如果有人让Claude Code做某件事，有时候他们只是希望它直接干活，因为他们可能很擅长写提示词。
So for example, like, if someone asks you Claude Code to do something, sometimes they just want them to do the work because they're, like, maybe a very good prompter.
有时候他们希望……
And sometimes they want...
其实他们并不擅长写提示词。
Like, actually are not good at prompting.
你明白我的意思吗？
You know what I mean?
然后你需要……
And you need...
就是，智能体需要去澄清一下，对吧？
Like, the agent needs to, like, clarify, you know?
所以这就像是一个很好的分界线。
And so that's like a good split lane.
而那个提问工具，差不多就是沿着这个方向来划分的，就是说，你觉得你自己有没有能力直接指挥这个智能体？
And to ask you the question tool sort of like splits along that side where like, or do you feel like you're good enough to instruct the agent as it is?
还是说智能体能够像……
Or is the agent able to like...
智能体是不是需要提取更多需求，跟你多协作，真正了解你的偏好？
Does the agent need to like pull out more requirements and like collaborate with you more and really understand your preferences?
总的来说，我认为几乎所有人都更偏向后者，而不是前者，也就是说，他们的问题更模糊，而且他们知道的比他们想要的少……
I, on the whole, believe that pretty much everyone is more on the latter than the former that they, like, have more ambiguity and they know less than they want...
比他们自认为对问题的了解要少。
Than they, like, think they know about the problem.
但，实际上……
But, like, real...
这其实是个界面设计问题，要让这个过程变得容易。
It's like interface design problem to make that easy.
你明白我的意思吗？
You know what I mean?
所以，如果你在设计一个问题，比如……
And so, like, if you're designing a problem, like...
或者你在处理一个问题的时候，像什么schema啊、调用栈啊这些细节其实都很重要。
Or if you're going through a problem, like, you know, things like what's the schema or, like, what's the call stack and things like that are really important.
你知道，细节和设计都很关键。
You know, like, the details and the design are important.
理想情况下，你希望在开始写代码之前就先把一些难题想清楚。
Ideally, you want to figure out some of these, like, hard problems ahead of time before starting implementation.
对，这就是为什么它们叫未知数。
And, yeah, that's why they call, like, unknowns.
所以我觉得在智能体编程里，搞清楚你的未知数永远是一项核心技能。
And so I think that this will forever be, like, a skill in agentic coding is like figuring out your unknown.
所以呢……
So, like...
因为就算模型特别聪明，它也得知道你想要什么，对吧，而且你是有自己的偏好的。
Because even if the model is, like, super intelligent, it, like, needs to know what you want, you know, and, like, you have preferences.
你需要把那些……
Like, you need to sort of, like, pull the...
把它提取出来。
Pull that out.
所以这就是，我觉得我现在的方向……
And so that's, like, I think how I'm...
我在推动的方向。
What I'm pushing.
接下来的问题是，代理如何与你互动？
The question then is, like, how does the agent interact with you?
我认为HTML一直是实现这一点的主要方式。
And I think that has been HTML, has been, like, the big way of doing that.
我们最近也加入了工件功能。
And we've recently added artifacts.
关于工件，我觉得我们其实做得不太好，比如……
And artifacts, I actually think we've done a bad job of like...
或者说，我在解释如何充分利用它们方面做得不好。
Or like I've done a bad job of like explaining how to use them fully.
我们有很多属性功能。
We have a lot of property capabilities.
它们有相关联的数据库，你知道吗？
They have a database associated with them, you know?
所以每个工件都可以存储和写入持久数据。
And so every artifact can store and write persistent data.
它们可以反馈给Claude，你知道吗？
They can, like, feed back into Claude, you know?
所以，有一件事人们还没做，而我想鼓励的是，就是仪表盘工件的想法。
And so, like, one thing that, you know, like, people are not doing yet that I'm trying to, like, encourage is, like, this idea of a dashboard artifact.
比如，让Claude长期负责一个项目。
So you have, like, Claude working on a project long term.
也许是个看板之类的。
Maybe it's like a Kanban or something.
它可以把看板数据存储在自己的数据库里。
It can store that Kanban data in its database.
多个Claude可以通过artifact MCP访问这些数据。
Multiple Claude's can access that data via, like, the artifact MCP.
而且，那个...
And, like, that m...
那个artifact也可以和那些Claude交流。
That artifact can, like, talk to those clouds as well.
所以，那个...
And so, like, the...
我们基本上是在构建基础组件，让你能通过artifacts拥有这种生成式界面，从而让你能更充分地展示来自代理的丰富细节。
We're basically building the primitives for you to be able to have this, like, generative interface via artifacts that will, like, let you surface more of that rich detail from the agents.
我觉得现在几乎所有关于代理的问题都是这样的：你以为你知道自己想要什么，但实际上你并不真正清楚。
And I think that, like, almost everything with agents right now is, like, this problem of, like, you think you know what you want, but you don't really know what you want.
而且代理需要很多细节，与它们进行循环协作非常重要。
And, like, the agents need a lot of detail, and collaborating with them in the loop is really important.
所以artifacts就是我们试图在这方面发展的方式。
And so artifacts are, like, the, like, way that we're trying to evolve there.
但还有很多工作要做，因为这比选择题复杂得多。
But there's a lot of work to do because it's so much more complicated than, like, a multiple choice question.
你懂吧？
You know?
这里面涉及的细节多得多，比如图表、代码片段、模式，或者任何跟那个问题相关的东西。
There's a lot more, like, detail in terms of, like, diagrams and code snippets and schemas or, like, like, whatever it is for that problem.
但基本上，Artifacts 更像是那种更 AGI 导向的提问方式。
But, like, artifacts is, like, the more AGI pilled way of, like, you know, doing ask you the question, basically.
所以……
So...
关于这些，我有一点不太明白……
I think one thing that's unclear to me about these...
就是 Artifacts 这块，什么样的反馈应该通过 Artifact 来传递，什么样的反馈应该通过 Claude 聊天来传递？
The the artifact stuff is, like, what feedback should go in through the artifact and what feedback should go through a Claude chat?
因为更 AGI 导向的做法是把所有东西都直接喂给 Claude。
Because the more AGI pilled one is to just feed everything to the Claude.
我觉得更偏向AGI的做法是通过artifact来交互。
I think the more AGI pulled one is to go through the artifact.
比如……
Like...
而且我觉得，从长远来看，artifact会成为你和整个系统之间的接口。
And I I think that, like, we sort of imagine in the limit, I think that artifacts will be your interface into the harness.
你懂吧？
You know?
你可以在这个实时的计划或工作文档上直接评论。
You can, like, comment on this, like, live, like, document of your plan, of the work.
你可能会看到多个agent，不同的agent在同时处理。
You can see maybe, like, multiple agents and different agents are doing this.
而这个artifact是为当前的工作量身定制的。
And that artifact is built for the current work that you're doing.
所以每个artifact都会有点不一样。
And so, like, each one has, like, slightly different.
我觉得从基础设施的角度来看，我们还在逐步实现这个目标。
I think we're still, like, getting there from, like, an infrastructure perspective.
不过，是的，我觉得为你的工作环境提供实时界面，可能就是未来的方向。
But, yeah, I think, like, on the fly interface for your harness is probably where things are are headed.
有没有一种版本，它是从命令行或聊天中抽象出来的，然后你……
Is there a version of it that's an abstraction from CLI or chat and you...
因为现在很多情况是，你在和Claude Code交互。
Because right now, a lot of it is, okay, you're interfacing with Claude Code.
你会收到HTML，用来做原型。
You're having HTML given back for a mock up.
内容相当丰富。
It's pretty rich.
还有图表。
There's diagrams.
工件就是把这些连接起来的方式。
Artifacts are ways to connect these together.
为什么不干脆都用这种方式呢？
Why not just do everything that way?
那就变成了要区分，推理是在哪里发生的？
Then it becomes, like, separating out, like, where is the inference happening?
智能是在哪里实现的？
Where is intelligence happening?
工作是在哪里完成的？
Where is the work happening?
你明白吗？
You know?
我觉得这有点像，就是那种区别，比如……
Like, I think this is kind of, like, difference between, like...
或者说本地和云端之间的一些差异。
Or like some of the distinction between local and cloud.
所以我觉得现在，如果你用Claude Code，它就像是本地的。
And so, I think right now, if you use Claude Code, it's like local.
而且，比如说，你可以启动远程控制来获得一些云端行为，或者你也可以在云端启动Claude Code。
And, like, you can spin off remote control, for example, to get some cloud behavior, or you can spin off Claude Code in the cloud.
我们正在朝着一个方向前进，不再是Claude的...
We're moving towards a place where instead of Claude's...
比如，你给本地的Claude发消息，它在本地启动一个会话并执行，而是更像你有一个在云端运行的Claude，你给它发消息。
Like, you message a local Claude, starts a session locally and it executes, to more like you have a Claude that you message that's in the cloud that's running.
它可以运行本地会话，也可以运行云端会话。
It can run like local, or like cloud sessions.
这有点像Claude Tag的工作方式。
This is kind of how Claude Tag works.
但随着时间的推移，我们也会加入本地操作的能力。
But, like, over time, we'll add, like, local hands as well.
