The One Skill Stopping You From a $300K+ AI PM Offer

频道: Aakash Gupta
视频: https://www.youtube.com/watch?v=iNFE_u5hw3U
原文语言: en
统计: 共 26 轮

[0:00]

I have just finished AI interviews at Uber, Atlassian, Cisco, Roblox, and a few other companies. One of the fastest growing AIPM interviews is the wipe coding interview where you are essentially designing a product and taking it all the way to a prototype. I've been hearing from more and more PMs I'm coaching that they're encountering a new PM interview round. This isn't the traditional product sense round, and it isn't an engineering coding round either. It's a hybrid of the two and I'm calling it the vibe coding round. AIPM roles at companies like Google, Meta, OpenAI, Anthropic, pay anywhere from 300,000 to north of a million dollars. I've been interviewing AIPM leaders like Jiona Zen, the CPO at Laurel AI, who literally ask you to screen share how you use AI in the interview. If you're nervous about such an upcoming interview, then you need to watch today's episode. Every video out there on vibe coding and PM interviews is about building a from scratch new product. But the questions you're actually most likely to see are building features on existing products. Today's episode will go through a mock interview for you to actually see what it's like if you have to build an existing feature for a new product, show off your AI prototyping and vibe coding skills in an interview live. Very excited to be interviewing you today. We'll be doing this interview for about 45 minutes and for about 30 minutes of this, what I'd love for uh you to do is a wipe code uh a feature or capability and I'll be giving you a prompt but would love to see how you go about the process of white coding this out. Okay. Um and today we're going to be building a feature for LinkedIn. So assume that you are a product manager at LinkedIn and of course LinkedIn really cares about you staying in touch with your network and what we are also realizing is people are generally apprehensive of doing so and we're

我刚刚跑完了 Uber、Atlassian、Cisco、Roblox 还有另外几家公司的 AI 岗面试。现在增长最快的 AI PM 面试题型之一，就是 vibe coding 面试——你要设计一个产品，并且一路把它做到 prototype。我辅导的 PM 里，越来越多人跟我说他们碰到了一种新的 PM 面试轮次。它不是传统的 product sense 轮，也不是工程师的 coding 轮，而是两者的混合体，我管它叫 vibe coding 轮。Google、Meta、OpenAI、Anthropic 这些公司的 AI PM 岗位，薪酬从 30 万美元一直到 100 万美元以上。我采访过不少 AI PM 领域的负责人，比如 Laurel AI 的 CPO Jiona Zen，他们是真的会让你在面试里共享屏幕，展示你怎么用 AI。如果你正为这样一场面试发怵，那今天这期你一定得看。市面上所有讲 vibe coding 和 PM 面试的视频，都是在从零做一个全新产品。但你在真实面试里最可能遇到的，其实是给现有产品加功能。今天这期会带你走完一场完整的 mock interview，让你真切看到：当你要在一个已有产品上做一个功能、还要在面试现场展示你的 AI prototyping 和 vibe coding 能力时，到底是什么样子。今天非常高兴能面试你。这场面试大约 45 分钟，其中差不多 30 分钟，我希望你来 vibe code 一个功能或者能力。我会给你一个 prompt，但我最想看到的是你 vibe coding 的整个过程。好的。今天我们要给 LinkedIn 做一个功能。假设你是 LinkedIn 的产品经理，LinkedIn 当然非常在意你能和你的人脉保持联系；同时我们也发现，人们普遍对主动联系这件事有顾虑，我们希望

[1:53]

hoping that AI can reduce some of these barriers. So help me build or conceptualize a feature and then of course take it to a prototype where you within the constraints of LinkedIn are doing a better job of staying in touch with your network. Does that sound good? Okay. So, let me understand this interview correctly. In about 30 minutes, you want to see how I actually vibe code and prototype a solution where we're helping people stay in touch with their network with AI. That is correct. And of course, this has to be an AI first while at the same time, I would like you to see you using AI to build a prototype. Okay. Fascinating. What is the goal of this feature? Why are we building this feature versus any of the other things we could be building? Yes. So, what we have found out is that of course users come to LinkedIn to consume content, but that is more of a passive activity. We'd want users to be actively engaged with the rest of their network, which involves a certain cognitive load or barrier. And we are hoping to bring down this cognitive load and barrier for users so as to then increase retention and engagement with the platform in general. Got it. Just to repeat it back, as I understand the mission and vision of LinkedIn as a whole, you guys have talked a lot about how you want to create economic opportunity. People get economic opportunity by retaining on the platform and engaging with their network. And so the goal is to move them from content passive consumers to active engagers in their network. Correct. Awesome. I'm going to take like two to three minutes if that's okay just to like gather my thoughts. Is that fine? Sounds good. Thank you.

