A 3x CPO's Claude Code Skills for Building an AI-Native Product Team

频道: Aakash Gupta
视频: https://www.youtube.com/watch?v=Eo26_4JcyNA
原文语言: en
统计: 共 72 轮 · Oji 40 · Aakash 28

[0:00] Oji

The era of monoskilled professionals is dead. So I'm going to demonstrate the future of product builders.

只会一种技能的专业人士，这个时代已经结束了。所以今天我要演示一下，未来的产品建造者是什么样子。

[0:07]

Meet Aji Udway, former CPO at Typform, Kalanley, and Parable as well as former head of product of creation and innovation at Twitter. How much of this thinking like viability gates and sharp problem tests should we really be outsourcing to LLM? Is LM basically the first draft just to get some thinking going and then we improve and react? You really do need to look through all the output. Make sure that it makes sense for you. We do this all the time because you know I think it's hard to fully trust 100% LLMs.

来认识一下 Oji Udezue，他曾任 Typeform、Calendly 和 Parable 的 CPO，也做过 Twitter 创作与创新方向的产品负责人。——像 viability gate、sharp problem test 这类思考，我们到底该有多少外包给 LLM？LLM 是不是基本上只负责出第一稿、先把思路跑起来，然后我们再改、再反馈？——你确实得把它输出的东西整个过一遍，确认对你来说是成立的。我们一直都这么干，因为我觉得很难 100% 完全信任 LLM。

[0:38] Aakash

Can you open up that scaffolding skill and we can see what is inside it and how it looks.

你能不能把那个 scaffolding skill 打开，让我们看看里面到底装了什么、长什么样。

[0:42] Oji

Yes. So it's it's multi-sklls here. You can see that it has a master skill.md and then it has a bunch of subsklls where it would do market research. I've been a product manager for 25 years and I've never really felt that coding was worth my time until now.

可以。它其实是一组 skill。你能看到有一个总的 skill.md，下面挂着一堆子 skill，比如做市场调研的。——我做了 25 年产品经理，一直到现在，才第一次觉得写代码这件事值得我花时间。

[1:03] Aakash

Before we get into today's show, please take a second to check that you're subscribed on YouTube and following on Apple and Spotify podcasts. If you want access to all of my favorite AI tools, I've gotten them to give you an entire year of their paid plans. Check out bundle.acg.com akashg.com for an entire year of bolt new air table speechify descript magic patterns linear dovetail arise and mobin and now into today's show OG what are people going to learn today if they stay till the end

进入今天的正片之前，花一秒钟确认一下：你已经在 YouTube 订阅、在 Apple 和 Spotify 播客上关注了我们。如果你想用上我最喜欢的那些 AI 工具，我已经替你谈下来整整一年的付费套餐——去 bundle.aakashgupta.com 看看，Bolt.new、Airtable、Speechify、Descript、Magic Patterns、Linear、Dovetail、Ario、Mobbin，全都是一整年。好，进入今天的正片。Oji，如果大家听到最后，他们能学到什么？

[1:36] Oji

well people are going to learn the most important thing they're going to learn is that we won't just focus on code skills the landscape of product right now or coding or engineering or the shipyard whatever you want to call it is full of repos in GitHub with skills to token max or token minimize or to change costs all on the coding layer. But we know that a tech company a successful tech company is really three layers. It's the software and the hardware. It's the product which is about customers and about the business model. And it's about the business. They're thinking about allocating resources to the rest of the chain. What you're going to learn is how to take the raw harness like cloud code and apply not only coding skills but product skills, product judgment on tap and business skills that will help you make the very best decisions about how to build a successful product.

大家能学到的最重要的一点是：我们不会只盯着 coding skill。现在整个产品圈——或者说写代码、做工程、我喜欢叫它 shipyard（交付现场）——GitHub 上全是各种 skill 的 repo：把 token 拉满的、把 token 省下来的、改成本结构的，全堆在 coding 这一层。但我们都知道，一家科技公司、一家成功的科技公司其实是三层：一层是软件和硬件；一层是产品，也就是客户和商业模式；还有一层是生意本身，是怎么把资源分配到整条链路上去。你今天要学的，是怎么拿 Claude Code 这种原始 harness，不只装上 coding skill，还装上产品 skill、随时可调用的产品判断力，以及商业 skill——它们能帮你在「怎么做出一个成功产品」这件事上做出最好的判断。

[2:38] Aakash

So I'm really excited about this. Where should we start?

我对这个特别期待。我们从哪儿开始？

[2:43] Oji

So Aash where I want to start is uh something I think is really important. So we see product mind consults with lots of big companies big and small. And the things that we get asked to do is first of all come in and reconceptualize a product as much more AI native. Well, very quickly what happens to us is that we get pulled into the shipyard like how people are organized, new skills for people in a AI era. How how do things work together? How do people work together? And what we're seeing immediately is the developers are speeding up very quickly, especially if they're early adopters. And then we see everyone else being a bottleneck. And particularly we see PMs who are not speeding up their product judgment, speeding up their orchestration skills to match the new speed of the engineers. So what I'm about to show you is uh we made a set of product judgment skills not just code skills but product judgment and business skills that product managers can use to think at the business layer at the product layer while not sacrificing code fidelity and things like testing and quality and so on. The best instantiation of this is the pro new project scaffolding skill. There are all kinds of skills here. We have things like finding the aha moment for your new product, whether it's aic or not. We have things like figuring out whether there's a sharp problem. But the scaffolding skill is special because it starts with you describing a business problem and then it really makes decisions, helps you make decisions whether it does market research for you, tells you whether it's a viable problem and how to solve it if it's not a viable problem. Uh makes architectural decisions for you based on asking you questions.

Aakash，我想从一件我觉得特别重要的事讲起。Product Mind 给很多公司做咨询，大公司小公司都有。人家找我们做的第一件事，通常是进去把一个产品重新构想成更 AI-native 的形态。但很快我们就会被拽进 shipyard 里——人是怎么组织的、AI 时代每个人需要什么新技能、各个环节怎么咬合、人跟人怎么协作。我们立刻看到的现象是：开发在飞快提速，尤其是那些早期采用者。然后其他所有人都成了瓶颈。特别是 PM——他们的产品判断力没有提速，编排（orchestration）能力也没有提速，跟不上工程师的新速度。所以我接下来要给你看的，是我们做的一整套产品判断力 skill，不只是 code skill，而是产品判断和商业判断的 skill，让产品经理能在商业层、产品层上思考，同时又不牺牲代码的严谨度——测试、质量这些都不丢。这套东西最好的一个体现，就是 new project scaffolding skill。这里各种 skill 都有：有帮你找新产品 aha moment 的，不管这产品是不是 AI 的；有判断问题够不够 sharp 的。但 scaffolding skill 很特别，因为它的起点是你描述一个商业问题，然后它真的会做决策、帮你做决策：替你跑市场调研，告诉你这个问题是不是值得做；如果不值得做，那该怎么解。它还会通过问你问题，替你把架构决策定下来。

[4:32] Oji

figures out how to test the the thing right based on those architectural decisions and even sets up you know continuous integration and continuous delivery for you. So we're going to work directly in cloud code not in cloud desktop or anything just in the same way that uh developers work. So what I'm going to tell this thing right now to scaffold a new project. One of the things I'm obsessed about is the fact that people won't read any code. So what I want to build is either a SAS or an agent that helps vibe coders, people who don't have a lot of software experience, helps them figure out if their code is any good. Security, robustness, complexity, simplicity, all of that tells them the story of their code because no one is looking at it even in a corporate environment. And so hopefully if we tell it to scaffold this, it will run all the processes of the scaffolding for us. I'm going to hit enter. If you pause and read it, you'll see that it's just a a first idea. And so what uh this is going to do is it's uh it will run a viability gate. It would take a look at some frameworks and try to figure out if this project is even viable at all. And so that first gate um is still sort of uh loading the skill. There you go. It's an 11step workflow.

它会根据这些架构决策，推导出这东西该怎么测，甚至连持续集成、持续交付都替你搭好。我们全程直接在 Claude Code 里干活，不用 Claude Desktop 之类的，跟开发者的工作方式一模一样。我现在就让它 scaffold 一个新项目。我一直很执着的一件事是：根本没人会去读代码。所以我想做的，要么是一个 SaaS，要么是一个 agent，帮那些 vibe coder——那些没多少软件经验的人——判断自己的代码到底行不行：安全性、健壮性、复杂度、简洁度，全都讲清楚，把他们代码的来龙去脉讲给他们听。因为根本没人在看代码，哪怕在公司环境里也没人看。所以理想情况下，我让它 scaffold 这个东西，它就会替我们把 scaffolding 的全套流程跑一遍。我按回车。你要是暂停下来读一读，会发现我输进去的就是一个很粗的初步想法而已。接下来它要做的是跑一道 viability gate——它会调一些框架出来，判断这个项目到底成不成立。第一道 gate 现在还在加载 skill……来了，是一个 11 步的工作流。

[5:52] Oji

That makes sense because it's a very powerful skill and it'll skip straight to the viability gate. Let's see.

这说得通，因为这个 skill 本身分量很重。它会直接跳到 viability gate，我们看看。

[6:00] Aakash

Can you open up that scaffolding skill and we can see what is inside it and how it looks.

你能不能把那个 scaffolding skill 打开，让我们看看里面到底装了什么、长什么样。

[6:05] Oji

Oh yeah, let's do that. So if you look at the skills, you'll see a bunch of the product uh product skills um sharp problem. So the first one is yes. So it's it is it's multi-sklls here. You can see that it has a master skill.md and then it has a bunch of subsklls where it will do market research. It will actually fetch skills for you, new skills for you if you need it for your project type. It will set up testing do you know and you know the core skill is to basically do viability gate. So let's go back to see what it's doing here. And so what it's trying to do is to do web search. So I'm going to tell it that it can absolutely do web search in order to find out what's going on. So it's conducting market research. So I think I interrupted it. So I'm going to continue. And it's going to conduct market research on the idea. And the next thing it's going to do is to write a product brief. And once it's done with the product brief, it will start to get to the code level at all. Uh claw takes its time a little bit, but what it's doing is that it's spinning up a web search session. Sometimes it would do that through Chrome. Uh if you have Chrome tools installed, sometimes it'll do that itself with HTTP.