而且，本地操作意味着，如果那个代理在线，它就能访问你的电脑，并且在那里工作。
And and so, like, local hands will be the ability for that agent to access your computer if it's online, you know, and be able to, like, work there.
所以它可以启动很多不同的子代理。
And so it it can spin off many different sub agents.
它可以，比如，传递这些……
It can, like, commute those...
子代理之间可以互相沟通。
Sub agents can communicate with each other.
这就是工件发挥作用的地方，基本上用来展示所有的工作成果。
And that's where the artifact comes in to display all of that work, basically.
所以你可以想象，比如……
So you can imagine, like, the...
你在把这些东西分开。
You're separating out these things.
所以有，比如，表面UI显示，那是一个工件，托管在某个地方，还有数据库等等。
So there's, like, the surface UI display that's an artifact and hosted somewhere and has a database and everything.
还有推理智能，对吧，发生在云端，你不用担心关电脑什么的。
There is the inference intelligence, right, that's happening on the cloud, and you don't have to worry about shutting off your computer or whatever.
然后还有，比如，手那种……
And then there's the, like, hands kind of, like...
而且它可以在本地运行。
And it can be local.
也可以在远程沙箱或者任何你需要完成工作的地方运行。
It can be in, like, a remote sandbox or wherever you need your work to be done.
这其实就是现在Cloud Code体验的拆解。
That's, like, unpackaging, like, the Cloud Code experience right now.
我们现在就像……
We're, like...
现在所有事情都集中在一个地方发生。
Right right now, it all happens in one place.
那你觉得多人协作这方面会怎么发展？
So How do you see, like, the multiplayer side of that?
比如说团队想用这种方式工作。
So say teams wanna work in this way.
现在这种方式很个人化，但你觉得未来多人协作会是什么样？
Right now, it's very individual, but how do you see the future of multiplayer?
比如现在，有个叫Claude Tag的版本。
Like, right now, guess there's Claude Tag, which is a version.
但是
But
我们正在推出Projects。
We're launching projects.
Projects就是那个……
And so projects is the...
就是类似Claude Tag的那种抽象概念，但用在我们的Claude产品上。
Like, this abstraction that's kind of like Claude Tag, but on our Claude products.
你可以给它发消息，它就能像Claude Tag那样，比如派生出子代理之类的。
So you can message it and, like, it will do the Claude Tag like stuff, like spinning off sub agents.
所以我们觉得，在多人协作方面……
So we think with multiplayer...
Claude Tag可能更天然地支持多人协作，因为它就在Slack里，权限什么的都已经设置好了。
Like, Claude Tag is, like, a little bit more native multiplayer because it's just, like, in your Slack and the permissions are all figured out and stuff like that.
但我确实认为多人协作是故事的重要部分，而且这需要更好地整合起来。
But I do think multiplayer is, like, an important part of the story and, like, that that will need to get tied together more.
可以想象这会变得多复杂，比如你有手，但现在别人电脑里也有手。
Like, can imagine how complicated it gets when you're like, oh, you have hands, but now you have other hands in other people's computers too.
而且你需要给它们授权，或者你有你的MCP，别人有别人的MCP，怎么弄清楚怎么用它们。
And, like, you need to, like, permission them or, like, you have, like, your MCP and someone else's MCP and how do you figure out how to use them.
这会变得相当复杂。
It gets, like, quite complicated.
而Claude Tag在解决这些问题上做得很好。
And Claude Tag does a good job of like sanding down all of these issues.
所以当你使用...
So that like when you have...
Google Docs，它怎么访问Google Docs？
Google Docs, how does it access Google Docs?
它可以通过共享的Claude MCP访问，或者如果它没有权限，也可以通过你的本地凭据访问。
Like, accesses through the shared Claude MCP or it can access through your local credentials as well if it doesn't have access.
不过，是的，我觉得Claude Tag就是我们的多人协作产品，对于那些天生就需要多人参与的事情来说，真的很有用。
But, yeah, I I think Claude Tag is our multiplayer product, and it's really useful for these, like, things that are inherently multiplayer.
比如，像值班的时候，事故处理本质上就是多人协作的。
Like, okay, like, on call, for example, incidents are inherently multiplayer.
你想@一下Claude。
You want to tag Claude.
你想让多个人登录。
You want multiply build log in.
你希望它能找到上下文。
You want it to be able to find context.
我觉得每当我做某件事，想要隐私或安全，或者想让别人来审查的时候。
I think whenever I'm, like, working on something and I want, like, privacy or security or, like, I want other people to review it.
你知道，这样真的很不错，比如……
You know, it's really nice to, like...
我会为每个项目建一个频道，然后比如@法务，说，嘿，我想发布这个。
I'll have a channel per project and I'll, like, at legal, for example, be like, hey, like, I want to ship this.
你能，比如说，这里是……
Can you, like, like, here's the...
Claude 什么都知道。
Claude knows everything.
你懂的，直接跟它聊就行。
You know, just chat with it.
这样法务就能得到准确的答案，比如具体什么内容会进入代码，我就不用一直盯着了。
And that way, legal gets precise answers, you know, on, like, what exactly is shipping into the code, and I don't need to be in the loop.
所以我觉得，多人协作越来越普遍了，是的，每个人都能和 Claude 一起参与。
So I think, like, multiplayer is getting, like, more and more, like, yeah, everyone can participate with Claude.
我觉得 Claude Tag 就是那个产品，项目一开始可能是单人操作，但之后会慢慢扩展。
I think Claude Tag is, like, that that product and, like, projects will start off single player and will, like, you know, expand.
我觉得有个问题，可能涉及两个问题，一个是身份，一个是隔离的单位。
I think there's a question about, like, maybe dual questions about identity and the unit of isolation.
Claude Tag，你特意把它做成了独立的身份。是的。
Claude Tag, you specifically chose to make it its own identity Yes.
这算是个有争议的选择。
Which is like a controversial choice.
其实还有别的做法。
There's there's other ways to do it.
Claude Projects，听起来可能跟ChatGPT的Projects差不多，就是……
Claude Projects, probably it sounds like, you know, if it's anything like ChatGPT projects, it is...
你知道，隔离的关键在于那些工件、那个云端实例，大家都在上面协作。
You know, the the isolation is that artifacts, that cloud instance, everyone's collaborating on this.
它……
It...
它会……
It'll...
听起来，它应该是……
It sounds like, you know, it should it should be...
比如，如果你在某个东西上非法协作，那个频道就应该是一个项目。
Like, if you're if you're collaborating illegal on on on a thing, like, that channel should be a project.
现在还不是，但它……
Like, it's not yet, but it...
这是自然的下一步。
That's the the natural next step.
我觉得，在Claude Tag里，实际上……
I mean, like, I think in Claude Tag, it's effectively...
就像Claude Tag，你基本上得自己安排。
Like like, Claude Tag, you have to sort of do your own arrangement, basically.
所以Claude Tag，对，每个频道你可以随便命名，我就这么命名。
And so Claude Tag, yeah, each channel is, like, you can name it as you want, and I name Yeah.
基本上每个功能
Like, each feature, basically
作为一个频道。
As a channel.
我觉得，有一些跨……
I I think, like, there is some trans...
就是说，什么时候算转移不太清楚，因为比如说……
Like, it's unclear when there is transference because let's say it is...
如果你有一个同事
If you have a coworker
被标记在这些东西里，那确实算转移，因为那是同一个人。
Who is tagged in all these things, yes, there is transfer because it's the same person.
但用Claude的话，就不太清楚是不是一定算，嗯，不，你根本不知道……
But with Claude, it's unclear if it's, like, necessarily, like, well, no, you don't know any...
你不知道其他那些东西。
You don't know about the other stuff.
你只应该用这些东西。
You should only use this stuff.
这就像那个冰山一角的梗，对吧，你可以……
It's like the tip of the iceberg meme, right, where you can, like...
这基本上就是我们花了那么多时间搞的东西。
This is what we spent so much time on, basically.
就像，有无限的表面积，比如，好吧，你想让Claude...
Is, like, there is, like, infinite surface area of, like, okay, you want Claude's to...
不是无限的，但就像，那是，表面空气...
Not infinite, but, like, it was, surface air...
有很多，比如，表面积需要去弄清楚，比如，权限和可见性，还有，你知道，就像，你怎么能...
A lot of, like, surface area to figure out of, like, permissions and visibility and, like, you know, like, how can you...
怎么...
How...
就像，Claude尽可能好地、尽可能安全地运作，你知道吗？
Like, Claude operate as well as you can, as safely as you can, you know?
显然，这对我们非常重要，因为，你知道，就像，我们代码库的安全非常非常重要。
And obviously, this is very important to us because, like, you know, like, like, security for our code base is very, very important.
所以我们在这方面投入了很多时间。
And so we've put a lot of time into this.
有太多，比如，边缘情况你可以发现，就像，哦，对，这个频道里的Claude有不同的权限，但它可以给另一个频道发消息，那它不能通过这种方式窃取数据吗？
There's so many, like, edge cases you can figure out where it's like, oh, like, yeah, this Claude in this channel has different permissions, but it can message another channel and can't it exfiltrate data that way?