AI 能降低其中的一些障碍。所以请你帮我构思或者设计一个功能，然后当然要把它做成 prototype——在 LinkedIn 这个产品的约束条件下，让人们更好地和自己的人脉保持联系。这个题目 OK 吗？好的，那我复述一下确认我理解对了：大约 30 分钟内，你想看我实际怎么 vibe code、怎么把一个解决方案做成 prototype，这个方案是用 AI 帮人们和自己的人脉保持联系。没错。而且当然，它必须是 AI first 的；同时我也希望看到你用 AI 来搭这个 prototype。好，很有意思。这个功能的目标是什么？为什么我们要做这个功能，而不是其他任何我们可以做的事情？是这样，我们发现用户来 LinkedIn 是为了消费内容，但那更像是一种被动行为。我们希望用户能主动和自己的人脉互动，而这件事本身带有一定的认知负担或者说心理门槛。我们希望把这个认知负担和门槛降下来，从而提升平台整体的留存和参与度。明白了。我再复述一遍：按我对 LinkedIn 整体使命和愿景的理解，你们一直强调要创造经济机会。人们获得经济机会的方式，就是留在平台上、和自己的人脉互动。所以目标是把他们从被动的内容消费者，变成在自己人脉网络里主动互动的人。对。太好了。如果可以的话，我想花两三分钟整理一下思路，行吗？没问题。谢谢。

[3:56]

Okay. I don't even think I needed two to three minutes. So, I think 30 seconds is fine cuz I'd rather just talk through this all together. Does that work for you? Let's do it. All right. So, let me share my screen cuz I know you wanted to see how I do all this, right? Well, actually, I'm going to start in a lowfi way [laughter] without AI or anything because I think it is important. The way I think about using AI is AI can do the middle 60% of the work, but the 20% before the initial thinking is still important to do as a human and then the initial refinement or the final refinement is really important to do as a human. So I want to start as a human. And here's the thing, we only have like 28 more minutes, right? So, I just want to spend like five or six minutes almost speed running what would be a typical 30 minute product sense interview that I know you guys also conduct and I did in one of my other rounds. And then I want to go to the screen share part. And how I want to do that is like think about what is the PRD per se for the solution we choose. Prototype it across a couple different tools. Choose a tool iterate on that one and maybe even build out backend at the end. Does that sound good? That sounds reasonable. Yes. All right. So, I haven't thought about this any further than just putting this user tag. So, let's think about it out loud together. Right. I think that there's sort of a 99% 1% users of LinkedIn like you said.

好，其实我觉得我都不需要两三分钟，30 秒就够了，因为我更想边想边跟你一起讲。这样可以吗？那就开始吧。好的。那我来共享一下屏幕，因为我知道你想看我具体怎么做，对吧？不过其实我打算先用一种很土的方式开始，先不用 AI 或者任何工具，因为我觉得这一步很重要。我对使用 AI 的理解是：AI 可以做中间那 60% 的活儿，但前面那 20%——最初的思考——依然要人来做，最后的打磨、终稿的收口，也非常需要人来做。所以我想先以人的方式开场。而且我们只剩大概 28 分钟了对吧？所以我只想花五六分钟，几乎是用倍速跑一遍一场典型的 30 分钟 product sense 面试——我知道你们也会考这个，我在另外一轮里也做过。然后我就切到共享屏幕的部分。我的打算是：先想清楚我们选定方案的 PRD 大概长什么样，然后用几个不同的工具分别做 prototype，选定一个工具、在它上面迭代，最后可能还会把 backend 也搭出来。这个安排听起来 OK 吗？听起来挺合理，可以。好。其实我目前的思考就只到打上这个用户标签这一步，所以我们一起边想边说吧。我觉得 LinkedIn 的用户大致可以分成 99% 和 1% 两类，就像你说的那样。

[5:23]

And so, I just want to graft off the context you already gave me. 99% of people they consume content. 1% they write content. I guess you know this this point is kind of orthogonal even to how much they engage with network. Another vector that we might want to think about is probably like in any given month, you know, some crazy percent, who knows if it's 99 or what, but like I'd guess it's a little less than 99, 95% don't engage with their network and 5% do, right? And so what we really want to do is we want to start to think about across those 95% who don't engage in the network, who are those people and potentially who do we want to build for? And so I would start to think about okay well what are the various buckets that we should put together for these people? And so, you know, some of them I think a really important and relevant dimension is going to be whether they're job searching right now, which again is probably only like 5% or something. And the reason I'm putting numbers is just cuz I want to get a sense of how big these groups are, right? So 5% are probably job searching right now. And then, you know, 80% probably want to, you know, keep up their professional profile for when they job search. And that's where like networking might have certain use case that might be different from job searching. And then there's probably like 15% of people I'm just trying to come up with a me sort of bucketing at first who like couldn't care less about their career. You know, [laughter] on the low end, we have the people who aren't career optimizers. On the high end, we have people who are already too rich. Like I think it was there was the joke was that uh the new CEO of Apple, he didn't even have a LinkedIn, right? And so there's people on both sides that couldn't care less about their career. And so what I'm seeing right away by putting some numbers on this is, you know, this bucket here is going to be really interesting to

我就顺着你刚才给我的背景往下推。99% 的人是消费内容的，1% 的人是生产内容的。不过我猜这个切分跟他们跟人脉互动多少其实是两个正交的维度。另一个我们可能想考虑的维度是：在任意一个月里，会有一个很夸张的比例——不知道是不是 99%，我猜稍微低一点，大概 95% 的人完全不和自己的人脉互动，5% 的人会互动，对吧？所以我们真正要做的，是去拆解那 95% 不互动的人里，都是些什么人，以及我们到底想为谁做产品。那我就开始想：这些人应该分成哪几个桶？其中我觉得一个非常重要且相关的维度，是他们现在是不是正在找工作——这部分大概也就 5% 左右。我之所以要给出数字，是因为我想对每一组有多大有个体感。所以 5% 大概是正在积极找工作的。然后大概 80% 是想维护好自己的职业形象、为将来找工作做准备，这里的 networking 场景就和正在找工作的人不太一样了。剩下大概 15%，我只是先粗略分一下，是那种压根不在乎自己职业发展的人。在光谱的一端，是不做职业优化的人；在另一端，是已经太有钱了的人。我记得有个梗说，Apple 新任 CEO 连 LinkedIn 账号都没有，对吧？所以两端都有一批完全不在乎自己职业的人。给这些桶标上数字之后我马上就看出来了：这个桶——