好，我们看看。你打开 skills 目录，就能看到一堆产品类的 skill，比如 sharp problem，第一个就是它。对，它确实是一组 skill：有一个总的 skill.md，下面挂着一堆子 skill——做市场调研的；它还会替你去抓 skill，如果你的项目类型需要新的 skill，它会给你抓回来；它还会帮你把测试搭起来。而核心 skill 基本就是跑 viability gate。我们回去看它现在在干嘛。它想做的是 web search，所以我告诉它完全可以联网搜，去查清楚现在这个领域是什么情况。它正在做市场调研。我刚才好像打断它了，那我让它继续。它会针对这个想法做市场调研，接下来会写一份 product brief。写完 product brief 之后，它才会开始下沉到代码层。Claude 会花一点时间，但它现在做的事是起一个 web search 会话。有时候它会走 Chrome——如果你装了 Chrome 工具的话；有时候它自己用 HTTP 去抓。

[7:43] Oji

Um, and of course, um, I didn't give Claude all my permission, so it's there. So, yeah. So, the viability gate, this is going to be interesting. It's going to tell me whether this thing is worth my time or worth your time very specifically. And you have to understand that what it's really doing is taking frameworks from product management from our book and it's trying to see is this thing worth spending time on. Some of the dimensions include things like the frequency of the problem, whether we have a clear customer. Um, and there we go. Let's run that. So what does it what does it tell us actually with the viability gate? Viability gate pass proceed. Zero weak, three strong, three moderate, not a silent pass. Three moderates are flagged as d-risking agenda. Now step four, market research. And if we go into the repo, we will start to see this document show up. So if we go away from skills and go to code memo one um it'll start to actually write for us what it's finding out from doing a web search. So right now we know that the it's usually if there three weak things it will it will it will kill it. will recommend you stopping but according to it uh the viability gate is strong. So it has like six dimensions revenue technical feasibility differentiation competitive landscape target user definition and the problem clarity and urgency. So if we go back to its uh its work, it's literally going in and doing deeper market research for you now. It's scanning competitors, is scanning pricing, is scanning what people have done in the space.

当然，我没有把所有权限都给 Claude，所以它卡在那儿了。好，viability gate，这部分会很有意思。它要告诉我这东西值不值得我花时间，更确切地说，值不值得你花时间。你得明白，它真正在做的事，是把产品管理的框架——从我们那本书里来的框架——拿出来，去看这件事值不值得投入时间。它的维度包括问题出现的频率、我们有没有一个清晰的客户，等等。好，出来了。我们跑一下。viability gate 到底告诉我们什么？viability gate 通过，可以继续：零个弱项、三个强项、三个中等，但不是无条件放行——那三个中等项被标成了 de-risking（降风险）待办清单。现在进入第四步，市场调研。我们进 repo 就能看到这些文档陆续冒出来。从 skills 切到 codememo 那边，它会把 web search 查到的东西真的写下来给我们。目前我们知道，通常只要有三个弱项，它就会把这个想法毙掉，会建议你打住；但按它的判断，这次 viability gate 是强的。它一共有六个维度：营收、技术可行性、差异化、竞争格局、目标用户是否清晰，以及问题的清晰度和紧迫性。回到它的工作区，它现在真的在替你做更深的市场调研：扫竞品、扫定价、扫这个领域里别人已经做过什么。

[9:59] Oji

And so right here, it's it's created it. It's done your market research package for you. Um, I set it so that it's not narrating itself in full, but you can see it has a market overview. It has trends shaping the space. It has direct or near competitors and what they're charging for. It has key takeaways. It has adjacent products. This is really, really comprehensive, guys, about what it's doing for you. And now is writing a product brief, right? Uh we have tools that help you write a PRD. This is literally your new project brief where it codifies that into a problem into target customers val core value proposition success criteria that will help you know whether you will succeed or fail and then starts to write goals for you that will help you sort of gauge where it's going. Okay, now it's done that. It's done some of the business level skills and what it's working on now is trying to find out how to scaffold the rest of the project in code for you. So now to the code layer, you can see that code memos core is an LLM analyzing repos for claw with current correct current model ids.

看，它做出来了，市场调研包已经替你做完了。我设置成它不会把自己每一步都完整念一遍，但你能看到里面有市场概览、有正在塑造这个领域的趋势、有直接竞品和相近竞品以及他们各自怎么收费、有关键结论、还有周边产品。各位，它替你做的这些东西是真的非常全面。现在它在写 product brief。我们有专门帮你写 PRD 的工具，而这一份就是你的新项目 brief：它把想法固化成问题定义、目标客户、核心价值主张、成功标准——这些能让你判断自己会成还是会败——然后开始帮你写目标，让你能衡量事情走到哪一步了。好，这部分做完了。商业层面的一些 skill 它已经跑完，现在它在琢磨怎么把项目剩下的部分用代码 scaffold 出来。所以现在进到代码层了，你能看到：CodeMemo 的核心是一个用来分析 repo 的 LLM，跑在 Claude 上，而且用的是当前正确的 model id。

[11:30] Aakash

So the scaffolding skill is basically an orchestrator skill. So, it seems like it's calling some of these other skills, creating documents, walking through this 11step process.

所以 scaffolding skill 本质上是一个编排型（orchestrator）skill。看起来它在调用其他这些 skill、生成文档，一步步走完这 11 步流程。

[11:39] Oji

Yes, it it it's So, if you look at these skills here, they're all really dense. They have built-in product frameworks that uh some of them we've written from our book, Building Rocket Ships. Some of them are um well-known product frameworks uh that we've included in part of this. Um, building rocket ship is actually pretty comprehensive and it's a very good resource. But what it does is it combines some of these things. The most important one it combines actually the sharp problem test. So it's doing an internal sharp problem test and it's orchestrating a few more skills within that in terms of new product scaffolding. Well, the thing that's special about this one is that by the time it's done, you will have a full repo. So you can see that it's created a prototypes folder. It's created your package of thinking. It's also created your first milestone folder for you to get started. By the time it's done, architecture is settled. Continuous integration is settled. And if you give it its GitHub repo, it can start the project and push the make the first push into GitHub for you. Why is this important? When you are a PM and when you are a vibe coder, the things that are difficult for you are all the things, right? It's doing the right kind of market research, figure out if it's the right thing to work on, but then immediately the where you hand off to developers is also a problem. is like what's the architecture? What's the uh you know I've seen a lot of vipers who don't have continuous integration. They don't know how to test the code because they don't know how to do that. So this sets it all for you so that you have the bones of a really solid product and a really solid code base to start to work on. I think that's the power of the scaffolding.

对，就是这样。你看这里这些 skill，每一个都非常密实。它们内置了产品框架，有些是我们从自己那本书《Building Rocket Ships》里写出来的，有些是业内公认的产品框架，我们也一并收了进来。《Building Rocket Ships》本身其实相当全面，是个很好的资料源。而这个 skill 做的事，是把这些东西组合起来。它组合进来的最重要的一个，其实就是 sharp problem test。所以它内部会跑一遍 sharp problem test，并且围绕新项目 scaffolding 这件事，再编排调用好几个别的 skill。这个 skill 特别的地方在于：等它跑完，你手里会有一个完整的 repo。你能看到它建了一个 prototypes 文件夹，建好了你那一整包思考成果，还替你建好了第一个 milestone 文件夹，让你可以直接上手。等它跑完，架构定了，持续集成也定了。如果你把 GitHub repo 给它，它可以直接把项目开起来，替你完成第一次 push。这为什么重要？当你是 PM、当你是 vibe coder，难的地方是所有地方：怎么做对的市场调研、怎么判断这到底是不是该做的事；紧接着，交接给开发的那一环也是个问题——架构是什么？我见过很多 vibe coder 根本没有持续集成，也不知道怎么测代码，因为他们压根不会。而这套东西把这些全给你定好了，让你一开始就有一副非常扎实的产品骨架和一个非常扎实的代码库。我觉得这就是 scaffolding 的威力。

[13:43] Aakash

Here's a quick word from our sponsors. Here's what I did last week that changed how I think about AI coding tools. I built a full internal dashboard in bolt.new. Took me about 2 hours. Looked good, worked well, had all the functionality the team needed. Then I did something you can't do in most AI coding tools. I sent the actual source code to my engineer. In base 44 in lovable, in most of the tools, the code lives inside the platform. Your engineer can't pull it. They look at your prototype, say cool and rebuild the whole thing from scratch. Bolt at new is different. The code is pullable. real production-grade front-end code. My engineer opened it, reviewed it, and started building on top of what I'd already made. No rebuild, no start from scratch. The prototype became the product. I use Bolt.new every single day. I built my entire paid cohort [music] platform on it. 75 product managers log into something I made in Bolt. Try it yourself. Bolt. Akash. That's bolt.new/ aos. Here's the problem with most AI tools. They just sit there waiting for you to prompt them. You [music] do all the remembering, they do the typing. Ari from Ariso flip them. It connects to my email, calendar, Slack, and it works the other way around. It comes to me. Monday morning, there's a briefing waiting. Before a call, it reminds me what we covered last time and what I [music] promised to send. When a follow-up slips, it nudges me before the other person has to. Every commitment, every thread, every [music] contact tracked and organized without lifting a finger. And because it has continuous memory, it [music] actually gets smarter the longer I use it. It's not a chat window. It's one AI partner that knows my whole working life and it keeps me prepared for all of it. If you're tired of being your own chief of staff, try [music] Ario.ai/ aos. That's a r iO.ai/ a kh.

插播一段赞助内容。说说我上周做的一件事，它改变了我对 AI 编程工具的看法。我用 Bolt.new 做了一个完整的内部 dashboard，大概花了 2 小时。样子好看，跑得也顺，团队要的功能全有。然后我做了一件在大多数 AI 编程工具里做不到的事：我把真正的源码发给了我的工程师。在 Base44、在 Lovable、在大多数这类工具里，代码是锁在平台里面的，工程师拉不出来。他们看看你的原型，说一句「不错」，然后整个从零重写。Bolt.new 不一样，代码是能拉出来的，是真正生产级的前端代码。我的工程师直接打开、review，然后在我已经做好的东西上继续往上建。不用重写，不用从零开始，原型直接变成了产品。我每天都在用 Bolt.new，我整个付费训练营平台就是用它搭的，75 个产品经理每天登录的就是我做出来的这个东西。你自己试试：bolt.new/aakash。再说说大多数 AI 工具的毛病——它们就干坐在那儿，等着你发指令。记东西的活全是你干，它们只负责打字。Ario 把这个关系翻了过来。它接上我的邮箱、日历和 Slack，反过来主动来找我。周一早上，简报已经等在那儿了。开会之前，它提醒我上次聊了什么、我答应过要发什么。有件跟进的事快掉了，它会赶在对方来催之前先推我一下。每一个承诺、每一条线索、每一个联系人，全被追踪和整理好，我一根手指都不用抬。而且因为它有持续记忆，我用得越久，它其实越聪明。它不是一个聊天窗口，而是一个了解我整个工作生活的 AI 搭档，随时让我把所有事都准备好。如果你厌倦了给自己当参谋长，去试试 ario.ai/aakash。

[15:25] Aakash

Can we scroll through that file a little bit just to see how it's put together?

我们能把那个文件往下翻一翻吗？看看它到底是怎么组织的。

[15:30] Oji

So which particular file?

具体是哪个文件？

[15:32] Aakash

The scaffolding skill.