或者，比如，你能不能……
Or, like, can you, like...
如果它用了你的MCP，然后去给别人发消息呢？
What if it uses your MCP and then messages someone else?
像这样的问题太多了，我们真的花了很多功夫去解决。
Like, there's, like, so much, and we've, like, really put a lot of work into sending it down.
确实下了很多功夫。
Lots of work.
好的。
Okay.
Fable。
Fable.
Fable，你写了两篇好文章。
Fable, you wrote two good articles.
我是说，你写了很多好文章，但比如那篇《Fable 实地指南》和《构建 Claude Code》那篇。
I mean, you've written many good articles, but on, you know, field guide to Fable, building Claude Code.
我很好奇，从你看到的来看，Anthropic 外部那些顶级用户有没有什么共同的模式？
I'm curious from what you've seen, is there any common patterns that you see in, like, top users at Anthropic externally?
比如，怎么用 Claude Code 才能发挥最大效果，有什么最佳实践吗？
Like, what are best practices for getting the most out of Claude Code?
我觉得最核心的技能就是提示词，提示词非常重要。
The, like, meta skill I say is, like, prompting is, like, very important.
你懂吧？
You know?
而且，那个……
And, like, that...
我觉得这可不是随便说说的，因为很多人觉得提示词不重要。
Like, I I think this is, like, not trivial to say because I I think a lot of people are like, oh, prompting doesn't matter.
他们觉得随便说一句话，Claude 就能搞定。
It's just like, I can just say a sentence and Claude will do it.
但我觉得提示词真的就是那种……
And I think prompting is really this, like, this...
这就像公开演讲，你知道的，或者写作之类的。
It's like public speaking, you know, like, or writing or something.
而且是针对特定的听众，那个听众就是Claude。
And for a specific audience, and that audience is Claude.
你需要建立对Claude的心智模型，了解它是怎么思考、怎么运作的。
And you need to, like, build a mental model of Claude and how it thinks and how it works.
所以，使用Claude Code最重要的技能就是拥有这个心智模型，对吧，了解Claude擅长什么、能一次性完成什么、不能做什么。
And so that's, like, the most important skill in working with Claude Code is, like, having this mental model, right, of Claude and, like, what it can do well, what it can one shot, what it can't.
所以很多人看到提示词时，他们只是……
And so so many people when you see prompting, they're just like...
他们的提示词很短，但他们有非常好的对Claude和代码库的心智模型，所以一切都很轻松。
They're short prompts, but they have such a good mental model of Claude and of, like, the code base and things like that that, like, it's effortless.
你懂我的意思吗？
You know what I mean?
但这就像技能上限很高。
But it's like high skill ceiling.
所以，像那种花很多时间在提示上，建立对代理工作方式的直觉和心智模型，真的非常重要。
So, like, that work of, like, you know, spending a lot of time prompting and building mental models of how, you know, an intuition for how the agents work is really important.
然后我觉得，接下来就是之前我们提到的那些未知的东西，就是能够发现你的……
And then I think, like, the next thing is, like, the unknown stuff we talked about earlier where it's like being able to find out, like, your...
那些你不知道的，或者你还没写下来的东西，去学习不同的事物。
What you don't know or what you haven't written down, learning about, like, different things.
我觉得随着Claude能做的事情越来越多，你遇到那些超出你熟悉范围、而且你知识储备很少的事情的可能性会非常高。
I think as Claude can do more and more things, the likelihood of you doing something out of distribution for you and, like, you have low domain knowledge on is very, very high.
你明白吗？
You know?
而且你越能学会那些词汇来更好地给Claude下达指令，这就变得特别重要。
And the more you can, like, learn the vocabulary to be able to prompt Claude, it becomes really important.
所以，我觉得最重要的未知就是那些“未知的未知”，就是那种，我根本不知道还有这种东西存在的情况。
And so, like, think the most important unknowns are the unknown unknowns where you're like, I I just, like, don't even know that this exists.
没错。
Exactly.
我觉得这就像地图和实际地形的区别，对吧？你的提示词就是地图，而实际地形就是代理需要完成的工作。
I think that's, like, illustration of, like, the map and the territory, right, where you're like, okay, this is my prompt, and the territory is, like, the actual, like, work that the agent needs to do.
如果你非常精确，你就能给出更精确的指令。
And if you are, like, very precise, you can give more precise things.
比如说，在设计方面，我就不太精确。
So, like, for example, in design, I'm not very precise.
我不是设计师。
I'm not a designer.
所以我会说，给我八个不同的设计稿吧。
So I say, like, you know, give me, like, eight different mock ups.
但如果我是设计师，我可能会说，嘿，这里有一些参考网站。
But if I was a designer, maybe I'd be like, oh, hey, here's some reference sites.
我想要这种字体，这种风格，还有这几个组件，放到Figma的MC板上展示出来。
Like, I want this type of font and this type of, like, look to it and here's, like, a few different components to to, like, visualize here the Figma MC board to bring in, you know, like...
所以，有了这些语言，你就能精确得多。
And so you can just be so much more precise with that language.
如果你不是设计师，基本上你就得去学那个语言，或者去了解那些你不知道自己不知道的东西。
And if you're not a designer, you just need to, like, try and learn the language basically or learn the unknown unknowns.
我觉得这适用于所有事情。
And that's true of, like, everything, I think.
你越能通过和Claude合作来学习事物是怎么运作的，你的提示词就会越好。
Like, the more, like, you can work with Claude to learn, like, how things work, the better your prompting will be.
我觉得另一个好例子是游戏设计。
I think another good example of this is, like, game design.
你知道，很多人会说，哦，我现在可以随便写个游戏了。
You know, like, where a lot of people are like, oh, like, I can vibe code a game now.
然后他们发现，这游戏不好玩。
And they're like, it's not fun.
你懂我意思吧？
You know what mean?
而且，游戏设计的关键在于，每一个选择背后都有很多……
And and, like, it's just like the thing about game design is, like, every one of these choices has, like, a lot
很多变体。
of Variations.
其实没什么技巧可言。
Not a lot of like craft to them.
所以就是，哦，好吧。
So it's like, oh, okay.
比如，你在做一个飞行游戏时，飞机的操控感……
Like, when you're making a a flying game, the feel of the paint...
飞机以及它对你操作的反应方式，有很多讲究。你知道，一个游戏设计师可能会花好几天在这上面。
Plane and the, like, you know, way it responds to your controls has a lot of, like like, know, a game designer would spend, like, days on that.
你明白我的意思吗？
Do you know what I mean?
对我来说，
To me,
那就是品味。
that's what taste is.
对，嗯。
Like Mhmm.
确实，从一千个数学上说得通的答案里，挑出来的那个。
It is, like, from the possible space of 1,000 mathematically valid answers Yeah.
就是人类会喜欢的那一个。
Is the one that is the humans will like.
是的。
Yes.
我觉得“品味”这个词让我有点纠结，因为你说得对，但每个人对它的定义都不一样。
I think with taste, I'm, like, torn on this word because I think you're right, but everyone has different definitions.
而且它听起来有点像技术含量低，或者有点精英主义，好像只有某些人才有品味。
And it it sounds sounds kind of like low skill or, like, elitist almost where you're like, oh, like, are certain people with taste.
就像……
It's like
品味就是我自己说的那个品味。
taste is what I call taste.
没错。
Exactly.
这些人没有品味。
These guys don't have taste.
对。
Exactly.
哦，就像工程师没有品味，而我，作为创始人，有品味。
Oh, like, an engineer doesn't have taste like I, the, like, founder, have taste.
你懂我的意思吗？
Do you know what I mean?
我觉得这其实不是真的。
And I think that's actually not true.
我觉得工程师对这类特定问题很有品味，你知道，每个人都有对特定问题的品味。
Like, I I think, like, the engineers have a lot of taste for these particular, like, problems, you know, and I think everyone has taste for particular problems.
我觉得，像Jason Liu，就像...
I think, like, Jason Liu, like like...
展开剩余字幕（还有 480 条）
比如说，要有品味，你得先吃。
Say, like, in order to, have taste, you have to eat.
你懂吗？
You know?
我特别喜欢那种感觉，就是，你得做很多事情。
And I really like that where it's like, okay, you have to like do a lot of things.
你得不断尝试，搞清楚你想要什么，喜欢什么，然后建立起自己的领域词汇。
You have to like iterate and figure out what you want, what you like and, you know, like build that like domain vocabulary.
然后当你写提示词的时候，你就是在把这些都综合起来，玩得很开心。
And then when you're prompting, you're like synthesizing all of that for fun.
当别人说得比你好时，是不是很烦？
Isn't it annoying when someone else says it better than you?
就是，靠，
It's just like, fuck,
我得永远引用这家伙的话了。
I have to quote this guy forever.
以后都得一直引用Jason Liu的话了。
Having to quote Jason Liu forever.
他会喜欢这个的。
He's gonna love this.
所以我得到了认可。
So I I get props.
而且
And
有时候其实还不止这样。
sometimes it's actually not even that.
有时候就是凭直觉。
Sometimes it's just intuitive.