[7:19]

play with people who want to keep up with their professional life. And the job searching bucket might end up being the easiest to capture. So this is the biggest. This is the easiest to capture. And so I think we can focus on both these as focus areas. And we don't need to necessarily like say, "Oh, we're only going to use one." So, if we look at both of these as focus areas, let's like bold these. Look at these, you know? So, we can just talk about these briefly cuz we're just speed running through this whole process. Job searching. I mean, these people are looking for referrals, right? Referral is like how you get a modern job. These people are looking to create connections at target companies. These people are looking to talk to people ahead of them in life. So it doesn't even really mean ahead of them. Ahead of them in career, right? So maybe this person is a PM at Google and I really want to be a PM at Google. Even though I'm not targeting referral, I want to talk to them so I can learn how they did things. They also might be like trying to reach out and like warm up their own network, right? Like their alumni and different things like that. So I think we got a good idea of the context of a job searcher. So I'm going to move on to the professional profile. Does that work for you? Sure. That works great. Okay, perfect. So, if you want to keep up your professional profile, it's almost like different now, right? Because you don't care as much about referrals. You do care a lot about talking to people ahead of your career.

也就是想维护自己职业生活的人——会非常值得深挖。而正在找工作那个桶，可能是最容易切进去的。所以：这个最大，那个最好切。我觉得我们可以把这两个都作为重点方向，没必要非说「我们只选一个」。那既然这两个都是重点，我们把它们加粗一下。好，我们简单过一下，因为整个过程我们是在赶进度。先说正在找工作的人。这些人在找 referral，对吧？referral 基本上就是现代人拿到工作的方式。这些人想在目标公司里建立连接。这些人想跟走在自己前面的人聊——不是人生走在前面，是职业上走在前面。比如说那个人是 Google 的 PM，我特别想成为 Google 的 PM，就算我不是冲着 referral 去的，我也想跟他聊聊，学他是怎么做到的。他们还可能想主动去激活、去暖一暖自己已有的人脉，比如校友之类的。我觉得找工作这个人群的画像我们已经比较清楚了，那我转到「维护职业形象」这一组，可以吗？可以，很好。好，那如果你的诉求是维护职业形象，情况就几乎完全不一样了，因为你没那么在乎 referral，但你非常在乎跟职业上走在你前面的人交流。

[8:42]

So, that's almost like there's also an arrow to this bucket for that. Then we have to think about, well, what are the other people who want to keep up with their professional profile? Why do they what I almost am trying to do is think about what are some of the triggers, right? because as we build this product, there's going to be some trigger or some use case or why would they care about keeping in touch with their network, right? Maybe um they're sort of passively job searching. I think this might be one of the most easy people like people who want to for instance build out more connections at their target companies, talk to people ahead in their career. So then they end up very similar to this bucket. And so we see two two overlapping sort of needs across these groups already. Then we say, okay, are there any unique needs for wanting to keep up a professional profile? Um, you might want to like uh well, I'm just trying to think outside the box even though these aren't all Yeah. One of the reframes I can probably suggest here, Akash. Um, is I'd love for you to focus a little bit more on what are the hurdles that you see that stops people from connecting to their network. Right. Like I like where you are spending time on the users. I think I have a good idea of where you're going with this, but if you could talk about what is it that really stops people or makes it awkward for people to reach out. Fair to assume that they are soft searching all the times. That's I think is a is a is an okay assumption to move forward with. Um but I'd love to hear how do you really want to get into that problem statement here. Perfect. So we'll kind of focus on the soft job searchers and thanks for keeping me going fast here on the users. That'll be where we focus on the users. So what you're asking next is like what are the problems, right? What are the problems that we want to overcome? So one of the

所以从这个桶也可以拉一条箭头指向那边。接下来我们得想：还有哪些人是想维护自己职业形象的？他们为什么要这么做？我其实想找的是一些触发点（trigger），因为我们做这个产品的时候，一定得有某个触发点、某个使用场景，来解释他们为什么会在意跟人脉保持联系，对吧？比如说他们处在一种「被动找工作」的状态。我觉得这可能是最容易切入的一群人，比如那些想在目标公司多建立几个连接、想跟职业上走在前面的人聊聊的人。这样一来他们就跟前面那个桶非常像了。所以我们已经能看到，这两组人之间有两个重叠的需求。然后我们再问：想维护职业形象的人，有没有什么独有的需求？嗯，你可能会想……我只是在尽量跳出框架想，虽然这些不一定都——对，Aakash，我这里可以给你提一个重新聚焦的建议。我更希望你多花点时间讲讲：你觉得阻碍人们去联系自己人脉的障碍是什么？我喜欢你在用户分层上花的功夫，我大概能看出你要往哪走，但如果你能讲讲究竟是什么让人们不敢联系、或者让联系这件事变得尴尬，会更好。我们可以假设他们一直处在「被动找工作」的状态，我觉得这是一个可以往下推的合理假设。但我很想听你怎么真正切进这个问题定义。好的。那我们就聚焦在「被动找工作」的这群人上，也谢谢你帮我在用户这块加速。用户就定在这里。你接下来问的是：问题是什么，对吧？我们要克服的问题有哪些？其中一个