那个 scaffolding skill。

[15:34] Oji

Oh yeah. So the skill. So yeah, if you go to new project scaffolding, you have the main skill here and then

哦，那个 skill 啊。行，你打开 new project scaffolding，主 skill 在这儿，然后我们往下翻。

[15:43] Aakash

we scroll through that. How is it? Uh okay, so it's got each step defined and if we keep going like for instance the sharp test is it going to go crawl call that skill?

往下翻一翻。怎么样？嗯，好——它把每一步都定义好了。再往下，比如 sharp problem test，它是会跑去调用那个 skill 吗？

[15:59] Oji

Yeah. So for example, the workflow is to gather context. It will ask you questions if you don't have it. It'll find a project type. Is this an iOS app? Android. Because of my prompt, it was pretty clear what it was. Uh, it will look for reference projects around the internet, some of the best practices, and it will try to understand how to build this. Uh, it will run a viability gate. This is a compressed version of the SH problem state where it will go through run clarity, target user, competitive landscape, differentiation. the the the SH problem test is even more specific. The thing that differentiates that is that it tries really hard to focus on three times value that you create. That's not in here. This sort of looks at the market space and see is there a lane for you to be in this in this lane and the sharp problem test will do something even deeper than that. Um whether you will make money that's really what uh sharp problem space is. This is about is there a lane for you? And if there are three dimensions that are weak, it will just recommend that you don't do this at all. This is very important. Uh LLMs very rarely tell you no. And this skill will tell you absolutely no, don't do this. Um and I'll demonstrate that in a second. I'll show you. I'll tell it to scaffold a new thing that uh I think if we look at it, it might look on a boundary, but it will say it will probably say we can't do this. And then market research is huge.

对。比如说，整条 workflow 第一步是收集上下文——你信息给得不够，它就会反过来问你。然后判断项目类型：这是 iOS app 还是 Android？因为我 prompt 写得比较清楚，它一眼就知道了。接着它会去网上找参考项目、找最佳实践，试着搞明白这东西该怎么建。然后它会跑一道 viability gate。这是 sharp problem test 的压缩版，会依次过一遍：问题是否清晰、目标用户、竞争格局、差异化。sharp problem test 比这更细，它最大的不同是死磕你到底有没有创造出 3 倍的价值——那部分不在这里。这里更多是看市场空间：这条赛道里到底还有没有你的位置。而 sharp problem test 会挖得更深——你究竟能不能赚到钱，那才是 sharp problem space 真正关心的。这一层只管有没有你的一条道。而且只要有三个维度是弱的，它会直接建议你：这事儿别做了。这一点非常重要。LLM 是极少会对你说「不」的。而这个 skill 会斩钉截铁地告诉你：绝对不行，别做。等一下我就演示给你看。我会让它 scaffold 一个新东西，那东西我觉得看着是在及格线边上，但它大概率会说这个我们做不了。再往后，市场调研这块分量很重。

[17:28] Oji

Um, you know, we tell people like, you know, market research is weird. Like, you probably shouldn't do market research with only LLMs. Um, but you can do a lot 60% 70% with LM before you go find real sources. And, you know, I want to step back and also tell people, um, you know, the danger in in this kind of workflow is that you take it as gospel, right? You really do need to look through all the output, make sure that it makes sense for you. Uh we do this all the time uh because you know I think it's hard to fully trust 100% LLMs. So if you go here for okay we're still in skills so let's finish that and we can see so it will generate a perfect claude MD for you for this project. So if you look you'll see people spend a lot of time thinking okay how do I author my claude MD what's the thing in the front matter in the rest of it. This says for this project here is a claude MD that works best for you. uh it'll create a cloud MD that does the project well, hunts bugs, follows the patterns of the folders, follows the patterns of the prototypes, and it's perfect for this project. Um, and then it will create a folder structure. So, it will say here's where you need to put the documents, here's where all the milestone things. And by the way, all this is in the cloud MD.

我们一直跟大家说，市场调研这事儿挺微妙的——你大概不该只靠 LLM 来做市场调研。但在你去找真实信源之前，用 LLM 先干掉六七成是完全可以的。另外我想退一步提醒一句：这类 workflow 最大的危险，就是你把它的输出当圣旨。你真的得把所有产出都过一遍，确认它对你是成立的。我们自己也天天这么干，因为我觉得百分之百信任 LLM 还是很难的。好，我们还在 skills 这块，先把这段讲完——它会给这个项目生成一份完美的 CLAUDE.md。你看，很多人花大把时间琢磨：我这份 CLAUDE.md 该怎么写？front matter 里放什么？剩下的又写什么？而这个 skill 直接说：针对这个项目，这就是最适合你的一份 CLAUDE.md。它生成的 CLAUDE.md 能把这个项目做好、会主动抓 bug、遵循目录的规范、遵循 prototype 的规范，完全为这个项目量身定做。然后它会把目录结构建出来，告诉你文档该放哪儿、里程碑相关的东西放哪儿。顺便说一句，这些全都写进 CLAUDE.md 里了。

[18:49] Oji

So, this is basically a self-organizing repo already. uh if you're going to use it here where the prototypes go and for each folder it just tells you it tells cloud and it tells you exactly what goes into it. It creates a test folder for you and then crucially it sets up your continuous integration your quality system. So every time you check in all your code is tested and it sets up security. It starts to find things that you shouldn't your secrets are particularly get ignored. So it knows how to do that so you don't have leakage when you push things to um GitHub and it starts quality right it starts keep all your documents you know how to cat you know do bug classification how to learn it starts to catalog over time bugs that are made and how to learn from those bugs so there's a lot that goes into this and then some of the harder things that it offloads are the CI templates So it's claude MD pattern. So it gets the perfect claw MD. So these are all called within that skill. And so if you go back here, uh what you have is essentially we started out with a very clean empty uh folder, but this is all you have now, right? And so it gets ready for you to commit your first is commit your first checkin on this project. Okay. So, why don't we try this? Why don't we try another project? Okay. Um, that may or may not, we won't say anything. May or may not pass the viability gate. Should we do that?

所以这基本上已经是一个自组织的 repo 了。prototype 放哪儿它都给你定好；每个文件夹它都写明白——既是说给 Claude 听，也是说给你听——里面到底该放什么。它给你建好 test 目录，然后关键的是，它把持续集成、把你的质量体系也一并搭起来了：你每次提交代码，所有代码都会跑一遍测试。它还配好了安全这块——它会去找那些你不该提交的东西，尤其会把 secrets 加进 gitignore。它知道该怎么处理，所以你 push 到 GitHub 的时候不会泄密。质量也是从一开始就有：所有文档留档，怎么做 bug 分类、怎么从中学习，它会随着时间把犯过的 bug 编成目录，再从这些 bug 里学。所以这里面的东西相当多。再难一点的活儿它会外包出去，比如 CI 模板，比如 CLAUDE.md 的写法模式，这样你才能拿到一份完美的 CLAUDE.md。这些全都是那个 skill 内部调用的。回到这边看，我们一开始是一个干干净净的空文件夹，而现在你手上已经是这么一堆东西了。到这一步，你就可以给这个项目提交第一次 commit 了。好，那我们要不要试一把？要不要再来一个项目？这个项目嘛……我们先不说，可能过得了 viability gate，也可能过不了。要试吗？

[20:29] Aakash

Let's do it.

来吧。

[20:31] Oji

Okay. Okay. So, let's do this. Run the scaffolding. scaffolding run the scaffolding skill on a new problem. Use a new folder for it in this one in this ripple. Okay, let's do that. And so I am going to try to solve a standard problem. I'm going to try to ask it build something for a project called Standup Zero that sort of uses Slack comments and creates a daily standup digest and ask it to run a viability gate first. Let's see what happens.

好，那我们来。跑 scaffolding——「在这个 repo 里，对一个新问题跑 scaffolding skill，给它单开一个新文件夹」。行，就这么来。我要试一个很常规的问题：我让它做一个叫 Standup Zero 的项目，大概是抓 Slack 里的评论、生成每日站会摘要，并且要求它先跑一遍 viability gate。看看会发生什么。

[21:29]

So while this is running, how much of this thinking like viability gates and SH problem tests should we really be outsourcing to LLM? Is LM basically the first draft just to get some thinking going and then we improve and react? Yeah, I I the the very specific thing that we are trying to solve is that if you start in chat and start with, hey, I have an idea, you know, unless you prompt really well and unless you sort of sold on adversarial prompting or even some people go as far as they will check with multiple different kinds of LLMs, what you're going going to see is it will tell you that you have a good idea. [snorts] The thing about this is that it's built in. It has an you know it has a built-in sort of responsibility to tell you whether you're smoking crack or not. And I don't I mean that jokingly tell you whether your idea makes any sense. And it doesn't just ground it in the model sense. it grounds it in a really clear framework that has evals that goes through very systematically. Um so I think that's the main difference. Now should you trust it completely? I've already said you shouldn't. Um it produces real artifacts that you should go check. So it'll create a viability gate document. You should look at that viability gate document and see if you agree with it. See if you can do extended research with it. But it's your first draft. It's your first thing out of the gate that has the ability to say yes or no and here's where you need to go.

Aakash：趁它在跑我问一句——viability gate、sharp problem test 这类思考，我们到底应该把多少外包给 LLM？LLM 是不是基本上只负责出个初稿、把思路带起来，然后我们再改、再回应？ Oji：对。我们想解决的其实是一个非常具体的问题：如果你是在聊天框里开局，上来说「我有个想法」，那么除非你 prompt 功夫特别好，除非你懂对抗式提问（adversarial prompting），甚至有些人会拿好几种不同的 LLM 交叉验证——否则你听到的结论一定是：你这想法不错。而这套东西的区别在于，它内建了一份责任，就是告诉你你是不是在瞎扯。我这么说是开玩笑——就是告诉你，你这个想法到底成不成立。而且它不是只靠模型的直觉来判断，它落在一套非常清晰、带 evals、按部就班往下走的框架上。我觉得这是最主要的差别。那你该不该完全信它？我刚才已经说了不该。它会产出真实的交付物，你得去检查。它会生成一份 viability gate 文档，你得去看那份文档，看你同不同意，看能不能基于它再做更深的调研。但它是你的第一稿，是你一出门就能给你一个「行还是不行」、并告诉你该往哪儿走的东西。

[23:16] Oji

Okay. So it doesn't have enough information. So it's asking me questions. What should standard be zero be built as? Let's call it uh slack typescript and slack. How far should this scaffold go? Let's go to the gate first and then pause because if the gate says no, then what what are we doing? Submit our answers. All right, let's see what it says. So, it's gone off and done some crawling, some competit competitive research as we saw, and it's trying to figure out if this is worth it. Um, I've read some of the the I've read some of the stuff it comes up with for the market research and it's really good. It will do a real basically as high quality as perplexity will do to go look for whether there's a space for you. Okay.