就像，你都没意识到自己想要什么，直到Amodei推出某个东西，你才觉得，哦，这个用起来立刻就更顺手了。
Like, don't realize you even want something till Amodei puts it out and you're like, oh, this just feels immediately better.
没错。
Exactly.
有一件事我一直在纠结，就是我觉得我有一半时间提问的方式，比如说我用语音。
One thing I go back and forth on is I feel like the way I prompt half the time, let's say I use voice.
有的人用语音，其他人也用语音。
Guys have voice, other people have voice.
这正好相反。
That is the opposite.
就是我自己在那儿唠叨个两分钟，按住功能键，然后松开，然后希望它能明白我的意思。
There's just like me rambling for, like, two minutes, pressing down the function key and then let go and then, like, hopefully, it figures it out.
而且很多时候，它确实能明白。
And oftentimes, it does.
但这不像那种结构化的提示词那么深思熟虑，比如那种像PRD或备忘录一样写得清清楚楚的沟通。
But it's not as thoughtful as, like, a structured prompt with, like, you know, well run communication as though it's a PRD or memo.
这是不是和大家平时的做法一致？
Is that in line with how people do?
这基本上就是两种模式的提示方式，有些提示词你花很多时间提前准备，有些你就随手一写。
This is, like, basically bimodal prompting where there's some prompts where you spend a lot of time upfront and other prompts you just dash it off.
我觉得语音不一定就不好。
I don't think the voice is necessarily low.
我觉得关键在于提示里包含多少信息，你知道，就像模型可以……
Like, I think it's like more like how much information is in the prompt, you know, like like the model can...
你可以中途加几句话，比如‘哦，其实我改主意了’，它也能完全跟上。
Like, you can and and like add some sentences and be like, oh, like, actually, I changed my mind like in the middle of the prompt and it will be able to follow that perfectly.
你懂我的意思吧？
You know what I mean?
所以我觉得文本的具体格式没那么重要，但重要的是……
So I think the, like, actual format of the text is less important, but then, like, the ability to...
就是里面包含多少信息。
Like, how much information is in it.
而且我觉得用语音的时候，很多时候，回到人类客服互动的场景，对很多人来说，说话比打字容易多了，你明白吧？
And I think for voice, a lot of times, you know, going back to, like, kind of human agent interaction and, like, for a lot of people, it's just way easier to talk than to like type, you know?
而且我……
And I...
如果这样能从你那里得到更多信息，那就更好了。
If that gets some more information out of you, like that's better.
从某种程度上说，感觉就像是给模型尽可能多的上下文。
At some level, it feels like just giving the model as much context.
对。
Yes.
在开始之前过度提示是一种最佳实践。
Over prompting before you kick off is a best practice.
我不知道。
I don't know.
很多时候，比如我第一次尝试Fable的时候，我花了整整三十分钟，认真写了一个很长的提示。
A lot of the times, like when I was first trying out Fable, I spend a solid thirty minutes, like really crafting a long prob.
我觉得这是模型运行时间越来越长的结果，对吧？
This I think is a response of models running for longer and longer, right?
在它们运行过程中，还是有点难以调整它们。
It's still a little difficult to nudge them as they're in in, like, you know, in the loop.
但我就是凭直觉花更多时间在启动第一个提示词上，然后反复调整它。
But I I just, like, intuitively spend more time kicking off that first prompt and working with it a lot.
我个人觉得，如果我是一个软件工程师，比如说自己开个创业公司，我大概会主要……
My personal opinion is that if I was a software engineer, if I was, like, you know, just running my own startup, for example, I think I would mostly fit...
坚持最多20倍。
Stick to a max 20 x.
你懂我的意思吧？
You know what I mean?
比如，验证和代码审查可能是两回事。
Like, maybe verification and so code review are kind of, like, separate things.
但我看到很多情况是，人们在做这种事时遇到速率限制，就像，哦，它做了很多工作，然后你说，哦，我不喜欢这样。
But I think, like, what I see a lot of times is people hit rate limits when they're doing this sort of like, oh, like, it did a lot of work and you're like, oh, I don't like this.
比如，你能不能撤销这个，重新做一遍？
Like, can you, like, undo this and redo it?
然后你就在这个模型本来可以做到的事情上反复折腾，如果你当初多花点时间或者提供更好的上下文的话，你明白吗？
And and then you're, like, iterating on this, like, thing that the model could have done if you had, like, spent more upfront time or given a better context, you know?
但结果却是，有点像……
And instead, it's sort of like...
你就会说，不行。
You're like, nope.
不喜欢那个设计。
Don't like that design.
试试这个，或者，你是不是搞砸了之类的？
Try this or, like, do you mess this up or something like that?
然后这就会消耗你更多的使用量。
And then that just eats up so much more of, like, you know, your usage.
所以我觉得，这可能是提高效率的一个关键点。
And so that's, like, I think maybe, like, a key, like, to boost for, like, efficiency as well.
对，我觉得，上下文，不仅仅是关于目标是什么的上下文，你知道，这很好。
And, yeah, I think, like, context and not just, like, context on, like, what the goal is, you know, I mean, it's good.
比如，你是在做原型，还是做正式产品？
Like, are you building a prototype or is it, a production thing?
比如，哪些地方可以投入计算资源，哪些地方不能投入？
Like, where can you spend compute or where can you not spend compute?
我觉得有时候你得给模型授权，或者说不授权，因为它并不直观地知道你想在这项任务上投入多少。
Like, I think you have to give the model permission or, like, not permission to do things sometimes where, you know, like, it doesn't know intuitively how much you want to spend on this task.
你可以用努力程度来控制这个。
And you can you can use effort for this.
所以我正在写一篇关于这个的博客文章，就是说，如果你想要……
So I'm working on a blog post about that where it's like, you know, if you want...
比如……
For like...
我们看到努力程度基本上随着任务的复杂性而增加。
We see the effort scales with basically the complexity of the task.
所以在安全方面，高努力程度会带来更多成果。
So for security, effort gets, like, way more results.
比如，高努力和低努力相比，结果会……
Like, high effort versus, like, low effort gets...
这会大大改变评估结果。
Like, changes the evals a lot.
但对于软件工程来说，影响不是特别大，因为大部分精力都花在验证和边缘情况测试上，诸如此类。
But for software engineering, it doesn't change it a huge amount because effort is mostly spent on the verification and the, like, edge case testing and things like that.
所以，能够给模型这样的指导，比如，嘿。
And so, like, being able to, like, give the model that guidance of, like, hey.
这个问题，我觉得你需要花大量时间去验证和做边缘情况测试。
This problem is something that I think I want you to spend a lot of time verifying and edge case testing.
你明白吗？
You know?
那模型混合使用呢？
How about model in the mix?
你知道，有Opus和Fable配合effort。
So, you know, there's Opus and Fable with effort.
还有Haiku也在其中。
There's also Haiku in there.
虽然还不完全是这样，但已经很接近了，我觉得前沿模型几乎会在所有方面都占据主导地位。
It's not quite true yet, but it's very close where I think the Frontier models will be Pareto dominant over, like, almost everything.
你知道，也许……
You know, like, maybe...
而且有时候我觉得，Opus 可能会占据主导地位。
And and sometimes I think I think Opus might be Pareto dominant.
你明白我的意思吗？
Do you know what I mean?
比如，我觉得这取决于情况怎么发展，如果是新版本的 Opus 的话。
Like, I think depending on, like, how things, like, shake out if it's like a newer version of Opus.
但我认为，越来越明显的是，智能模型能用更少的 token 完成简单任务，比其他模型更高效，这主要是因为验证。
But I think that, like, increasingly, it's just going to be, like, the smart model is going to be able to, like, do the simple task for less tokens than this, like, the the other models basically because of verification.
有了验证，在极限情况下，你的模型就不需要验证了。
With verification, in the limit, your model doesn't need to verify.
如果模型是完美的，它只需要做一次工作，然后就像，好的，你……
If it's a perfect model, it just does the work once and it's like, okay, like, you...
我做到了。
I did it.
你知道吗？
You know?
而且随着Fable的发展，我就觉得，哥们，你根本不需要启动Chromium去截图这些东西。
And increasingly with Fable, I'm like, I'm like, dude, you don't need to spin up Chromium and screenshot all of these things.
就像，我看到了。
Like, I I see it.
就像，你做到了。
Like, you you did it.
所以很多……
And so a lot of the...
在更高努力程度下，你会花更多那些token去验证。
At higher effort, you spend more of those tokens verifying.
但如果你在处理更简单的问题，而很多软件工程其实就是这样，你知道的，比如在Fable里，低和中等稳定性下，它可以花更少的token去验证。
But if you're working on simpler problems and a lot of software engineering is, like, you know, like, well, like, in Fable, like, low and medium stability, it can spend less tokens verifying.
而且随着模型越来越聪明，它们基本上就能直接说，好了。
And as the model gets smarter and smarter, they will just be able to, like, alright.
搞定了。
Done.
你懂吧？
You know?
比如，我可以为了保险起见跑一下Lint，但我心里知道代码是能通过的。
Like, I can run the Lint for sanity's sake, but, like, I, like, know it Lints.
你明白我的意思吗？
You know what I mean?
其实你连那一步都不用做。
Like, you don't even need to do that.