[10:28]

big things is like you even forget how you knew someone. You forget why you might need to talk to someone. Like you don't even know the use case of why you might need to. Maybe, you know, an example of this just to give is like you're trying to do a deal with their company or you're going to a conference that they're going to be at or um you are solving a problem that they also solve, right? Um they also solved in their job. Uh you so that's one area. I think there's just this whole amount of like product friction that of course we can explore and that's probably going to be like pretty fruitful area for us to explore in terms of prototype design. So I think like here we can just do an audit together. We can go look at the whole flow together. Let's do it really fast. So let's switch over to the product and let's go take a look. Right. So as I jump in, right, if I hit something like my network tab, okay, this takes me to invitations. So then there's this catch-up feature. Look at this. We've got an empty state in our catchup feature. This to me immediately is like, okay, as we go through the audit, we could go through all the steps of the flow and that could be an interesting first area to put at. So I've dropped it here into the miro. Then you know, job changes, birthdays. Look at this. Okay, so we've just because I've turned this off, I see.

很大的问题是，你连自己当初是怎么认识这个人的都忘了。你也忘了为什么需要跟他聊。你甚至不知道自己找他能有什么用。举个例子：你可能正想跟他公司谈一笔合作，或者你要去参加一个他也会去的会议，或者你正在解决一个他在自己工作里也解决过的问题。这是一块。另外我觉得还有一大堆产品层面的摩擦，这些当然可以去挖，而且对我们做 prototype 设计来说，这大概是一块很有产出的地方。所以我觉得我们可以现在一起做一次审视（audit），把整个流程走一遍。我们快点来。那我们切到产品里看一看。好，我进来之后，点一下「我的人脉」这个 tab——好，它跳到了邀请列表。然后这里有个「叙旧」（catch-up）功能。你看，我们的 catch-up 功能是个空状态。这对我来说马上就说明：我们把流程每一步都走一遍的话，这可能就是第一个值得下手的地方。我已经把它记到 Miro 里了。然后你看，工作变动、生日。看这个。哦，是因为我把这个关掉了，我懂了。

[11:53]

So let me turn this on for a second. So that's so weird. like this is tied to our notifications. Like why is our updates section tied to our notifications? So that's like another audit that we could do. To continue the audit, I'll go ahead and turn it on. And so let's go ahead and see what happens if I turn this on. This is so weird because I have to turn on inapp notifications in order to get access to job changes. Super odd. Okay. And then it ends up saying no recent updates and then grow your network drops you back here. So basically that button is just completely broken. Right? So we can see that the current workflow for the product has a lot of friction in even discovering who to talk to and there's probably a lot of fruit we can have just from optimizing that. Then let's go ahead and assume that I figured out who to talk to because we revealed it correctly. So I'm connected to this guy. He's member of technical staff at Anthropic. when I click in, I don't know how or why I'm connected to this guy, when we even connected, right? So, I'm totally lacking in this information. So, I think there's almost like two elements that have come out of the audit so far. Area one is that we aren't recommending who to talk to very well. So, that's one area we need to improve on. And then the second area we need to improve on is once you get to somebody giving you the information to talk to them and then I know we've kind of walked back our AI writing comments but I think there's probably room for a feature like that and let's check right I want to take a second to talk to you about the fourth cohort of LAN PM job. I trained 30 students in cohort 1, 50 students in cohort 2, and 75 students in cohort 3. And I am bringing back [music] the program for cohort 4. It starts in August and it lasts 3 months where you're going to have intense sessions, a [music] Monday morning session where I go over your resume,

那我把它打开一下。这太奇怪了，这个东西居然跟我们的通知绑在一起。为什么「动态更新」这一栏要跟通知绑定？这又是一个可以审视的点。为了把 audit 继续下去，我把它打开。我们来看看打开之后会发生什么。这真的很怪，我必须打开应用内通知，才能看到工作变动的信息。非常奇怪。好，然后它显示「暂无近期更新」，接着「拓展人脉」又把你甩回了这里。所以那个按钮基本上是完全坏的，对吧？由此可见，当前这个产品流程，光是在「发现该跟谁聊」这一步就有大量摩擦，仅仅是把这块优化好，可能就有很多果子可以摘。那我们接着假设我已经搞清楚该跟谁聊了，因为系统正确地把人推给了我。我跟这个人是好友，他是 Anthropic 的 member of technical staff。我点进去，却完全不知道我是怎么、为什么跟他连上的，也不知道我们是什么时候连上的。这些信息我完全没有。所以我觉得，到目前为止这次 audit 已经暴露出两块问题。第一块是：我们在推荐「该跟谁聊」这件事上做得很差，这是需要改进的一块。第二块是：当你定位到某个人之后，要给你足够的信息去跟他聊。我知道我们在 AI 代写评论这件事上其实是往回收的，但我觉得类似的功能还是有空间的，我们来验证一下。对了，我想花一点时间跟你聊聊 Land PM Job 第四期。第一期我带了 30 名学员，第二期 50 名，第三期 75 名。现在我要把这个项目带回来做第四期了。第四期从八月开始，为期三个月，会有非常高强度的课程安排——周一早上有一场我逐份过简历的 session，