好，它信息不够，所以在反问我。Standup Zero 要用什么来做？那就写 TypeScript 加 Slack 吧。这次 scaffold 要走到哪一步？先走到 gate 就停——因为如果 gate 说不行，那我们还折腾什么。提交答案。行，看看它怎么说。它已经跑出去爬了一圈，做了我们刚才看到的那种竞品调研，正在判断这事到底值不值得做。它做出来的市场调研我读过一些，质量真的很好。它基本能做到 Perplexity 那个水准，去帮你查这个市场里还有没有你的位置。好。

[24:22]

And does this apply only to new products or what about like features for existing products? That's a really really good uh question. So uh because we realize you know a lot of vibe coders will say I could never build before I need to build. So this is perfect for that. Um when I started the first demo of code memo I used these skills to do it but really most PMs don't start brand new projects. So we have features for them. So for example, this is called the vet a feature. I have a feature idea. Should I even bother building it? I have opportunity cost of other feature ideas. And so this particular skill will take a feature idea and tear it apart. Look for anti- patterns, confidence. So it's like a scaffolding skill, but very specific for a feature within a product. Um, and I think this one is actually even more interesting because this one actually focuses very hard, not just on the lane, but whether it's worth building at all amongst other features you have. So the sharp problem test is baked in really hard into this.

Aakash：这套只适用于全新产品吗？那已有产品上的新功能呢？ Oji：这个问题问得太好了。我们意识到，很多 vibe coder 会说「我以前根本做不出东西，我需要能把东西做出来」，那这套对他们是完美的。我当初做 Code Memo 的第一版 demo，用的就是这些 skill。但说实话，大多数 PM 并不会从零起新项目，所以我们也给他们准备了针对 feature 的那一套。比如这个叫 Vet a Feature：我有一个功能想法，我到底该不该费劲去做它？我还有别的功能想法，做这个的机会成本是什么？这个 skill 会把一个功能想法拆开来撕一遍——找反模式、看置信度。所以它像 scaffolding skill，但专门针对已有产品里的某个 feature。我觉得这个其实更有意思，因为它不光看有没有赛道，还死磕一件事：在你手上这一堆功能里，这个到底值不值得做。所以 sharp problem test 在这里嵌得特别深。

[25:42] Aakash

So the key lesson here for any PM is like create a skill like veta feature. You can grab OG's which we're going to link down below in the description. But fundamentally, you need to work with LLMs to actually figure out is this a problem worth solving? Is this the right problem space? And before we used to do that all by ourselves or skip that step entirely. And this is really helping us have a thought partner make sure we actually do it every single time.

所以对任何 PM 来说，这里最关键的启发就是：去做一个像 Vet a Feature 这样的 skill。你可以直接拿 Oji 的——我们会放在下面简介的链接里。但根本上，你需要跟 LLM 一起去搞清楚：这个问题值不值得解决？这是不是对的问题空间？以前我们要么全靠自己闷头想，要么干脆把这一步跳过去。而现在这等于给了我们一个思考伙伴，保证我们每一次都真的把这一步做了。

[26:09] Oji

That is correct. That is correct. We we want you to think about the full stack of the problem. again business product and code together. Um you know when you work in triads or quads or whatever that is because of the people around you you are if you're lucky thinking about all those things together and you're like is this worth solving as a PM and the the the developer is like oh here's how we're going to do it and the product marketing is like here's how we tell the story I'm going to go off and do research. What we're doing is compressing all that into one really smart orchestrator and builder can get started really fast. It doesn't mean you don't need specialists, by the way. Big companies always need specialization. That's a given. But in small startups, in smaller companies, and even as you scale up to big companies, the ability for people to one person holding a critical skill, say product, to really have agents that help them with the first draft of these things is incredibly powerful.

没错，就是这样。我们希望你把问题的整个 stack 一起想——商业、产品、代码放在一起想。你在 triad、quad 这种小组里干活的时候，因为身边有这些人，运气好的话你是把这几件事一起在想的：你作为 PM 在想这事值不值得解决，开发在想我们该怎么做出来，产品市场在想这个故事怎么讲、我去做调研。我们现在做的，就是把这一整套压缩进一个非常聪明的编排者兼建造者里，让你能极快地起步。顺带说一句，这不代表你不需要专才。大公司永远需要专业分工，这是必然的。但在小创业公司、在规模没那么大的团队，甚至你一路长成大公司之后，让一个握着关键技能（比如产品）的人手上有一批 agent 帮他把这些东西的初稿做出来，威力是大得惊人的。

[27:13] Aakash

Quick thought experiment for you. Is there anything in this video you should be trying on your own? If there is, try it. Take a screenshot, post it on LinkedIn X, and tag me. I'd love to see what you're learning. Now, a quick word from our sponsors before we get into the back half of the pod. I used to think I had a retention problem. Turns out I had a messaging problem. I was sending the same onboarding emails to every [music] new user, whether they activated on day one or never logged in again. I had no idea who was slipping or why. Customer.io changed that. [music] Every message I send is now based on what users actually do in the product. Someone hits a key activation moment, they get nudged to the next one. Someone goes quiet, [music] they get a different path entirely. Their AI agent makes it fast. I describe the campaign I want and it builds the full journey for me. Triggers, timing, [music] copy, even branching logic. And when I want to know how something is performing, I just ask the agent directly and it tells me what to do next. They also have an MCP server [music] which means AI tools like Claude can see directly what's happening in your customer.io workspace. Your segments, your customer data, [music] your attribution, all of it. So instead of explaining your business context every time you need help, Claude already knows it. Notion used customer.io [music] IO to personalize their onboarding and hit nearly 50% open rate, improved conversion by 6 to 7% with localized [music] campaigns, and pushed open rates up another 20% through AB testing. The idea is simple. Customer.io helps you deliver more impact from every message you send. If you're a PMR founder and your onboarding is still one-sizefits-all, try Customer.io at customer.io. Are you looking to land your next product management job? I am accepting a group of just 30 product managers into a 12-week cohort led by me where every Monday for 90 minutes I help

给你出个小思想实验：这期视频里，有没有什么是你该自己动手试一下的？如果有，就去试。试完截个图发到 LinkedIn 或 X 上，@ 我一下，我很想看看你学到了什么。下面先插播一段赞助商内容，然后我们进入这期播客的后半段。我以前一直以为自己有个留存问题，结果发现我有的其实是一个「信息传达」问题。我给所有新用户发的是同一套 onboarding 邮件——不管他第一天就激活了，还是从此再没登录过。我压根不知道谁在流失、为什么流失。Customer.io 改变了这一点。我现在发出去的每一条消息，都基于用户在产品里的真实行为。有人触达了某个关键激活时刻，就会被推向下一个；有人沉默了，就走一条完全不同的路径。他们的 AI agent 让这件事变得很快：我描述我想要的 campaign，它就把整条 journey 给我搭出来——触发条件、时机、文案，连分支逻辑都有。想知道某个东西跑得怎么样，我直接问 agent，它会告诉我下一步该干什么。他们还有一个 MCP server，意味着 Claude 这类 AI 工具可以直接看到你 Customer.io workspace 里的情况：你的分群、你的用户数据、你的归因，全都能看。所以你不用每次求助都重新解释一遍业务背景，Claude 已经知道了。Notion 就用 Customer.io 做 onboarding 个性化，打开率接近 50%，靠本地化 campaign 把转化提升了 6% 到 7%，再用 A/B 测试把打开率又往上推了 20%。逻辑很简单：Customer.io 帮你让发出去的每一条消息都更有效。如果你是 PM 或创始人，而你的 onboarding 还是一刀切，去 customer.io 试试。另外，你在找下一份产品经理的工作吗？我要收一批人，只收 30 位产品经理，进入一个由我亲自带的 12 周训练营。每周一 90 分钟，我会带你

[28:53] Aakash

you through your job search, creating your candidate market fit, updating your LinkedIn, updating your base resume. You're going to get personalized feedback and one-on-one mentorship sessions with my co-teers, Ankit Fermani, who is an AIPM at Atlassian and was a group product manager at Meta, Prasad Ready, who is a CPO and has been in product for over 26 years, as well as my other live instructor, Bar Jorski, who's going to run another 90-minute session per week where we really help you deliver on all of the deliverables in an actionable way and get you custom resume feedback, custom LinkedIn feedback. This program worked extremely well in cohort number one, which is just finishing up. 40% of the cohort got a job before the cohort even ended. We got jobs at places like OpenAI and Enthropic. So, if you want to get a higherp paying PM job, be sure to check out my landp.com cohort. The next cohort starts in February, runs through the end of April. The next time I'm opening up a cohort is in May. So, if you want coaching from me to land a PM job, this cohort is a no-brainer. It is a premiumbumppriced product. It is more expensive than the average product out there, but the return is huge. Most people who join the cohort see a salary raise anywhere from $10 to $100,000 in the first year. And so the ROI will be there within a year. And we guarantee two plus interviews. So if you don't get two interviews after completing the 12-week program and following all the steps, we will refund the money to you.

走完整个求职过程：找到你的候选人-市场匹配（candidate market fit）、更新 LinkedIn、更新你的基础简历。你会拿到个性化的反馈，还有跟我的联合讲师做一对一辅导——Ankit Fermani，Atlassian 的 AI PM，此前是 Meta 的 group product manager；Prasad Reddy，做过 CPO，在产品这行干了 26 年以上；还有我的另一位直播讲师 Bar Jorski，他每周会再带一场 90 分钟，帮你把所有交付物真正、可落地地做出来，并给你定制化的简历反馈和 LinkedIn 反馈。这个项目第一期效果极好，第一期马上就要结课了：40% 的学员在训练营还没结束就拿到了工作，去了 OpenAI、Anthropic 这样的地方。所以如果你想拿一份薪水更高的 PM 工作，一定去看看我的 PM 求职训练营。下一期 2 月开课，一直上到 4 月底。再往后开放报名要等到 5 月。所以如果你想让我带着你拿下一份 PM 工作，这一期完全不用犹豫。它定价偏高端，比市面上一般的产品贵，但回报非常大。大多数学员在第一年里的涨薪幅度在 1 万到 10 万美元之间，所以一年之内 ROI 就回来了。而且我们保证至少两个面试机会——如果你完整走完这 12 周项目、每一步都照做了还拿不到两个面试，我们把钱退给你。

[30:20] Aakash

So it's a no-brainer. Check it out at landpob.com. And now back into today's episode. Do you know how to take an AI product from idea to development to evaluation to deployment and eventually to scale? That's exactly what product faculty's AIPM certification helps you do. I even took the course myself. You'll learn directly from Rohan Varma, the product lead working on codecs at [music] OpenAI. You'll go deep into AI prototyping, evaluations, agents, AI native workflows, cloud code, open claw, latency, cost, guardrails, rag, routing, fine-tuning, and production systems. You'll even build your own AI product as your capstone with unlimited one-on-one support. So, if you want to stop just learning AI and actually build AI products that work, join product faculty's AIPM certification on Maven. 5,000 plus students have graduated and they have 1,000 plus reviews. Use code AKOS550 to get $550 off your enrollment.