而且这样会比那些较小的模型节省很多token。
And that will be so much more token efficient than, like, the the smaller models, basically.
我们这边有没有什么好的做法，可以判断我们是不是用了过多的effort？
Is there a good practice on our side that we can use to see if we're using too much effort?
我特别讨厌在这种事情上浪费时间。
Like, I I freaking hate wasting time on that kind of stuff.
我明白你的意思。
I know what you mean.
我觉得，嗯……
I think, like...
所以在这篇博客文章里，我大概的分配是，代码审查和安全应该设置得更高一些，基本上是这样。
So in in this blog post, my rough distribution is, like, code review and security should be, like, higher max, basically.
然后，软件工程设置按领域来分。
And, like, software engineering settings per domain.
我觉得，如果你做的是UI之类的东西，那就设低一点。
I think, like, if you're doing, like, UI or something like that, like, low.
中等的话，我觉得是你正在构建一个API，你想确保覆盖足够的边缘情况。
And medium, I think, is you're building, like, an API and you want to make sure, like, you cover enough edge cases.
你懂吧？
You know?
所以我觉得，就像我说的，建立这种关于这些分布如何运作的心智模型，确实是工作的一部分。
And so I think building, like I said, that mental model of, like, how things work across these distributions is, like, yeah, part of the job.
这更多是靠直觉还是评估？
This is more intuition driven or eval?
因为我猜这也会变化。
Because I'm guessing this would change as well.
他有评估。
He has evals.
所以我在博客里做的是，基本上把所有的终端基准评估都过了一遍。
So what I did in the blog post is I go over all of the terminal bench evals, basically.
大概有70个问题，我正……
So there are, like, 70 problems, and I'm...
表明，比如，在安全问题上，它做得更多。
Show that, like, okay, you know, like, in the security problems, it does more.
然后我也会看一些转录，主要是看……
And then I also, like, look at some of the transcripts just in terms of, like, how...
它会回答什么？
What does it answer?
它会忘记什么之类的吗？
What does it forget or something?
很多时候，我还有一个提示技巧，就是让它写决策笔记或实现笔记，因为基本上在它遇到的每个评估问题中，它都会想到正确的解决方案，但决定不去做。
And a lot of times, this is another prompting tip I have is, like, asking it to make decision notes or implementation notes because in basically every eval problem that it faces, it thinks about the correct solution, you know, and decides not to do it.
你知道，就像，哦，这是答案。
You know, it's like, oh, like, here's the answer.
如果我这样做会怎么样？
What if I did this?
然后它就想，哦，可能不行，对吧？
And then it's like, oh, probably not, you know?
然后继续往下走。
And then keeps going.
这就是大多数失败的情况，你明白我的意思吧，在更高的最大级别上。
And this is like the majority of the failures, you know what I mean, at like a higher max level.
模型很少会真的不知道怎么做某件事。
It's very rare that the model just doesn't know how to do something.
如果你有这些实施笔记，那你就可以去检查。
If you just have these implementation notes, then you can review.
然后你可以说，哦，其实我希望你做那件你没做的事。
And you can be like, oh, actually, I want you to do this thing that you didn't do.
模型在整体上越来越擅长把这些暴露出来。
The models are getting better at surfacing that overall.
比如，我在Fable 5.1的转录里看到，当它输出这个时，它也会说明自己的决策过程。
Like, I see in the transcripts of Fable 5.1, like, when it does this output, it will call out its decision making as well.
但在框架里把这些做得更明确会更好。
But making this more explicit in the harness is better.
现在我们也允许你修改框架，这样你就可以，比如，添加一些
And now we're, you know, allowing ways of you modifying the harness so you can, you know, like, you know, add some
我们可以说
We can say
在那方面提供帮助的能力。
helpability there.
所以我想提两件你提到的事情，我觉得它们其实超出了提示词的范畴。
So I do wanna call out two things that you mentioned that I think actually exist outside of prompting.
第一是，咱们可以说，那个提示词太重要了，以至于不应该放在提示词里。
One is actually, like, let's let's call it the the prompt that is so important that it shouldn't be in a prompt.
它其实在Claude MD或Agents MD里，对。
It is actually in Claude MD or Agents MD Yeah.
也就是目标。
Which is like goals.
比如你的处境、你的目标、你想要的那些东西。
Like, your your situation, your goals, the things that you want the the the thing.
第二是决策日志或实验日志，或者任何你希望在当前会话之后还能保留的追踪记录，用来完成那些事情。
And then the second of all is the decision log or the experiment log or whatever log of traces that you might want to actually survive the current session to to do those things.
这些就像是外部因素……
Those are, like, externalities that...
没有标准。
There's no standard.
没有……
There's no...
不像技能那样。
It's not like skills.
不像MCP那样。
It's not like MCP.
没有标准。
There's no standard.
它就是一个Markdown文件。
It's it's just like it's a markdown file.
首先，是这样吗？
First of all, is that right?
Claude MD会被淘汰吗？
Is Claude MD going away?
你之前明确表示不喜欢 agents.md，但你还是要做。
You have a documented dislike of of agents m d, but you're gonna do it.
好的。
Okay.
所以，我的意思是，agents。
So, I mean, agents.
md……
M d...
就是说，我们还是要做的。
Like, we're we're gonna do it.
我觉得不同的模型之间差异很大。
I I think it's just like different models are very different from each other.
你明白我的意思吗？
You know what mean?
但我发现维护不同的模型真的很麻烦。
But I realize that it's, like, such a pain to, like, maintain different ones.
你明白我的意思吗？
You know what I mean?
而且，是的，随着模型越来越好，它们完成简单任务的下限也更高了。
And, yeah, like, as the models get better and better, the floor of how they accomplish the simpler task is better.
所以我认为最终，Claude。
And so I do think in the limit, Claude.
Md会消失，而且可能不需要太久。
Md goes away and maybe not even, like, that far.
我觉得现在开始一个新项目，不用Claude可能更好。
Like, I I think, like, I think that right now, it might be better to start a new project without a Claude.
Md。
Md.
对。
Yes.
我觉得，如果你看到非常重复的失败模式，你可以把它们加到你的Claude里。
I think that, like, maybe if you see very repeated failure modes, you add them to your Claude.
嗯。
Md.
真正棘手的是，这每个模型都不一样。
The really tough thing is that this changes per model.
所以，如果你已经添加了一堆失败模式
And so, like, if you've added a bunch of failure modes
或者，甚至像FableMD。
or, like, like, even FableMD.
你需要Opus MD。
You need Opus MD.
或者……
Or...
嗯，甚至Fable 5.1和Fable 5之间，你知道，这挺烦人的。
Well, even Fable 5.1 versus Fable five, you know, like, it is annoying.
我不是说，你知道，我们不是故意这么做的。
Like, I'm not, you know, like, we don't, like, do this on purpose.
你知道吗
You know
我的意思？
what I mean?
这就是模型的工作方式。
It's just like how the models work.
所以，可能Fable 5有Fable 5.1没有的某种故障模式。
And so, like, maybe, like, Fable five had this, like, failure mode that Fable 5.1 doesn't.
如果你一直带着这个记录各种故障模式的运行日志，可能会过度限制Claude，你明白吗？
And if you keep this context with running log of a bunch of different failure modes, they will probably over constrain Claude, you know?
所以，我们其实刚刚为技能添加了评估插件。
And so this is like, we we just actually added evals plugins for skills.
现在你可以评估一个技能是否更好。
And so now you can eval if a skill is better.
我觉得是我们团队的Daisy做的这个。
I think Daisy on our team did this.
所以，是的，我们正在努力解决这个问题。
And so, yeah, this is like, we're trying to work on this.
我们知道这还是要花一些token，而且你知道，它并不完美，但我们是在尽力帮助解决这个问题。
We know it's like, you still have to spend tokens on it and, like, you know, it's not it's not perfect, but it's like, we're trying to help out with this problem.
所以……
And so...
关于提示词，我想给的一个建议是，我经常跟别人说，足够高级的提示词和足够高级的高管沟通其实没什么区别。
And as far as prompting goes, the the one one tip I wanna offer is, it's something I I have told people a lot is sufficiently advanced prompting is indistinguishable from sufficiently advanced executive communication.
所以我其实提到过……
So I've actually referred to...
这是Heavybit的一个高管沟通工作坊，是我职业生涯中见过的最好的。
This is a executive comms workshop from Heavybit that is the best I've ever seen in my career.
他们教了一个叫SCQA模型的东西。
And they teach this thing called the SCQA model.
直接去谷歌搜一下就行。
Just just Google it.
这确实是个东西。
It's a it's a thing.
比如，人们做提示工程已经几十年了。
Like, people have done prompting for decades.
只不过它叫高管沟通。
It's just called executive communications.
就像当一个人需要向组织架构图下层的成千上万人传达信息时，这就是你要做的。
It's like when one person has to communicate to thousands of people down the org chart, this is what you do.
所以，情境、复杂点、问题、答案，这就是写备忘录的方式。
So situation, complication, question, and answer is how you write the memo.
但显然，有时候你没有答案，但至少可以把情境、复杂点和问题列出来，里面也有一些例子。
But, obviously, sometimes you don't have the answer, but you can actually at least list out the SC and Q, and then they have some examples in there.