[13:57]

behavioral interviews, LinkedIn. On top of that, Bart Jorski is going to be teaching [music] you the PM fundamentals in 2026. How to write AI PRDS, how to AI prototype with cloud code, all of the key skills you need to freshen up your knowledge [music] for this market. And Ang Vermani is going to be teaching you AI product management. He is an AI product manager at Uber and he is going to teach you how to build AI features that actually work successfully. On top of that, Prasad Ready is going to be doing one-on- ones with you for [music] market reviews, LinkedIn review, candidate market fit review. So it is a full package. It is three courses in one for one low fee. So join at landpob.com. So then the next area that there might be friction is if I'm messaging somebody. So if I go to message her, what happens when I hit write with AI? Introduce yourself. Okay. So it takes quite a while, but we do have an AI feature. This is good. So I think part of it might be we have a pretty strong AI feature. What if we bring that forward and get it to people sooner? So pretty strong number of m dashes in there do scare me though Akash and yeah true means that is not at all humanized plus it is not pulling in the right context from either your past interactions or your broader relationship with that person. Yes, I was thinking about some sort of past interaction public um [snorts] widget. So I'm glad you brought that up. Uh so Just double click on that a little bit. Akash, see the the thing that we think with LinkedIn is that the true value of your network lies in relatively weak ties, which is why the person that you potentially selected here is someone that you haven't touched base with in a bit, right? Or in some cases, you don't even remember who that person is, right? And that the problem is not that you forget, it's just that you don't have a nonawward reason to reach out.

……行为面试、LinkedIn。除此之外，Bart Jaworski 会给你讲 2026 年的 PM 基本功：怎么写 AI PRD、怎么用 Claude Code 做 AI prototype，所有你在这个市场里需要重新打磨的关键技能。还有 Ang Vermani 会讲 AI 产品管理——他是 Uber 的 AI PM，会教你怎么把真正跑得起来的 AI 功能做出来。另外 Prasad Reddy 会跟你做一对一：市场情况复盘、LinkedIn 主页诊断、候选人与市场的匹配度评估。所以这是一整套打包——三门课合一，只收一份很低的费用。来 landpmjob.com 报名。

好，那接下来可能有摩擦的地方，是我给别人发消息的时候。我点进去给她发消息，点「Write with AI（用 AI 写）」会发生什么？「介绍一下你自己」。嗯，等了挺久的，不过我们确实有这个 AI 功能，这是好事。所以我觉得其中一个点可能是：我们本身有一个挺强的 AI 功能，那能不能把它往前挪、让用户更早碰到它？

不过这里面破折号（em dash）多得有点吓人啊，Aakash。

对，确实——意思就是这一点都不像人写的；而且它也没有把正确的上下文拉进来，不管是你跟这个人过去的互动，还是你跟他更大范围的关系。

对，我刚才就在想要不要做某种「过去互动」的公开 widget，所以你提到这个我很高兴。

呃，那你就着这个再往下挖一点。Aakash，我们在 LinkedIn 内部的看法是：你这张人脉网真正的价值在于那些相对弱的关系（weak ties）。所以你刚才选的这个人，多半是你有一阵子没联系过的人，对吧？有些情况下你甚至都不记得这人是谁了。问题不在于你忘了，而在于你找不到一个不尴尬的理由去联系他。

[15:54]

And your audit definitely revealed this as a pretty big gap which is you are being prompted as it is almost a notification that you're receiving and that notification is meant to be an organic reason for you to reach out but doesn't do a great job. On top of that, cold outreach itself is hard and doesn't pull in the contextuality. And I think that's the frame that I'd love for you to carry through the interview. Both the fact that it's awkward to reach out for a non-genuine reason and B that cold outreach is generally hard for people. That is what we have observed with our users. Awesome. Okay. I'm glad we were thinking alike. Uh let's do that. So, let's just go ahead and assume that I was going to do a full exploration of the problem and solution space and prioritize the problem and solution, but it sounds like you're saying let's just move forward with a solution that's kind of solving this problem along the lines of reaching out to weak connections with context and a relevant message um easily. I' I'd love to see more of the prototyping part. Akos, that's why I am sort of pushing in this direction. Perfect. Yeah. So to build a good prototype, we need to clearly create the right context now. And so now we can go into well, what does the PRD for this look like, right? And so some of the key elements that we need to put into here are what is the um the overall goal that we're trying to drive economic opportunity by um getting back together with your weak connections. And the goal really is or some of the non goals first, right? The non- goals, like you just said, are like spam messages cuz people are going to leave our platform or out of context messages or messaging people you don't know that but are just connected to. And then sort of like the metric of success around this, as you said, is going to be there's going to be like leading and lagging metrics. So, the leading metrics we're going to care about are like people

你刚才那轮 audit（体验走查）确实把这个大缺口挖出来了：系统提示你去联系，形式上几乎就是你收到的一条通知，而这条通知本来该给你一个自然而然去联系的理由，但它做得并不好。除此之外，cold outreach（冷启动式主动联系）本身就很难，而且没有把上下文带进来。我希望你在接下来的面试里就抱着这个框架往下走：A，为了一个不真诚的理由去联系别人很尴尬；B，冷联系对大多数人来说本来就难。这是我们在用户身上观察到的。

太好了，我们想到一块儿去了。那就这么办。那我们就直接假设：我本来是要把问题空间和方案空间完整探索一遍、再给问题和方案排优先级的，但听你的意思是，我们直接推进到一个方案——大致就是让用户能轻松地、带着上下文和一条相关的消息，去联系那些弱关系。