所以这完全是笔不用想的买卖，去 landpob.com 看看吧。好，我们回到今天这一期。你知道怎么把一个 AI 产品从想法一路做到开发、评估、部署，最后跑到规模化吗？Product Faculty 的 AIPM 认证课教的就是这件事，我自己也上过。你会直接跟 Rohan Varma 学——他是 OpenAI 负责 Codex 的产品负责人。课程会深入讲 AI 原型、evaluation、agent、AI 原生工作流、Claude Code、OpenClaw、延迟、成本、guardrails、RAG、路由、微调，以及生产环境系统。毕业项目还要你自己做出一个 AI 产品，全程有不限次数的一对一辅导。所以，如果你不想只是停留在「学 AI」，而是真想做出跑得起来的 AI 产品，就去 Maven 上报 Product Faculty 的 AIPM 认证。已经有 5000 多名学员毕业，1000 多条评价。用优惠码 AAKASH550 可以减 550 美元。

[31:18] Aakash

So maybe we can turn over to the other one that's ran and you can show us like how you apply your PM judgment on top of what it's written.

要不我们切到另一个已经跑完的那个？你演示一下，你是怎么在它写出来的东西之上叠加你自己的 PM 判断力的。

[31:28] Oji

Yeah. Okay. So that's a good point. So, um I'm back in the first one and let me just make sure that it's uh done. So, I'm going to ask it to continue and share the final results in a table for me for me and so that we can know that it is done. Done. And we're going to get out of So yeah, so here's the result. Okay, so it's done market research. It's decided what's viable quad structure. So the way to think about this now once you're here is to go back into this and start to take a look. Uh I've worked with developers in the past who I had to insist that they actually do code review even if they're working with um even if they're working with uh AI. And so the place you start first is these two documents, right? The other ones are sort of boilerplate. So I'm going to create some space here. So this is the market research document.

嗯，好，这个点问得好。我先回到第一个，确认一下它到底跑完没有。我让它继续，然后把最终结果整理成一张表给我，这样我们心里有数它确实跑完了。好，完事了。我们退出来……对，结果就在这儿。它做完了市场调研，判断了可行性，也把 Claude 的目录结构搭好了。到这一步之后，正确的姿势是回过头去，一份一份看。我以前带过一些开发，得反复跟他们强调：哪怕你是跟 AI 一起写的，code review 该做还是得做。所以第一站就看这两份文档，剩下那些基本都是模板化的东西。我把界面腾点空间出来。这一份是市场调研文档。

[32:46] Aakash

And then I think if we doubleclick, if you like two-finger click on the market research,

然后我觉得你可以双击一下——或者用双指点一下那个 market research 文件。

[32:51] Aakash

Mhm. I think you can open as preview so it looks a little better, right?

嗯。可以选 open as preview（预览方式打开），这样看着舒服一点，对吧？

[32:57] Oji

Okay, let's do that.

好，那试试。

[32:58] Aakash

There we go. Perfect.

这就对了，完美。

[32:59] Oji

Okay. Well, you know a VS code trick that I don't. So that's that's good on you cache. [laughter] All right. So, uh, market overview. It sits at intersection of AI code. So, look, I don't want to read through all this. What the first thing is to take this for yourself and for PMs, PM leads in your space and see if it covers all the ground that you need. Uh the things to look out for is does it mention the kinds of tools you expect it to mention? Well, these are a really good set for this kind of tool because these are your competitors. Uh does it mention in each of these categories all the things that you wanted to mention? uh you take a sense check of these things like if it mentions tools that are old, tools that are not functional, tools that are sort of decorate, you shouldn't you shouldn't trust it. If it mentions tools that are current uh adjacent things that you were thinking about as you were thinking about the problem, but they didn't actually solve the problem for you, then you are on the right track. Um, take a look at this differentiation map for example. Does it seem right? No. Yes. Where does code memo sit? Does it seem right for you? And so this is your first scan for credibility. Imagine that you are doing a a PRD review in a room full of PMs. This is what you essentially should be doing with this one. [snorts] And then it actually writes a PRD for you. And so this is the real PRD review.

行。你还知道一个我不会的 VS Code 小技巧，可以啊 Aakash。（笑）好，先看市场概览。它说这个产品处在 AI 和代码的交叉点上。我不打算逐字念完。第一件事，是把这份东西拿给你自己、拿给你所在领域的 PM 和 PM leader 看，检查它覆盖的面够不够。要盯什么？盯它有没有提到你预期它该提到的那些工具。这里列的这一组就挺靠谱，因为这些正是你的竞品。再看每个分类底下，你想到的东西它是不是都写进去了。你要做一次常识校验：如果它提的是过时的工具、根本跑不起来的工具，或者只是充数的摆设，那你就不该信它；反过来，如果它提的都是当下真在用的工具，还包括你琢磨这个问题时脑子里冒出来过、但其实并没有真正解决你问题的那些相邻方案，那说明方向是对的。再比如看这张差异化地图，它排得合理吗？CodeMemo 应该落在哪个位置？符合你的判断吗？这就是你的第一遍可信度扫描。你可以想象自己正站在一屋子 PM 面前做 PRD 评审——对着这份东西你要做的就是这件事。然后它还真给你写了一份 PRD，那才是真正要评审的东西。

[34:37] Oji

So is the stuff in the market consistent with what it's writing for you. So let's scan this very quickly. Code is now produced faster for anyone than anyone can read it. AI codegen has split authorship from understanding. The person who wrote an app via prompts often cannot evaluate whether it's safe, correct, or production ready. This bites two groups hard. Solo founders ship AI generally they cannot audit. Corporate developers onto an unfamiliar code base. So it came up with all this by itself. And if you stare at this and say, "Shit, I would invest in that. I would spend time on that. It nailed exactly what I want." Or you should be asking yourself, "What did it miss? What persona did it miss?" For example, what I notice here is that it says the primary is a nontechnical founder. Makes sense. Small company. Sure. Um, it's probably not a big company founder though, right? They have more resources, more tools. So, I'm glad that it said non-technical founder, but it didn't qualify it by the size of the company. For example, um, it sort of doesn't think the corporate developer is the strongest thing, right? Antony who needs evolving story, which is simply out of scope to protect focus.

所以关键是看：市场调研里的结论，和它给你写出来的 PRD 对不对得上。我们快速扫一眼。它写道：今天代码被生产出来的速度，已经超过任何人能读懂的速度；AI 代码生成把「写代码」和「看懂代码」这两件事拆开了；靠 prompt 攒出一个 app 的人，往往判断不了这段代码到底安不安全、对不对、能不能上生产。这件事扎得最狠的是两类人：一是独立创始人，他们发出去的是自己没法审计的 AI 生成代码；二是刚接手一个陌生代码库的企业开发者。这些全是它自己想出来的。你盯着这段看，如果反应是「靠，这个我愿意投钱、我愿意花时间，它把我想说的全说中了」，那就对了；否则你就该反过来问：它漏了什么？漏了哪类人群？比如我注意到，它写主要用户是非技术背景的创始人——合理；小公司——也没问题。但大公司的创始人大概率不是目标吧？人家资源更多、工具也更多。所以它写「非技术创始人」我很满意，但它没有再按公司规模限定一层。另外它其实不太看好企业开发者这条线，把这类人列成了 anti-persona，理由是这类需求需要一套持续演进的方案，为了保住聚焦，直接判成 out of scope。

[36:02] Oji

That's makes sense, I guess. But I I read the output and it was it mentioned that it's much lower value proposition for the corporate developer because they have so many more tools to do these things. And so the vibe coder is a real target for this value proposition. Paste a GitHub URL upload a vioded folder and in under a minute get a plain English verdict on whether the code is production safe. So this starts to tell us that the value loop is really good is really powerful. Okay. Right. And so I would scan this and see is this the kind of highle product brief I would write? What's missing? What's missing if I wanted to present it to other people? And that's how to evaluate this.

这么说也讲得通吧。不过我把输出读下来，它的意思是：对企业开发者来说这个价值主张要弱得多，因为人家手上能干这些事的工具多得是。所以真正的目标用户是 vibe coder。贴一个 GitHub URL，或者上传一个 vibe coding 出来的文件夹，一分钟之内就给你一句大白话的结论：这段代码能不能安全上生产。看到这儿你就知道，这个价值闭环很扎实、很有劲。对吧。所以我会这样扫一遍，然后问自己：这是不是我自己会写的那种高层产品简报？缺了什么？如果我要拿去讲给别人听，还缺什么？评估的方法就是这样。

[36:48] Aakash

Cool. Let's go take a look at the other agent and see if what it did uh with our other idea.

不错。我们去看看另一个 agent，看它把我们那个想法做成了什么样。

[36:54] Oji

Okay, let's do that. So let's step back here. It looks like it is still crunching through the viability gate. So, let's go to that folder. And you can see it's called standup zero here. And it is crunch through the viability accept accept uh viability assessment. And so scorecard so it's weak, right? So this is the real sort of alpha from using a skill like this. It's going to help you separate out which features or in this case we showed new products that you should be building and shouldn't be building.

好，那就看。我们先退回来。看起来它还在跑 viability gate。我们进那个文件夹——你看，这个叫 Standup Zero。它已经把可行性评估跑完了。评分卡出来了：偏弱。这才是用这类 skill 真正的超额价值——它能帮你把该做的和不该做的功能分开，在这个例子里分的是该不该做这个新产品。

[37:45] Oji

Yeah. So here it says there were three moderate scores. So it's saying I'm not sure about this basically right. um you should proceed with eyes open and then it goes into the details right problem clarity and urgency is moderate um it's just workflow convenience it's not very deep uh target user definition isn't like isn't very strong for what I pasted in uh the competitive landscape is very strong is actually a bad thing there's a lot of competition here and so it's calling out all these things for you uh differentiation isn't very strong. So all all of this is basically pointing out to like it's going to be very hard for you to enter this market. It's not defensible for you. Um everyone can build exactly what you're building. So if you really want to continue this, you should maybe pause and think about it. Basically now I don't know about you, but I think that's worth a lot from start, right? This is you thinking hard about where you're going before you start thinking I can write code. Because you know what we see today is people just build. People build and then there's no customer for it. There is nobody for it. In fact, it's super puzzling right now because it feels like GitHub is the only place that people can express themselves when they have cloud code. They just build anything, put it on GitHub, and just languish with zero stars. If that makes any sense. Now, if you are someone who wants to make an impact in the world, this kind of thing will help you conserve your time, attention, and focus on all the right things.