所以，就给大家留点线索，如果他们想探索的话。
So just leaving breadcrumbs for people if they want to explore.
在继续之前，我想问你，还有没有其他被低估的技巧，或者人们能从Claude Code中获得很多价值但还没用上的方法？
I mean, before we move on, I wanna ask you any other underrated tips, ways people could get a lot of value from Claude Code that they're not using?
我觉得很多都在那个……
I mean, I think a lot of them are in the...
那个不太为人知的文档里。
This unknown, like, doc.
比如，我给了很多示例提示词，像用来头脑风暴，或者之后用来考你。
Like, I gave a bunch of example prompts, like, using it for brainstorming, using it to quiz you after.
我们其实加了一个“像对五岁小孩解释”的技能，这个提示词很短。
We added this, like, explain it like I'm five skill, actually, which is a very short prompt.
它甚至没有说“像对五岁小孩解释”。
And it doesn't even say explain it like I'm five.
就是……
It's like...
基本上，这个提示词的关键是“用几句话讲清楚大局”。
Basically, the keyword of this prompt is big picture is few words.
你知道，这大概就是核心。
You know, like, that's like the the main thing.
效果出奇地好。
And it is shockingly good.
你懂我的意思吗？
You know what I mean?
就像，你……
Like, you...
我觉得我基本上在推特上说过这个。
Like, I think I tweeted about this, basically.
就是那个斜杠 Eli five，你可以把它装成插件。
And it's like slash Eli five and, like, you can install it as a plug in.
但确实，它在直击要害、直接说‘对’这方面要强得多。
But, yeah, it's, like, way better at just cutting through the BS and being like, yeah.
完全正确。
Exactly right here.
所以图表看起来相当清晰。
So the diagrams are, like, quite clear.
我觉得关于 artifacts 的一个事实是，它们放了太多文字，结果人们根本不看这些 artifacts。
I think one of the things that is true with artifacts is, like, they put too much text in and people are not reading the artifacts.
你懂吗？
You know?
所以这个就简化了很多。
And so, like, this simplifies it a lot more.
对，这个其实是 Anthropic 的人在处理一些非常复杂的事件时，发现搞不清楚到底发生了什么，然后就有了这个。
And, yeah, this came out of, like, just people at Anthropic, like, going through very complicated incidents and being like, what is happening?
你懂吗？
You know?
所以我觉得这个很棒。
So this one I think is great.
我的版本其实是……
My version of this is actually the...
就是测试你的理解，给你几个选项，然后如果你选错了，你就会发现你脑子里想的和实际情况对不上。
It's like test your understanding, give you a few choices and then, like, if you actually get it wrong, you have a mismatch between what you think is happening versus what's actually happening.
我觉得这是那种大家都爱聊，但真正去做的人很少的事情。
I think this is one of those things that everyone loves talking about and then very few people really do.
比如，我……
Like, I...
你知道，我觉得……
You know, I think...
真的吗？
Really?
但大多数人就是不想被考来考去的，你懂我的意思吧？
But most people just don't want to get quizzed about something, you know what I mean?
可惜，我觉得这是我们确实需要……
Unfortunately, I I I think this is one of the like things that we need to like...
那反过来，不问你问题呢？
What's the opposite of ask you as a question?
在事情发生之前先问你问题。
Ask you as a question before the thing.
这是在之后
This is after the
的事情。
thing.
没错。
Exactly.
这是一个保持脚踏实地的好方法，比如，你真的知道自己在做什么吗？
It's a good way to stay grounded of like, do you even know what you're doing?
最糟糕的情况是，别人给你发来一堆乱七八糟的东西，而他们根本没搞清楚自己在问什么，或者输出是什么。
The worst case is when people send you slop and they haven't understood what they're asking for or what the output is.
然后就像，兄弟，我可不想读这个。
And it's like, dude, I don't want to read this.
你甚至知道这是什么吗？
Do you even know what it is?
所以，你知道，你给自己定个规矩，在发送东西之前，你至少要知道里面实现了什么。
So, you know, you make it a rule for yourself that before you send stuff, you should at least know what's implemented.
是的。
Yes.
但你可以把它做成一个模组，然后自己构建模组来确保你测试过它。
But so you could make this a mod and and you could build your own mod to, like, make sure you you test it.
所以，是的，我们可以
So, yeah, we can
来聊聊这个。
talk about that.
我们直接进入正题吧。
Let's get right into it.
什么是Claude Mod，这个图展示的是什么？
What is Claude Mod, and what is this diagram showing?
好的。
Okay.
Claude Mods基本上就是你可以自定义整个Cloud Code框架，然后我们要……
So Claude Mods is basically you can customize the entire Cloud Code harness, and we're going to...
如果你有需求，尽管告诉我们。
If you have request, we will, like, let you, you know, like, please let us know.
我们会不断增加更多功能。
We'll add more and more.
这个适用于命令行工具。
This works for a CLI.
也适用于桌面版。
It works for desktop.
也许将来会支持Claude Tag。
Maybe it will work for Claude Tag in the future.
这个我也不确定。
I don't know.
我们是想让这个功能非常实用。
Like, you know, like, we're trying to make this very, very sensible.
你可以看看这份参考表。
You can see this reference sheet.
我不想让大家被它搞得不知所措。
I don't want people to get overwhelmed by it.
你明白我的意思吗？
You know what I mean?
从宏观层面来说，你可以自定义测试框架的执行方式，也可以自定义它的界面。
At a high level, you can customize both the execution of the harness and the UI of the harness.
所以，比如你提到Boris的那个俄罗斯方块例子，那就是在自定义界面。
And so, like, you say on that Tetris example from Boris, that's like customizing the UI.
基本上就是在游戏里展示俄罗斯方块。
Like, showing, like like, basically Tetris in the game.
但是，比如说，你想做这样一件事，就是……
But, like, let's say that you wanted to do this thing where you had...
在每个项目之后测试你的假设，或者检验你的理解。
You tested your assumptions or, like, tested your understanding after every project.
那你要做的就是让Claude来制作这个插件。
What you would do is you would ask Claude to make this plugin.
基本上，每次提示之后都会运行一个分类器。
It would spin a classifier after every prompt, basically.
所以，在每一轮结束时，你会启动一个子代理，或者说一个分叉代理。
And so, like, at the end of each turn, you would spin off a sub agent or, like, a forked agent, basically.
分叉代理会保留提示缓存。
A forked agent is, like, maintains the prompt cache.
所以这就像……
So it's like a...
其中一个不太直观的地方是，你可以分叉并发出一个小请求，而且成本很低，因为整个提示缓存都已经准备好了。
Like one of the unintuitive things where you can fork and do like a little request, and they'll be very cheap because the entire prompt cache is like done.
然后你就可以问，这个任务完成了吗？
And so you can be like, has this task been completed?
对，这就是你做BTW那些东西的方式。
Like That's how you do BTW and all those.
底层的分叉代理。
The underlying fork agent.
对。
Yes.
但是……
But...
所以你可以……
So you can...
在分叉的子代理里，你可以问，比如，这个任务完成了吗？
In the fork sub agent, you can say like, has this task been completed?
如果完成了，就返回真。
If so, return true.
然后，在你的钩子或者插件模块里，抱歉，可能是在子代理里，你会说，如果为真，就给我出个测验。
And then in your hook or in your like plug in mod or sorry, like, in the sub agent probably, you'd say, like, if true, give me a a quiz.
你知道的，给我问题和答案。
You know, give me questions and answers.
然后，用JSON格式，你解析它，然后基本上显示在提示输入框上方，就是这些问题列表。
And then, like, in a JSON format, and then you'd parse it, and then you'd display above the prompt input, basically, this list of questions.
所以这有点消耗token，因为你知道，每次助手回合结束后都得做一次。
And so this is something that's, like, slightly token intensive because, like, you know, you have to sort of do it after every end of the assistant turn.
但这只是一个轻量级的分类，然后你可以得到这个测验，而且你会发现，Claude总会帮你搞定。
But it's like a lightweight classification, and then you can, like, you know, get this quiz, and then you'll see, like, you know, Claude will always do it for you.
你不需要记得去做。
You don't need to remember to do it.
我们聊过很多这样的技巧，对吧，比如那些实现笔记。
There are lots of these, like, tips that we've talked about, right, where it's like, oh, implementation notes.
现在你也可以为实现笔记添加一个工具。
You can also add a tool for implementation notes now.
所以我正在添加的这个工具，就是注册，我觉得我把它叫做“假设”，但也许我会改改。
And so, like, this tool that I'm adding is, like, register, like, I think assumption is what I'm calling it, but, like, maybe I'll change it around.
这是一个模组。
And this is a a mod.
然后你给它一个注册假设的工具，它就会维护一个列表。
And so, like, you give it a register assumption tool, and then it will keep a list.
它会……
It'll...
每次它这么做的时候，它都会保留一个……
Every time it does it, it'll keep a...
比如，添加到列表里，然后在最后，它会显示这些假设。
Like, add to the list, and then at the end, it will display those assumptions.
你明白吗？
You know?