我更想多看看 prototype 那部分，Aakash，所以我才往这个方向推。

完美。那要做出一个好 prototype，我们现在就得先把正确的上下文建立起来。接下来我们可以进入：这东西的 PRD 长什么样？需要写进去的几个关键要素——总体目标是什么：通过重新连上你的弱关系来创造经济机会。目标是……或者先说非目标（non-goals）：非目标就像你刚说的，比如垃圾消息，因为那会让人离开我们的平台；比如没有上下文的消息；比如给你其实不认识、只是加了好友的人发消息。然后是这件事的成功指标，正如你说的，会分成先导指标（leading）和滞后指标（lagging）。先导指标我们关心的是有多少人……

[18:15]

read the message, people respond to the message. I don't think we're actually reading their messages, but if we were actually reading their messages, we could even create an LLM judge that talks about did they do did they do something together. Um, so those would be the leading and then the lagging like you already talked about is ultimately we're trying to drive retention. So, we've got the overall hypothesis, we've got the context, we've got the non- goals. These I think are like the three non-negotiables to creating a good PRD. So now I think we're ready to prototype. Do you think so? Uh yes, I do think so. I'm just one quick note on the PRD itself. Uh, I think what might be helpful is as you look into the experience, it would be great to define what a weak connection looks like as well as when you say that we're going to reach out with the right context. How do you think about the context of what's already available on LinkedIn versus your broader relationship with the person? So that might be a something that we discover and tackle later in the discussion. Okay, cool. We can discover and tackle it, but at the high level, I guess what we want to define for the PRD is a weak connection is like somebody you're firstderee connected to um but haven't messaged or commented on their content in the last 3 months. Something like that. Sure. And then context available on LinkedIn. Like obviously we're going to have like if you have any shared employers um academic institutions, locations and then engagement data I think and then I don't know if we actually look at this but things like application data if you're interested in the same field. Mhm. So those would be the main things we understand. So all right let's move into we've got like our PRD and what we wanted to do is actually see like how I prototype.

……读了这条消息、回了这条消息。我不觉得我们真的会去读用户的消息内容，但如果我们真的读，甚至可以做一个 LLM judge 来判断他们后来有没有真的一起做成什么事。这些是先导指标；滞后指标就是你刚才已经说过的——我们最终是要拉动留存。

所以我们有了整体假设、有了上下文、有了非目标。我觉得这三条是写出一份好 PRD 的三个不可妥协项。那我觉得可以开始 prototype 了，你觉得呢？

嗯，我同意。关于 PRD 我插一句：我觉得在你深入体验之前，如果能定义清楚「弱关系」到底长什么样，会很有帮助；还有你说「带着正确的上下文去联系」——你怎么看待 LinkedIn 上已经现成的上下文，跟你和这个人更广泛的关系之间的区别？这个也许我们可以在后面的讨论里再挖、再解决。

好，可以后面再挖再解决。但在 PRD 这个高层面上，我想定义的是：弱关系 = 你的一度人脉，但过去 3 个月里你没给他发过消息、也没在他的内容下互动过。大概是这个意思。

可以。

然后是 LinkedIn 上现成的上下文，显然会有：共同雇主、共同的学校、地理位置，然后我觉得还有互动数据；另外我不确定我们实际上会不会看这些，但比如求职申请数据——你们是不是对同一个领域感兴趣。

嗯。

这些就是我们能掌握的主要信息。好，那我们往下走：PRD 有了，接下来我们真正想看的是我怎么做 prototype。

[20:14]

So let's go into that. And as you can see, I spent a lot of time on the the product journey before because I actually think that prototyping for the most part. Like if I look at it when I was doing it 3 years ago versus now, a lot of the skill has actually dissolved in that the models keep getting better and better. And so the upfront definition is more what matters. So, what I'm going to do since we only have a little bit of time here is I'm going to use a couple of my favorite tools to run uh a prompt. And so, what I'm going to do is I've created a folder, just a blank folder inside of clawed code. So, what I'm going to do is I'm going to build out the context for this folder. So the first thing we need to do, let me throw you a curveball questions here. I know you're assuming that this is a blank folder, but if you had a magic wand and if you could have a harness here, what are the top three things you'd want to have in this harness so as to allow you to prototype much faster? Actually, I could even pull those if you want. I can pull my own PM operating system, but there's three things. the design system, the overall strategy and prior features, and the overall um what was the third thing I was going to say? There are three things, the design system, like our overall context and our skill and a prototyping skill. Those are the three things I'd want in here. And so, we can actually just prompt it right now. Create a context library.

那我们就进入这部分。你也看到了，我前面在产品体验走查上花了很多时间，因为我其实认为，prototype 这件事本身……如果拿我三年前做的时候和现在对比，很多技巧其实已经消解掉了——模型越来越好，所以前期的定义反而更重要。

因为我们时间不多，我接下来会用几个我最喜欢的工具来跑一个 prompt。我在 Claude Code 里建了一个文件夹，就是个空文件夹。我要做的是把这个文件夹的上下文搭起来。所以第一件事……

我先给你抛个刁钻问题。我知道你现在的前提是这是个空文件夹，但如果你有一根魔法棒，可以给自己配一个 harness（脚手架），你最想往这个 harness 里放的前三样东西是什么，好让你 prototype 快得多？

其实我可以直接把它们拉进来——我可以把我自己的 PM operating system 拉过来。不过就是三样东西：design system；整体战略和历史功能；还有整体的……我第三个想说的是什么来着？就三样：design system、我们的整体上下文和我们的 skill，再加一个 prototyping skill。这三样是我想放进来的。那我们现在就直接 prompt 它：建一个 context library。