对。这里说有三项拿了中等分。翻译过来就是：这事我不太确定，你要做可以，但得睁着眼睛做。然后它把细节铺开：问题的清晰度和紧迫性，中等——这只是流程上图个方便，不是什么深痛点；目标用户的定义，就我贴进去的那点内容来说，不够扎实；竞争格局，非常强——但在这儿「强」是坏事，说明这条赛道竞争太激烈了。它把这些一条条给你点出来。差异化也不够强。所有这些加起来其实就一句话：你要杀进这个市场会非常难，而且守不住，谁都能造出一模一样的东西。所以你要是真想接着做，也许该先停下来想想。你怎么看我不知道，但我觉得这份东西在起步阶段就值很多钱——这是在你还没开始想「我可以写代码了」之前，先认真想清楚自己要去哪儿。因为今天我们看到的现象是：大家上来就造。造完了没客户，压根没人要。说实话现在这事挺让人费解的，感觉有了 Claude Code 之后，GitHub 成了大家唯一能表达自己的地方——什么都造，往 GitHub 上一扔，然后零 star 烂在那儿。你懂我意思吧。如果你是那种真想在世界上做出点影响的人，这类东西能帮你把时间、注意力和精力都省下来，花在真正对的地方。

[39:32] Aakash

So, where do we go from here? We promised that we'd help people become a builder PM. What's the next step once you've figured out the right idea? In this case, we're knocking out standup zero, but we're in favor of code memo.

那接下来往哪儿走？我们答应过要帮大家变成会造东西的 builder PM。选对想法之后，下一步是什么？在这个例子里，我们把 Standup Zero 毙掉了，选了 CodeMemo。

[39:45] Oji

Yes. So, let's go back to code memo, right? And I think we get to tell it to continue. All right. Fantastic. So, the next thing I would do is prototypes, right? And this is a fun one. What should the code memo look like, right? I want to establish the basic interaction model um that delivers value. And you first ask questions about user interacting with code memo. And then we will generate a set of X prototypes to evaluate what direction to take this in. Now this is very compressed right like you don't necessarily um start with prototypes but you do start with interaction like and by that I mean specifically does this have user experience does this is just is this just a chat interface talking to an agent like what does it look like how do you instantiate it and so uh I'm going to ask it to use its survey skill to help me make the choices So now that it says go, what I'm trying to do next is really to um start to build a baseline of the code, start to build a baseline of a product. Um if you look at these skills, there a few directions you can take this in. For example, the interesting one that I would do right after this is the customer discovery week. Okay, I'm going to ask it to use this skill to tell me how to get to better confidence and to answer the unknowns. Okay, so while it's doing that, [clears throat] I'll prep that question. So, I need you to develop my customer discovery plan by running the customer discovery week skills. So, on one hand, I'm trying to generate ideas of what the interaction is. On the other hand, I need to ask a bunch of questions. So let's tackle the questions about the interaction model first. So paste a good hub URL.

对。那我们回到 CodeMemo。我想我们可以让它继续跑。好，很好。接下来我会做原型，这一步很好玩。CodeMemo 该长什么样？我想先把那个「能交付价值的基本交互模型」定下来。所以我先让它就「用户怎么跟 CodeMemo 交互」问我一批问题，然后我们再生成一组原型，用来判断这东西该往哪个方向走。这里我压缩得很厉害——正常情况下你不一定一上来就做原型，但你一定要先定交互。我说的交互具体是指：这东西到底有没有一套用户体验？还是说它就是一个跟 agent 对话的聊天界面？它长什么样？你怎么把它启动起来？所以我打算让它调用 survey skill 来帮我做这些选择。它跑起来之后，我下一步真正要做的，是给代码打一个基线、给产品打一个基线。你看这些 skill，从这里可以往好几个方向走。比如我马上就想做的、比较有意思的一个方向，是 customer discovery week（客户访谈周）。好，我让它用这个 skill 告诉我怎么把信心提上去、怎么把未知项一个个答掉。趁它跑着，我先把那条指令备好——我需要你跑 customer discovery week 的 skill，帮我做出客户访谈计划。所以我这边一边在生成「交互到底长什么样」的方案，一边还得去问一大堆问题。我们先处理交互模型这批问题。第一题：贴一个 GitHub URL。

[42:46] Oji

Um so I don't like any of these. So what I'm going to say is I think it's both one and two. So I needed to paste a GitHub URL. But I'm going to connect it to GitHub because I need to analyze the code that I'm writing because I'm a VI coder. How's the verdict delivered on screen? This is a score. single vertic card, narrative scroll, conversational chat. So, I think that it needs to be a dashboard and drill down. Uh, once the verdict is on screen, what should the user be able to do? Depth of engagement. Uh, nope. I needed to I needed to do two and three. So, two and three. Uh, and four. Okay. in the first 60 seconds. This is interesting. It's getting to the aha. What was the core value the user came for? Two understanding. Submit those answers. All right. Uh, I need to develop my customer discovery skill plan or in the customer discovery skills on the product brief and the market research. Uh, do this in parallel while finishing up. It's going to spin up a thread finishing up the interaction model. Okay. So, I'm basically doing two things now. I'm grounding my sense of what interaction is, and I'm going to launch additional research to really figure out if I fully understand the problem.

嗯……这几个选项我都不满意。我要说的是，我觉得一和二都要。我需要能贴 GitHub URL，但我也要能连上 GitHub，因为我是 vibe coder，我得分析我自己正在写的代码。下一题：结论怎么呈现在屏幕上？选项是评分、单张 verdict 卡片、叙事式长滚动、对话式聊天。我觉得应该做成仪表盘加逐层下钻。再下一题：结论出来之后，用户能做什么？也就是参与深度。嗯，不行，我要二和三，二和三，再加四。好。下一题：最初的 60 秒。这题有意思，问的是怎么走到 aha 时刻——用户来这儿要的核心价值到底是什么？选二，理解代码。提交答案。好。接下来我要让它基于产品简报和市场调研去跑 customer discovery skill，做出客户访谈计划。让它并行做，同时把交互模型那边收尾。它会另开一个线程去跑。所以我现在等于同时在做两件事：一是把「交互到底是什么」这件事落地，二是再发起一轮调研，看我是不是真的把问题吃透了。

[44:34]

Now, I would ordinarily do the first one, the second one before the other, but I just wanted to demonstrate the two different ways. You can go straight to build or you can start to talk to customers about it. Okay. So what we're seeing is is summarizing the tension and so it's going to generate three self-contained HTML CSS prototypes. The skill itself tells it to do this prototype before you react. So it's going to go off now and it is going to basically start to write prototypes. So the last set of skills in the diagram we started with was really the business layer like is there a business viability we talked about in terms of new product for a feature there might be some more nuance there. Then this next layer is really that product layer right we're figuring out the prototypes what actually works and we're going to go through and find a prototype that might potentially actually demonstrate a viable solution so that we can test that. Is that right?

正常我会先做第二件、再做第一件，但我想把两条路都演示一遍：你可以直接去 build，也可以先去跟客户聊。好。现在能看到它在总结这里的张力，接下来它会生成三个自包含的 HTML + CSS 原型。是 skill 本身要求它「先做原型，再下判断」的。所以它现在就会跑出去，开始写原型了。那么我们一开始那张图里的最后一层 skill，其实是商业层——商业上到底成不成立；刚才我们是按新产品讲的，如果做的是一个功能，这里会更微妙一些。而接下来这一层就是产品层了，对吧？我们在这一层做原型、看什么真的行得通，然后把那个可能真能证明方案可行的原型挑出来，拿去做验证。是这样吧？

[45:36] Oji

Yes. Can you still hear me?

对。你还能听到我吗？

[45:38] Oji

Yeah. Yeah. So, um that's exactly right. We start out with the business problem. Is there a lane for us? Are we thinking about the customer correctly? There other business problems that you get to later. Pricing, access to market, what's your marketing plan? But what we're doing now is a product is like what's my customer discovery so I can dig deeper into the problem. What does this even look like? Uh just for you know PMS you know we get this like as PMs we own usually usually we own why is this a thing that we spend time on as an organization. What is it which is how does it instantiate? What does it look like? What does it feel like when it's solving the problem? How does it generate delight? And of course, we share this with designers. And so, we're still early in that process. Is this prioritized by my CPO? Assuming you're a VI coder, what won't we know? Um, what does it feel like? What does it look like? Um, these are all the things that you take to developers. But what you have now is that because you're doing it all in one shot inside cloud code, you can compress all of that within one cloud code session. You notice that we are not in cloud code desktop and co-work. You can do all this within one session. So what we have here is it's done some prototypes. What's that shortcut you you liked again? So while this is running in the background, I am going to show you some of the prototypes it's made. I really hope to God because we didn't install a a lot of design skills. Can you see this? Mhm.

嗯，对，完全正确。我们是从商业问题开始的：这里有没有我们的位置？我们对客户的判断对不对？后面还有别的商业问题——定价、能不能进入市场、推广计划是什么。但我们现在做的是产品这一层：我的客户调研该怎么做，才能把问题挖得更深？这东西到底长什么样？对 PM 来说，我们通常拥有的是「为什么这件事值得整个组织花时间」，以及「它到底是什么」——它怎么落地成型？看起来什么样？解决问题的时候用起来是什么感觉？它怎么制造惊喜？当然，这些我们会跟设计师一起做。所以我们还在很早的阶段：这件事我的 CPO 排优先级了吗？就算你是个 vibe coder，有哪些东西是我们还不知道的？它用起来是什么感觉？看起来是什么样？这些都是你要拿去跟开发聊的东西。但现在不一样的是，因为你是在 Claude Code 里一口气做完的，所有这些都能压缩进同一个 Claude Code session。你注意到我们既不在 Claude 桌面端、也不在 Cowork 里——这一切在一个 session 里就能做完。我们这边它已经做出了几个原型。你刚才说你喜欢的那个快捷键是哪个来着？趁它在后台跑，我给你看几个它做出来的原型。真心希望别翻车，因为我们没装多少设计类 skill。你能看到吗？

[47:30] Oji

Yeah. So, it's made a few things for us to take a look at. The way you should think about this is prototypes. You should just get a feel. This isn't real design yet. This is definitely about where do you start? So, here you see it. You paste in the GitHub thing here and it tells you the story of the code very quickly. Um, here more of the same but more data. Uh, logic up front and here's sort of a side panel where it does similar things. So what I usually do when I start to work on these things is I will have in practice I'll have shot CN I'll have UI UX promax skills. I'll have like I have like five or six design skills and I'll tell it to use one design skill to make a prototype. So if I have six design skills, it will make six prototypes. And I'll ask it actually because you can make hundreds and hundreds of prototypes. I'll ask it actually to combine two skills each. So it will make like six more skills but using two design skills at a time to do it. And so I can see what this could look like uh for us. But all this is baked into the skill directly. I'm going to go back to claude code. All right. And share that screen again. and then we'll take a look at the other task that we ask it to do. Okay.