我还在做的另一个模块是模型路由器。
Another mod I'm working on is model router.
就是内部的，比如，Claude 模型路由。
And so, like, internal, like, Claude model routing.
所以它是……
So it's...
这是……
This is...
我想说，我们默认不做模型路由的原因是，这其实是个难题。
I want to say the reason we don't do model routing by default is, like, it's a hard problem.
你懂我的意思吧？
You know what I mean?
而且，就像……
And, like...
你肯定会搞错的。
You will get it wrong.
你……
You...
就像……
Like...
你可能会不小心把Fable用在难题上，或者把Sonnet用在……
You will, like, accidentally use like Fable for a hard problem or Sonnet for
你有自动审批，但没有自动模式。
You have auto approvals, but you don't have auto mode.
嗯，你会有自动……
Well, you will have auto...
就是说，你没有那种自动路由，或者……
Like like, you don't have like auto routing or
什么的。
something.
模型选择的自动模式。
Auto mode for model picker.
对。
Exactly.
对。
Exactly.
所以我……
So I'm
就面临一个棘手的问题：这个功能要开放到什么程度，用户需要为此考虑多少？
getting the rough question of like, how much do you open this up and how much do people have to think about this?
比如你提到提示缓存和构建路由器，感觉你可以轻松构建一个按查询路由的模块。
Like when you talk about prompt caching and building a router, it seems like you could easily build a mod that routes per query.
而且我很快就放弃我的计划了。
And I'm just killing my plan very fast.
我想问的更多是，这样的产品文档看起来是什么样的？
I guess my question is more so like, what is like a product doc like this look like?
它是给谁用的？
Who is it for?
是给高级用户用的吗？
Is it for power users?
还是每个人都应该能够轻松使用？
Is it everyone should be able to go Definitely through
高级用户。
power users.
我觉得是高级用户，但Claude Code的特点是有很多人都是高级用户，你知道，因为分享东西很容易，比如你可以...
I mean, think it is power users, but like the nature of Claude Code is that so many people are power users, you know, because it's easy to share things, you know, like you can...
比如一个人就能做出一个不错的模型路由工具，不会老是破坏提示缓存，然后你可以把它们组合起来用。
Like one person can make a good model router thing that doesn't break prompt cache all the time and then you can like, you know, sort of compose them.
插件的另一个好处是它们可以互相连接和组合。
Another cool thing about the plugins is that they can hook into and compose with each other.
所以我有一个模组，它会在顶部创建一个模式选择器，任何插件都可以注册成为一个模式。
And so I have, like, a a mod that will, like, create a mode selector at the top, and any plugins can register to be a mode.
所以自动路由器也可以是一个模式。
And so, like, the auto router can be a mode.
或者你可以有一个像工件模式那样的模式，就是那种……
Or, like, you can have a mode that's like artifact mode where it's like...
它主要用工件跟你交流。
It it primarily talks to you in artifacts.
就像你可以在计划模式之间切换一样。
It's kind of like, you know, you you can toggle between plan mode.
你明白我的意思吗？
You know what I mean?
所以，你可以创建越来越多的这种模式。
And so, like, you can create more and more of these modes.
但创建模式的能力本身就是一个模组。
But the ability to create modes is in itself a mod.
你懂吗？
You know?
所以这里有很多丰富的内容，但我们确实想让它变得相当简单。
And so there's a lot of richness here, but we do want to make it fairly easy.
我们希望做到……
We want to be...
让你可以直接安装别人的。
Make it so that you can just, like, install someone else's.
你可以和Claude聊天，我们会确保它理解像提示缓存这类东西的细微差别，这样它就能提醒你。
You can you can chat with Claude, and, you know, we'll, like, make sure that it understands the nuances of things like prompt caching and stuff so it can, like, warn you.
我觉得这对Claude来说并不是特别复杂的行为。
This is, like, not extremely complicated behavior for Claude, I think.
但我们应该掌握制作模组的技巧。
But we should have just a good skill on how to make mods.
是的，我们看看进展如何。
And, yeah, we'll see how we go.
我确实认为这是可变软件的一个预览，就像生成式软件一样，你可以安全地进行定制。
I I do think that this is, a preview of, like, mutable software, you know, and like how, like, generative software, just like you can customize safely.
如果启用的话，你可以定制任何软件。
If enabled, you could customize any piece of software.
我认为越来越多的应用理想上应该这样做。
And I think that more and more apps ideally do something like this,
你知道吗？
You know?
顺便说一句，我们还有一条很酷的推文，讲的是无限金钱按钮，就是让你的SaaS可以被代理使用。
And by the way, we have we have another cool tweet about how, you know, there's the infinite money button, which is, like, make your SaaS consumable by agents.
我觉得可变软件很有趣，而且其他人也尝试过这样做。
I think mutable software is interesting and, you know, other people have also tried to do it.
我觉得难点在于，当你能做所有事情的时候，人们就会……
I think the hurdle comes when you can do everything, then people...
用户会……
Users get...
容易感到困惑。
Tend to get confused.
所以通常，真正管用的就是那种单一、明确的流程。
So usually, the stuff that works is just like one opinionated flow.
而这个是偏向于不那么明确的方向。
This is in the side of less opinionation.
就是说，给高级用户更多权力。
It's just like, well, more power to power users.
而且我觉得这可能是由AI解锁的，比如你可以直接提示你想要什么。
And I think probably unlocked by AI where, like, you could just prompt for whatever the thing it is.
或者可以有一个技能来提供这些明确的意见，你知道的，
Or there can be a skill that gives the opinions, you know,
对。
and Yeah.
然后……
And then...
所以，稍微了解一点TypeScript和构建系统这些东西的话，最接近的……
So knowing a little bit about, like, TypeScript and build systems and all these things, the closest...
我……
I...
我其实很好奇，做这个的团队，我不知道你跟他们熟不熟。
I'm actually very curious that the the team who worked on this, I don't know how close you were to them.
他们有没有从像Babel、Webpack这些老牌构建系统里汲取灵感，因为听起来很像，比如插件生态，那些可以互相组合的东西。
If they drew any inspiration from build systems like Babble, Webpack, all these all these, old school things, because it sounds very similar, like, the plug in ecosystem, those things where they can compose to each other.
怎么说呢，我对技术细节不太深入，但我知道这是Bund团队和Claude Code团队合作的结果，所以……
I mean, I'm not deep in the technical details, I do know it was a collaboration with someone on the Bund team and someone on the Claude Code team and so......
系统麦加。
System Mecca.
没错。
Exactly.
这真的很令人兴奋。
It's it's very exciting.
但现在，智能体就能实现这种非常复杂的软件扩展性。
But, yeah, like, agents can just do this very complicated sort of, like, extensibility into your software now.
所以，这也是另一个理由，比如……
And so, yeah, like, you know, another reason to, like...
如果你在运营一家初创公司，你完全可以告诉Claude，嘿，我们能不能做一个扩展系统？
If you run a startup, like, you can just prompt Claude and be like, hey, like, could we make an extension system?
那会是什么样子？
Like, what would that look like?
你懂吧？
You know?
而且我真的很想……
And I just really want to...
比如，你以前有钩子（hooks）
Like, you had hooks in the past
还有插件，所有这些。
And plug ins, all these things.
那么，模组具体能做什么，而这些是那些东西做不到的？
So what specifically will mods be able to do that those things could not do?
其实，我们内部最初把这个叫做函数钩子。
Internally, we were originally calling this function hooks.
所以，这大概能让你有点概念，钩子基本上就是注册一个事件，然后调用一个脚本。
And so, like, that's, like, gives you a little bit of an idea where, like, hooks sort of register a, like, an event to happen and then, like, a script to call, basically.
而这个基本上...
And this basically...
在TypeScript运行时内部运行。
Inside of the, like, TypeScript runtime is running things.
所以，你能得到一些好处，比如，作用域里有很多东西，比如，这个对话有多少轮，用了多少token，等等。
And so, like, you get some benefits of just, like, it has a bunch of things in the scope with, like, for example, like, how many turns is in this conversation, right, like, how many tokens have been used, like, etcetera.
比如，有哪些消息，诸如此类。
Like, what are the messages, things like that.
所以有很多消息可以用。
So has a bunch of messages that can be used.
然后基本上就是更多的钩子。
And then it's just, like, a lot more hooks, basically.
所以我们有，你知道，像……
So we have, you know, like...
或者很多……
Or a lot of...
更多，你知道，功能……
Lot more, like, you know, funk...
就是你可以注册的那些东西。
Like, things you can register on.
然后你可以做……
And then you can do...
因为……
Because of the...
因为这一切都在进程内发生，你可以生成子代理，你知道的，带着四个上下文和背景之类的。
Because it's all happening in process, you can spawn sub agents, you know, with four contests and contexts and stuff.
然后，那会返回结果。
And, like, that will return.
你可以解析那些结果。
You can parse the results of those.
你可以……
You can...
你结构化输出，以便返回它们。
You structure output to sort of, like, return them.
然后你可以修改界面，这在hooks里是永远做不到的。
And then you can modify the UI, which you can never do in in hooks.
所以
So
所以修改界面，这就是你展示俄罗斯方块例子的原因。
So modify UI, this is why you showed the Tetris example.