[21:47]

Assuming you're a PM at LinkedIn using publicly stated info, create an AI prototyping skill and um create the LinkedIn design system. And for that, I'll just go ahead and give it a couple screenshots, just some of the ones we already even took in Miro, so that it I don't have to like go looking for more. So, I'm just going into Miro and dropping some of these images in here. Let me pull a couple more. So, you can't see this, but I'm over in Miro and I just grabbed another image. I didn't even place it like super carefully in the prompt because I found like especially Fable, which we're using right now, it can handle it. So, I've just put those in. So, it's this will be my harness, right? So, that's Claude code. I actually wouldn't um only use claude code. So what I found is that it can be useful to create one prompt and then use it across multiple other tools. So one of the other tools that I like to use is actually magic patterns. So let me go share that. And what I like about Magic Patterns is that it doesn't create a backend and because it's front end only, it's really fast. So what we can do is we can do something similar and here we can write out our prompt. And the easiest way to write out this prompt for a tool like Magic Patterns when it doesn't have a harness in my opinion is just to go to Claude Chat. So we'll go to Claude Chat and we'll create the prompt for this together. So, let's go into Claude chat. Here we are. Write a prompt for AI prototyping tools like lovable and bold and magic pattern. So, now it knows what we're doing. And then, how am I going to give it the context? I'm literally going to give it like this mirror that we just created together. So, I'm going to give it all of this context. If it's too small, it might not be able to read it. So, I'll like zoom in on some of this information. Then I'll come back over here. I'll paste some of these screenshots in so that

「假设你是 LinkedIn 的 PM，只用公开信息，建一个 AI prototyping skill，再把 LinkedIn 的 design system 建出来。」为此我顺手丢给它几张截图，就用我们刚才在 Miro 里已经截过的那些，这样我就不用再到处去找了。所以我切到 Miro，把这些图拖进来。我再多拉几张。你们看不到这个画面，我现在在 Miro 里，又抓了一张图。我甚至没有在 prompt 里把它们摆得特别讲究，因为我发现——尤其是我们现在用的 Fable——它是能处理好的。所以我就这么放进去了。这就是我的 harness。

这是 Claude Code。我其实不会只用 Claude Code。我发现一个做法很有用：写一份 prompt，然后拿它跨多个工具去跑。我喜欢用的另一个工具是 Magic Patterns。我分享一下。我喜欢 Magic Patterns 的地方是它不建后端，因为只做前端，所以特别快。我们可以在这里做类似的事，把 prompt 写出来。

而对 Magic Patterns 这种没有 harness 的工具，我认为写这份 prompt 最省事的办法，就是去 Claude Chat。所以我们去 Claude Chat，一起把这个 prompt 生成出来。进 Claude Chat，来了：「给 Lovable、Bolt、Magic Patterns 这类 AI prototyping 工具写一个 prompt。」这样它就知道我们在干嘛了。

那我怎么给它上下文？我就直接把我们刚才一起做的这块 Miro 白板给它。所以我要把这些上下文全给它。如果画面太小它可能读不清，所以我会把其中一些信息放大一点。然后我再切回来，把这些截图粘进去，好让它……

[24:07]

it really understands. Okay, this is um the task. And then I'm just going to give it a little more, right? Uh we want to generate diverse solutions that surface um weak connections with contextual messages that can help the user generate economic opportunity. Some of the surfaces will live on. So we kind of identified three, right? the homepage, the grow your network page, and profiles of your firstderee connections. And so now I'll say like write a PRD. The idea is that the AI prototyping tool can actually um brainstorm like different diverse solutions to. So that's how I would approach this problem. And then what I'd do is I'd honestly open up another tool too. So my third favorite tool for this kind of work is lovable. So now we've got a prompt being built in claude. We've logged into magic patterns. We've got a harness being built inside of claude code and we got lovable. So that'll be like our third sort of tool for this parallel prototyping workflow. So let's go into the first tool that we were working on which is cloud code and let's go see how cloud code is doing. So, it's created all of this for us. I can just double check because I always like to sanity check this. How did it do with the L this design system? Pretty good. How did it do with the context library? Pretty good. So, now I'm ready to hit it with the prompt. I think for the context library, it looks like it needs a little bit more information now that I see this. Like in this product section, it doesn't have all the context we just talked about. So one of the things that will help this context library is giving it those mirros. So I will give that in a second. But first let's go take a look at the claude prompt that we had created. So generate five distinct UI concepts for a feature that helps professionals reconnect with. Perfect. Perfect.

……真正看懂。好，这就是任务了。然后我再多给它一点：我们想生成多样化的方案，把弱关系连同带上下文的消息一起浮现出来，帮用户创造经济机会。这些方案会落在哪些界面上——我们大概识别出三个：首页、Grow your network（拓展人脉）页，以及你一度人脉的个人主页。然后我会说：写一份 PRD。思路是让 AI prototyping 工具真的能去头脑风暴出各种不同的方案。这就是我处理这个问题的方式。

接下来我还会老实说再开一个工具。我做这类活儿第三喜欢的工具是 Lovable。所以现在我们手上有：在 Claude 里正在生成的 prompt、已经登录好的 Magic Patterns、在 Claude Code 里正在搭的 harness，再加上 Lovable。Lovable 就是这套并行 prototyping 工作流里的第三个工具。