好，它给我们做了几个东西可以看看。你应该把它们当原型看——找个感觉就行，这还不是真正的设计，重点是「从哪儿起步」。你看这个：把 GitHub 链接粘进来，它很快就把这份代码的来龙去脉讲给你听。这边是类似的东西，但数据更多、逻辑放在前面；这边是个侧边栏，做的事情差不多。我实际工作时的习惯是：我会装上 shadcn，装上 UI/UX Pro Max 这类 skill，手上大概有五六个设计 skill，然后让它用其中一个设计 skill 做一版原型。有六个设计 skill，它就出六版原型。我还会让它每次组合两个 skill 来做——因为原型你可以做成百上千个——这样又能多出六版，每版是两个设计 skill 混着来的。这样我就能看到这东西对我们来说可以长成什么样。而这些全都直接烧进 skill 里了。我切回 Claude Code，再共享一次屏幕，我们看看刚才让它做的另一个任务。

[48:57] Oji

So, by the way, it it's asking me about the interaction model I want. We're not going to spend time figuring that out yet. Um, so we decline that. So, we're going to instead say the discovery plan agent finished. So, it created a customer discovery plan for us. So, we're going to go to docs and we're going to go to customer discovery plan and we're going to take a look at it. Okay. And it says the plan itself is complete and executable. Every script, the survey, the synthesis template and the week's time budget are below. What is missing is a one input of skill will not let a plan proceed. A confirmed list of recruitable people with real target cohorts um not adjacent contacts from the founders network. skill treats fewer than five real target contact as own validation signal. It means you may not have access to the market. Again, very important. Uh it says you can't just wing this. You need actual people to talk to. And if you tell me who they are, then I know that you're ready for this. And so what we can see here is the three-step interview process uh that it wants us to do. So let's see. It tells you how to recruit, what the gates are, where to go shop for uh targets, um the hypothesis you want to validate, step one, motivation. So, it walks through. In the book, we talk about this three-step process where the first step is figure out what questions to ask.

顺便说一句，它在问我想要什么样的交互模型。这个我们先不花时间定，直接拒掉。我们改去看 discovery plan 那个 agent，它跑完了——给我们生成了一份客户调研计划。进 docs，打开 customer discovery plan，看一下。它说：计划本身是完整、可执行的，所有访谈脚本、问卷、综合分析模板，还有一周的时间预算都在下面。缺的只有一项输入，而缺了这一项，这个 skill 不会放行：一份真能约到人的名单，必须是真实的目标人群，而不是创始人自己人脉里那些沾边的联系人。如果真实目标联系人少于五个，skill 会把这件事本身当成一个信号——说明你可能根本进不去这个市场。这点非常重要。它等于在说：你不能靠拍脑袋糊弄过去，你得有活人可以聊；你告诉我他们是谁，我才知道你准备好了。然后你能看到它要我们走的三步访谈流程：怎么招募、卡点在哪、去哪儿找目标用户、你想验证的假设、第一步动机……它一步步带着走。书里我们讲过这个三步流程：第一步是先搞清楚该问什么问题。

[50:36] Oji

Open-ended questions, no motivation, no direction, no prototypes whatsoever. So it walks you through that. Tell me about walk me through the last time. So it's walking through people's workflows. And then step two is when you take those things and say okay I now think I have the questions I need to answer. And then it will codify uh those questions that you answer. So the first step you learn what questions to ask. The next one you use those questions to ask and then the third you expand it. So it becomes data. So once you really solidify the questions you're asking, you can start getting sample of the answers. You turn it into a survey so you can quantify the responses that you get. So it builds all this for you ready to go. In this case, it's saying, listen, you you you sound like you're still winging it. You need very specific people. Go find them and I will help you uh validate that you have a real customer discovery plan.

开放式问题，不带动机、不给方向，也完全不拿原型出来。它就这么带着你走：跟我说说、带我过一遍你上一次是怎么做的——也就是把人的工作流走一遍。第二步，是你把这些东西收上来之后说：好，我现在大概知道我需要回答哪些问题了；然后它会把这些问题固化下来。所以第一步是学会该问什么，第二步是拿这些问题去问，第三步是放大——让它变成数据。等你把要问的问题真正定死，就可以开始收集答案的样本，把它变成一份问卷，这样你拿到的回答就能量化。它把这一整套都替你搭好，拿着就能用。而在我们这个例子里，它说的是：听着，你现在听上去还是在拍脑袋，你需要非常具体的人。去把人找出来，我再帮你验证你手上是不是一份真的客户调研计划。

[51:36] Aakash

All right. So, what I'm really hearing from these skills is that they're forcing you to do the PM fundamentals. And what they're doing is they're giving you the first draft of those things. And

好。我从这些 skill 里听到的其实是：它们在逼着你把 PM 的基本功做扎实。而它们给你的，是这些基本功的第一版草稿。

[51:47] Aakash

to be a true builder PM, you need to be able to do this at engineering speed. And that's why we're relying on these skills. We're not just doing it hand as we did step by step in the past.

而要当一个真正的 builder PM，你得能以工程的速度把这些事做完。这也是我们为什么要靠这些 skill——不再像过去那样一步一步手工来。

[51:59] Oji

Correct. So what we there's something we call the three-speed problem which is for the past 40 50 years whatever you want to call it the longest poll has always been uh development uh and it's long in many ways it takes time to actually write quality code it takes time to do quality um it takes time if you're in the cloud to make it robust not fail to do uh devops recover very quickly All of that is being cut down maybe by 10x maybe in 5 years by 20x but really the process of building is not just code. It's like why should we build build it get it to customers. So if this middle speeds up very quickly, you start to have almost an equation imbalance. And the reason it's imbalance is because the why should we build is customer bound. You have to talk to people. You have to talk to the market. And the other side access to the market, getting it into people's hands is also customer bound because you know Antropic has shipped what maybe like 60 features in 30 days. They're becoming recursive. Correct. Um, by the way, that's a whole other thing to talk about, which is how to build feedback into the skill. So, some of these skills build feedback deeply into the product automatically for you. [snorts] Uh, and we're trying to refine some of those things. We haven't shipped that one, but I don't use like we were fussing with slash commands just now. I don't know all the slash commands. So, I can't actually absorb all the features that they make, right? It's li I'm limited by that. And so you see them, I don't know if our our audience knows this, but Claude co-work was just completely redesigned, right? They shipped it and it seemed amazing, but there was low adoption. And so they've just sort of blended it back into chat, right, a bit versus having three tabs, code, co-work, and chat. Uh because you just it didn't stand alone by itself. So people are having to not just ship very quickly but make sure um they have the right thing.

没错。我们管这个叫「三段速度问题」。过去四五十年，不管你怎么算，最长的那根杆一直是开发：写出高质量的代码要时间，把质量做扎实要时间，如果你在云上，要做到健壮、不宕、DevOps 能快速恢复，同样要时间。现在这些正在被砍掉，可能 10 倍，五年内可能 20 倍。但真正的「造东西」这个过程不只是写代码，还包括「我们为什么要造它」，以及把它交到客户手里。所以当中间这一段突然变快，整个等式就失衡了。失衡的原因是：「为什么要造」这一端被客户绑住——你必须跟人聊、跟市场聊；另一端，进入市场、把东西送到用户手上，同样被客户绑住。你看 Anthropic，30 天里大概发了 60 个功能吧，他们已经开始递归了。对。顺便说，这本身是另一个大话题：怎么把反馈机制建进 skill 里。有些 skill 会自动帮你把反馈深深嵌进产品里。我们还在打磨这些，那个还没发。举个例子：我们刚才在折腾斜杠命令，我根本记不住所有斜杠命令，所以我其实吸收不了他们做出来的全部功能，对吧？我被这个卡住了。我不知道观众知不知道，Claude Cowork 刚刚被彻底重做了一遍——发出来的时候看着很惊艳，但采用率很低，所以他们又把它揉回聊天里去了，而不是像原来那样分成 code、Cowork、chat 三个 tab，因为它自己撑不起来。所以现在大家不只要发得快，还得确保自己做的是对的东西。

[54:16] Oji

And by situating this in the same skill set that your developers use that you can learn very trivially. Um you see a massive amount of speed up, right? Because this repo can be your developer's repo as well, right? It's not different, right? If your developer, if your founder and your developer is working on this, they can see all the business skills themselves. They can run it themselves or you can run them, right? So we are seeing people essentially collapse the separation of tasks of PM live in notion. uh designers live in Figma and collapsing it into one GitHub repo with everything and the coders will live here under source test uh all the security stuff playright uh test vi test you live up here but it's joint context in one repo for every member of the team who needs to touch this stuff and everything is always available to them at all times.

而当你把这套东西放进你的开发所用的同一套技能体系里——这套东西你学起来其实非常轻松——你会看到速度提升非常巨大。因为这个 repo 可以同时就是你开发的 repo，没有区别。如果你的创始人和开发都在这上面干活，他们自己就能看到全部这些业务类 skill，他们能自己跑，你也能跑。所以我们看到的是，大家把过去那种分工上的割裂直接压平——PM 活在 Notion 里、设计师活在 Figma 里，现在全部收进一个 GitHub repo：写代码的人在 source、test 底下，安全那一套、Playwright、Vitest 都在那儿；你在上面这一层。但它是同一个 repo 里的共享上下文，团队里每个需要碰这些东西的人，随时都能拿到全部内容。

[55:31] Aakash

Amazing. If people want to go grab these skills, can you show us where they can go grab them?

太棒了。如果大家想去拿这些 skill，能给我们看看去哪儿拿吗？

[55:37] Oji

Yes, I can at cash. So, I am going to share different screen. Um, so one of the fun things we've done at Product Mind as we start to build is start to collect these things. Some of them are private. They're not available yet. But if you go to the labs of product.com, don't just ignore these things. This is not a monetizable site yet or at all. This is just to show the products that we can actually share from our engagements. But everything you've seen actually can be accessed more trivially through what we call skills. So you can chat with it without going into cloud code. So that's the first thing. The second one and we have a lot of people you know who do this now. And the second one is a skills library. Uh the skills are open source. So you can come here and you can browse through them, read them. Here you have the new project scaffolding, vet a feature, vibe memo. I'm very very excited about this one. Vibe memo is a simple system. We installed it that captures your decisions as you build. So right now when you write code that's captured uh but vibe memo says capture the why. Why are we making these decisions in the code? So it keeps the why with all the decisions in the log for you. I find this very exciting. So every time we have a big codebase, we want to take a look at why did we even make that decision? What were we thinking? Vibe member captures it. Sharp problem test reverse fee.