那它是不是也适用于工件？
Does it also extend to artifacts?
我猜应该是的。
I assume it does.
你……
You...
其实，工件有点像另一种定制方式。
Like, artifacts are kind of like a different way of customizing it.
你知道，比如，我正在做的一个模组是仪表盘模组，它会提示Claude维护一个仪表盘，那就算是一个工件。
You know, like, you can definitely one of the mods I'm working on is, like, this dashboard mod, which will, like, sort of prompt Claude to maintain a dashboard, that's an artifact.
但它们有点，怎么说呢，稍微有点不同，或者说不是完全不同。
But they're kind of, like, slightly orthogonal or not orthogonal.
它们会以不同的方式相互配合。
They they compose with each other in different ways.
模组更像是，嗯，在你的Claude Code环境里，改变代理循环，你知道的，而UI只是一个附加的好处。
Like, mods are, like, a little bit more, like, in your Claude Code harness, changing the agent loop, you know, and, like, the UI is, like, an added benefit.
而工件就是，你想，嗯，从高层次看东西，非常交...
And then artifacts are just, like, you want to, you know, see things at a high level, very inter...
高度交互。
Highly interactive.
你知道，这些交互方式可以比TUI甚至我们的桌面端大得多。
You know, like, the affordances can be a lot bigger than, you know, like a TUI or even in our desktop.
我猜你会写一篇很好的博客文章来区分它们。
I'm guessing you'll have a good blog post on the differences.
因为现在，你也可以，嗯，做一个循环，输出到一个交互式仪表板的工件，但你也可以用模组来做。
Because right now, you can also, you know, make a loop that outputs to an artifact that's an interactive dashboard, but you can also do it with a mod.
只是需要思考一下如何做一个黑客...
There's just some thinking about making a hack...
在我们对那个环境不太了解的情况下，去修改它。
Hacking on a harness when we don't know much about the harness.
说到mods，我特别兴奋的一点是，云端代码有太多东西需要你记住。
Well, something I'm excited about with mods is, like, there's so much things of cloud cloud code that you just have to remember.
你懂我的意思吗？
You know mean?
你会想，哦，让我做这个，然后调用那个仪表盘缩放功能，它执行一个循环，诸如此类。
You're like, oh, like, let me do this and then let me call the dashboard scale that does a loop and things like that.
然后……
And...
或者，之后让我验证一下我的假设。
Or, like, let me test my assumptions afterwards.
我觉得，如果你用这些小分类器之类的东西来做所有这些事情，你会想，这些是我关心的。
And I think, like, if you do all of these things using these little classifiers and stuff, and you're like, these are the things I care about.
这是我想做的。
This is what I want to do.
你可以，比如……
You can, like...
你就不用记那么多东西了。
You don't have to remember as much.
还有一个我正在做的模组，就是“下一步”模组，它……
One more, like, mod I'm working on is a next steps mod that
我有个……
I have...
我刚想说，我有一个……
I was gonna say I have a
“下一步”技能。
next step skill.
我经常用“下一步”功能。
I always run next steps.
那它能访问你的技能吗？
And does it have access to your skills?
这就是那种让我觉得……
Like, this is one of those things where I'm like
我觉得是的。
I think so.
好的。
Okay.
我的意思是，
I mean,
它可能有特定的访问权限，
it probably specific access to
嗯，你觉得它总是有吗？
Well, think there's always has?
我觉得可能需要特定的提示，才能知道你的技能之类的。
Think there's, like, specific prompting, I guess, to, like, know your skills kind of.
就像，我觉得Claude有时候会在对话过程中忘记这些。
Like like, I think Claude forgets them sometimes throughout, like, the thing.
但不管怎样，这个想法就是，对，下一步行动，同时也会提醒说，嘿，这件事已经发生了。
But anyways, the idea of, like, yeah, next steps that also are like, oh, hey, this has happened.
用解释技能来告诉你发生了什么，因为这看起来挺复杂的，你知道吗？
Use the explain skill to explain to you what happened because this seems like quite complex, you know?
或者，对，用你的未知技能。
Or like, yeah, use your unknown skill.
看起来你像是在让模型去迭代这些小改动，感觉你可以把提示词写得更好。
It looks like you are like asking the model to like, you know, iterate on these small changes, it seems like you could prompt better.
你知道吗，比如，如果你这样做会怎么样？
You know, like, what if you did this?
所以，我觉得……
So, like, I think...
比如，在那里投入更多的计算资源。
Like, spending more compute there.
而且它应该总是以选择题的形式出现。
And it should always come out as multiple choice.
我们有一个……
We have we have a...
我有他的技能。
I I have His skill.
我的下一个技能是这样的。
My next skill is like this.
好的。
Okay.
完美。
Perfect.
你可以偷走它。
You can steal it.
但是，对我来说，这一切都是……
But, like, but, like, for me, it's all...
我觉得模型真的总是需要被提醒，你在这里想做什么？
I think models really always need to be reminded, what are you trying to do here?
看看整个对话记录，然后想，哦，这是最初的目标吗？
Look at the whole transcript and go, like, oh, was this original goal?
你的解决方案真的解决问题了吗？
Did your solution actually solve it?
你是不是偷懒了？
Were you lazy?
如果你偷懒了，也许是有原因的。
If you're lazy, maybe there was a reason.
也许你需要我的认可，也许你需要……
Maybe you needed approval from me, maybe you needed...
有两件事你想提出来。
There's two things you wanna suggest.
所以这有点像修改那个“问你问题”或“采访我”的技能。
So it's it's a little bit like the modification of the ask you the question or interview me skill.
所以这是下一步。
So it's next steps.
没错。
Exactly.
而且再说一次，用mods来做的好处是，你可以用第四个sub agent，这样它之后就不会留在上下文里了。
And and again, the benefit of doing it with mods is you can do it a fourth sub agent and so it doesn't remain in the context afterwards.
所以你有这种想法，比如，模型在执行任务，然后你几乎像一个监督者，你知道，就是确保你能把接下来的步骤做好。
So you have this like idea of like, okay, the model is doing its execution and you have this almost like supervisor, you know, like, that is like making sure that you can do like the next steps well.
所以……
So...
我确实有两个面板，而且我经常尝试有一个监督者的东西，保持高层上下文，然后实现部分，对。
I do have two panels and like, I I I often try to have a supervisor thing, keep the high level context and then the implementation Yeah.
细节放在另一个agent里。
Detail in in another agent.
我觉得随着模型的变化，很多东西都会被抽象化，你知道的……
I feel like a lot of this abstracts away as models change, you know, the...
半小时前，你提到了“驾驭工程的惨痛教训”，而现在我们感觉又走到了另一个极端。
Half an hour ago, you said bitter lesson of harness engineering and we're on the other extreme right now, I feel.
嗯，所以……
Well, so so...
对。
Exactly.
如果一切都可以自定义，那Claude Code到底是什么？
If everything's customizable, what actually is Claude Code?
这就是我昨晚跟你聊到的。
And which I talked to you about last night.
我觉得这其实是……
I mean, I think that this is...
我觉得苦涩的教训有点反直觉，你懂我的意思吗？
I think the bitter lesson is unintuitive, you know what I mean?
就是说……
In terms of like...
我……
I...
而且，我们在这里有点误用了“苦涩的教训”这个概念，就像……
Also, like, we're kind of misusing a little bit of the bitter lesson here where it's like...
这更多是关于扩展和计算能力之类的事情。
It's more about like scaling and compute and stuff.
但我觉得，有些东西就是……
But like, I think there is something where it's just like...
我在这里用它来打个比方，就是说那些工具很快就会过时。
I think I use it as an approximation here to say that harnesses go out of date very quickly.
你明白我的意思吗？
You know what I mean?
而且，怎么……
And, like, how...
但它们的变化方式并不直观。
But how they change is unintuitive.
你知道吗？
You know?
所以，最明显的例子就是从聊天到智能体，你得给它们全新的工具。
And so, like, the big obvious example is, like, from chat to, like, agents where you had to give them entirely new tools.
但是，我觉得这个新版本，哦，它能修改自己的工具集。
But, like, I think this new version of, oh, it can modify its own harness.
这就是一个自我工具集循环。
This is, an own harness loop.
这是利用它自身能力的一种方式。
It's, a way of using its capabilities.
或者，它能构建一个工件。
Or, like, it can build an artifact.
而且，我觉得我的理解是，模型越来越聪明，现在它们比一般的软件工程任务要聪明得多。
And, like, I think the way I think about it is, like, the models have more and more intelligence, and they're, like, so much more intelligent now than, like, the average software engineering task.
你看那些终端基准测试，它们能解决像雅可比猜想这样的问题。
Like, you look at the, like, terminal bench ones, and they're, like, solve, like, the Jacobian conjecture.
虽然不是真的，但你知道，那些任务相当复杂。
Not really, but, like, know, it's like, they're they're quite complex.
说实话，作为软件工程师，我可能都做不了这些。
Like, I would not have been able to do this really as a software engineer.
关于 Bayt 播客
Bayt 提供中文+原文双语音频和字幕，帮助你打破语言障碍，轻松听懂全球优质播客。