可以，Aakash。我换个屏幕共享。我们在 Product Mind 一边做一边攒的一件有意思的事，就是开始把这些东西收集起来。有一部分是私有的，还没开放。但你去 ofproduct.com 的 labs 页面看看——别急着划走——这不是个能变现的站，完全不是，只是把我们在客户项目里能对外分享的产品放出来。不过你今天看到的所有东西，其实都能用我们说的 skill 更轻松地拿到，不进 Claude Code 也能直接跟它对话，这是第一点。现在挺多人这么用了。第二点是一个 skill 库，这些 skill 都是开源的，你可以进来一个个翻、一个个读。这里有 new project scaffolding、vet a feature、vibe memo——vibe memo 这个我特别兴奋。Vibe Memo 是个很简单的系统，我们装了它，它会在你构建的过程中记录你的决策。现在你写的代码是被记录下来的，但 vibe memo 说：把「为什么」也记下来。我们为什么在代码里做这些决定？于是它把每个决策连同背后的原因一起留在日志里。我觉得这太带劲了。以后代码库一大，我们想回头看「当初到底为什么那么定？当时怎么想的？」，vibe memo 都替你记着。还有 sharp problem test、reverse fee……

[57:19] Oji

There's so many. Let me just point out a few more favorites. One is road map from strategy. So it'll construct a whole like strategy for you. Um another one is a listing machine. It will turn all the surfaces in your product into feedback tools that constantly collect information. Uh these are things that people spend years and years trying to figure out. We build it into a skill that gets you started immediately. Uh scope cutter, we all do this. We got to cut scope. And scope cutter gives you a very succinct way to reduce how much you do. There's even a simpler one which is once you want to build the first thing are you going to do an MVP which is just for learning or are you going to do an SLC which is actually to put in front of people for them to be amazed and how do you even do that and how do you cut and so on and so forth and of course advanced things like pricing like how should you design pricing all built in. So again, business and product level skills that merge with code skills in one place for you to become hyperproductive. Um, we think this is the new frontier. Um, I think code skills alone are becoming p. Uh, the other day I saw someone remove superhuman skills from their repo because the models are so good. Um, some of these skills are breaking. But [clears throat] these business skills, these product skills, these are still the things that give us the ability to build the right thing. So, I'm super excited about that.

太多了。我再挑几个我特别喜欢的。一个是 roadmap from strategy，它能替你把一整套战略搭出来。另一个是 listening machine，它会把你产品里的所有界面都变成反馈工具，持续不断地收集信息。这些都是别人琢磨好多年才摸出来的东西，我们把它做进一个 skill，让你立刻就能上手。还有 scope cutter——砍范围这事谁都躲不掉，scope cutter 给你一套非常干脆的办法，把要做的量降下来。还有个更简单的：你要做的第一版，到底是 MVP（纯粹为了学到东西），还是 SLC（真的拿到人面前、要让他们眼前一亮）？这该怎么做？怎么取舍？等等等等。当然还有更进阶的，比如定价该怎么设计，全都内置好了。所以还是那句话：业务层和产品层的 skill，跟代码层的 skill 合在一个地方，让你变得极其高产。我们觉得这是新的前沿。我觉得光有 code skill 正在变得没什么稀奇——前几天我还看到有人把 superhuman skills 从 repo 里删了，因为模型已经足够好了，有些 skill 就这么被废掉了。但这些业务 skill、产品 skill，仍然是让我们有能力「做对的东西」的那部分。所以我对这块超级兴奋。

[58:51] Aakash

And today we showed the harness as claude code in anti-gravity. Is there any particular reason you should be using anti-gravity or can you use any harness for cloud code?

今天我们演示的 harness 是跑在 Antigravity 里的 Claude Code。有什么特别的理由非用 Antigravity 不可吗？还是说 Claude Code 用什么 harness 都行？

[59:00] Oji

No, it's any harness. Um, I don't even particularly So, anti-gravity is a fork of VS Code. Uh, so it looks a little bit like cursor. It looks like VS Code. It does. That's just what I have on my screen. I'm running through a terminal and so I actually don't use any features of anti-gravity. It's just, you know, it's a nice interface, but nothing special about it.

没有，什么 harness 都行。Antigravity 是 VS Code 的一个 fork，所以看起来有点像 Cursor，也像 VS Code——本来就是。它只是刚好开在我屏幕上。我其实是在终端里跑的，Antigravity 的功能我一个都没用。界面挺好看，但没什么特别的。

[59:20] Aakash

So, if you're a product leader, right, you go ahead and you download these skills, you make these skills your own, you distribute them to the team, the team is starting to use them.

所以，如果你是产品负责人，你把这些 skill 下下来，改造成你自己的，再分发给团队，团队开始用起来。

[59:32] Aakash

You are one of the few people who has actually helped teams do that. So you've gotten to see what are the mistakes that people make? What are the problems or the pitfalls in rolling out this new way of working that people should know about so that they can avoid them?

你是少数真正帮团队落地过这件事的人。所以你都见过大家会犯哪些错？推行这种新工作方式的时候，有哪些坑或者陷阱是大家该提前知道、好绕开的？

[59:45] Oji

I think that the most important thing is shared context. You know, I heard um Boris um from Quad, right, talk about the fact that his quad.mmd is a very thin, it has only a few lines, maybe six lines. And what it's actually doing is that it's referencing a shared claude MD. Now, you should think about claude MD as some some kind of like base skill basically. And so centrally they author claude MD and optimize it like think about once a minute everything that everyone is learning is being put into that central cloud MD it's small it's short it contains essence it connects people to all the right tools within the organization so I think that the biggest lesson for putting this into an enterprise is don't let people fork these things willy-nilly, right? For example, I was working with an enterprise that had a ways of working document and then they had a very specific SDLC that was supposed to produce very specific artifacts. Okay. What I would do in that uh situation is to take the relevant skills here for example the uh scaffolding skill or the vet a feature skill and I would harmonize it with what they need to come out of it. Why would our product brief be different from what the organization wants? It shouldn't be right. Uh why would a market research document be different from what the product marketing team would produce or what they believe is right to produce?

我觉得最重要的是共享上下文。我听 Anthropic 的 Boris 讲过，他的 CLAUDE.md 非常薄，只有几行，可能就六行。它实际在做的事，是去引用一份共享的 CLAUDE.md。你可以把 CLAUDE.md 理解成某种基础 skill。他们在中心统一维护和优化这份 CLAUDE.md——你想想看，几乎每分钟，所有人学到的东西都会被汇进那份中心的 CLAUDE.md；它小、它短，装的是精华，还把人连到组织里所有该用的工具上。所以我觉得，要把这套东西搬进企业，最大的一课是：别让大家随随便便各自 fork。举个例子，我合作过一家企业，他们有一份「工作方式」文档，还有一套非常具体的 SDLC，要求产出非常具体的交付物。这种情况我会怎么做？我会把这里相关的 skill 拿过来，比如 scaffolding skill 或者 vet a feature skill，然后把它跟他们要的产出对齐。我们的 product brief 凭什么要跟组织想要的不一样？不该不一样，对吧。一份市场调研文档，凭什么要跟产品市场团队会产出的、或者他们认为该产出的东西不一样？

[1:01:40] Oji

And so we would essentially centralize this, modify it so that it creates artifacts that our organization needs and just prevent people necessarily from forking it themselves. Now to be clear, I like organizations where it's sort of open source and if people learn new things because you know the CPO doesn't know everything, you give people the license to innovate or if they innovate, they should innovate into a central place so everyone gets the innovation if that makes any sense. So I think that's the biggest thing is when AI rollouts are completely ungoverned in a way that doesn't have shared context because look at listen I think in science fiction we always sort of uh afraid of the hive mind right we're afraid of AIs collaborating because then they'll kill us but it turns out that we are the original hive mind the only reason that we've survived you know through the centuries as humans is because we learn from each other. And so this is very imperative when you roll out AI, try to force it into a a structure that one person learning means everyone has learned. That's one of the big powers of it.

所以我们的做法基本上是把这套东西集中管起来，改造成能产出我们组织真正需要的产物，同时尽量别让大家各自 fork 一份自己玩。当然说清楚一点：我其实挺喜欢那种开源式的组织——CPO 不可能什么都懂，所以你得给大家创新的许可；但如果他们真创新出了什么，那创新应该回流到一个中心位置，让所有人都能享受到，不知道我说清楚没有。所以我觉得最大的坑就是：AI 的推广完全没有治理，也没有共享上下文。你看，科幻作品里我们总是害怕 hive mind（蜂巢思维）对吧，我们怕 AI 互相协作，因为那样它们就要灭了我们。但事实是，我们人类才是最原始的那个 hive mind——我们能活过这么多个世纪，唯一的原因就是我们会互相学习。所以在推 AI 的时候这点特别关键：想办法把它塞进一个结构里，做到一个人学会了就等于所有人都学会了。这是它最大的威力之一。

[1:02:52] Aakash

OG, you've been generously shared these skill files. You guys can find the link in the description to this GitHub repo. As he said, you can make these your own. That's really the key. adding it in to make it your own and then deploying it as an open source repo for your team to use. This is the playbook. If you've been wondering, how do I become a more AI native team? How do I get my PM team to work at an engineering speed? We've just given all of that away for free. Thank you so much for being here.

Oji，你这次特别大方，把这些 skill 文件都分享出来了。大家可以在视频简介里找到这个 GitHub repo 的链接。就像他说的，你可以把它变成你自己的——这才是关键：把你们自己的东西加进去，让它长成你们的样子，然后作为一个开源 repo 部署给团队用。这就是那套打法。如果你一直在琢磨「我怎么才能让团队更 AI native？怎么让我的 PM 团队跑出工程师那样的速度？」——我们刚刚把这些全都免费送出来了。非常感谢你今天来。

[1:03:22] Oji

Thank you, Aash. And I I hope uh people become even more effective product builders and hyper creators with the stuff that we've made. Super happy to share.

谢谢你，Aakash。我也希望大家能靠我们做的这些东西，成为更高效的产品打造者、成为超级创造者。非常乐意分享。

[1:03:34] Aakash

I hope you learned as much from today's episode as I did. If you can do one thing that's totally free that would help the show, it would be to [music] check that you're following on Apple and Spotify podcasts. Check that you've left ratings and reviews on those platforms. Check that you're subscribed on YouTube. Leave a like and a comment on this video. and then share it with your friends. We're trying to make better and better podcasts. After 2 years, we think we've gotten something pretty good going. So, let us know what we can do to make it even better, who else we should interview, and we will put on the best shows we possibly [music] can. Finally, don't forget my offer for the bundle. You get an entire year of my paid newsletter, plus my favorite AI tools. Bolt, [music] new, air table, speechify, descript, magic patterns, linear, dovetail, arise, and mobin. That's $27,000 worth of value for just $150. So check that out at bundle.acushi.com if it interests you. And I can't wait to share our next episode soon.

希望你今天从这期节目里学到的，和我一样多。如果想帮这个节目一个忙，有件完全免费的事你可以做：去 Apple Podcasts 和 Spotify 上确认你已经关注了，顺手留个评分和评价；在 YouTube 上确认你已经订阅，给这条视频点个赞、留个评论，然后分享给你的朋友。我们一直想把播客越做越好。做了两年，我们觉得已经跑出点样子了。所以也告诉我们还能怎么改进、还该请谁来聊，我们会尽全力做出最好的节目。最后，别忘了我那个 bundle 优惠：你能拿到我付费 newsletter 的一整年订阅，外加我最喜欢的一批 AI 工具——Bolt.new、Airtable、Speechify、Descript、Magic Patterns、Linear、Dovetail、Arize 和 Mobbin。总价值 27,000 美元，你只要 150 美元。感兴趣的话去 bundle.aakashg.com 看看。下一期节目，我们很快见。