我们回到第一个工具 Claude Code，看看它跑得怎么样。它已经把这些都建出来了。我可以核对一下，因为我总喜欢做个 sanity check：design system 做得怎么样？挺好。context library 做得怎么样？挺好。那我现在可以把 prompt 甩给它了。

不过 context library 我现在看下来，好像还需要更多信息。比如这个 product 部分，就没有我们刚才聊到的那些上下文。所以能帮到这个 context library 的一件事，就是把那些 Miro 白板喂给它。这个我等下就给。但先去看看我们刚才在 Claude 里生成的那份 prompt。「为一个帮助职场人重新连接……的功能，生成五个不同的 UI 概念。」完美，完美。

[26:26]

Looks great. I'm loving this. Right. Can you Akash? Can you talk for a few seconds about what makes this a strong prompt? Yeah. really go into the details, but would love to get your thoughts on why what a what distinguishes a good prompt for prototyping from a not great prompt. It's actually the same as any prompt if I'm honest. There's like four things roughly. Number one, giving it a very clear definition of the steps of the task it should pursue. And so that's why it starts with generate five distinct UI concepts. Number two, defining your context. has really good definitions of the context. Number three, showing what you want the output to be. So, it has really good requirements and it has the output section. Number four, saying what you don't want. And so, I think it actually has all four elements of a pretty good prompt. So, I'm pretty happy with that. Sonnet did a decent job with this. So, then I'm ready to go into my tools. And what I'm going to do is hit it with this. Do you not want to give it some design guidance on this being for LinkedIn? So, in the prompt itself, does it say that if it it should have it says right at the top. My bad. Yeah, it has some information, but you know, it could be better. Honestly, you're right. So, let's go ahead and stop these prompts and let's let's iterate. A lot of times we do need to iterate on the prompt.

看着不错，我很喜欢。对了 Aakash，你能不能花几秒钟讲讲，是什么让这成为一个强 prompt？不用讲得太细，但很想听听你的看法：一个好的 prototyping prompt 和一个不怎么样的 prompt，区别到底在哪儿？

说实话，跟任何 prompt 都一样，大概就四件事。第一，给它一个非常清晰的任务步骤定义，所以它开头就是「生成五个不同的 UI 概念」。第二，定义好你的上下文——它对上下文的定义写得很到位。第三，说清楚你想要的输出是什么，所以它有很到位的 requirements，还有 output 部分。第四，说清楚你不想要什么。所以我觉得它其实四个要素都齐了，我挺满意的。Sonnet 这活儿干得不错。

那我准备把这个丢进我的工具里了。

你不想在这上面给它一些 LinkedIn 的设计指引吗？

prompt 里面……它是不是已经写了？它在最前面就写了。是我看错了。对，是有一些信息，但确实可以更好。老实说你是对的。那我们就把这几个 prompt 停下来，来迭代一下。很多时候我们确实需要迭代 prompt。

[27:58]

Let's iterate on the prompt to um include a I mean for what it's worth you already got the design system in cloud code right Akash so yes exactly but for lovable and bolt we'll probably need to and then I'm also thinking about this and I don't know if I really want the prototyping tool to generate the five diverse solutions I want um I don't even want sonnet to generate the five diverse solutions I want something like fable to generate generate the five- diverse solutions. Um, five diverse solutions now and we'll prototype one in each of the tools. So, I'm going to go ahead to go to a more powerful model now and do that brainstorming step of these diverse solutions here. You could do it in a prototyping tool. It kind of depends like people designers talk about are we code first or canvas first. You could go either way, like if you want to see it, but I kind of trust Fable more. So, I don't want to give too much away to those tools that they don't even sometimes show me which model they have underneath it. Yeah. So, while that's happening, let's go back into Claude Code. This is where we actually had um the design system already in place and so we can use the prompt that we originally got. And so let's give it that original prompt and that'll get to work. And now it becomes a little bit of a waiting um for Claude. And so that's why I like to have parallel systems. Anyways, we can go now into this prompt. It's starting to work on that for us.

我们来迭代这个 prompt，加上……

不过说实在的，Claude Code 里你已经有 design system 了嘛，Aakash？

对，没错，但 Lovable 和 Bolt 那边我们大概还是得给。另外我还在想一件事：我不确定我真的想让 prototyping 工具去生成这五个多样化方案——我甚至不想让 Sonnet 来生成这五个方案，我想让 Fable 这种模型来生成这五个多样化方案。先生成五个方案，然后我们在每个工具里各做一个 prototype。

所以我现在要换一个更强的模型，去做「多样化方案头脑风暴」这一步。你也可以在 prototyping 工具里做，这有点看情况——设计师会讨论我们到底是 code first 还是 canvas first，两条路都行，如果你想直接看到东西的话。但我更信 Fable 一些。而且我也不想把太多东西喂给那些工具，它们有时候连底下用的是哪个模型都不告诉我。

对。那趁它在跑，我们回到 Claude Code。这边 design system 已经就位了，所以我们可以用原来那份 prompt。把原始 prompt 给它，它就开工了。

接下来就有点变成等 Claude 了，所以我才喜欢搞并行系统。总之，我们现在可以进到这个 prompt 里，它已经开始给我们干活了。

[29:35]

Um, and what we can do, one of the workflows that I really like since that prompt is taking some time is to have it define the design system first before you give it any more information. So, what do I mean here? Learn the design system of LinkedIn. Simple as that. And then we can go into here. We can give it that one screenshot that we had created. We can put that screenshot into here. So we can hit it with that kind of a prompt. And I'll just give it the exact same prompt here in magic patterns. And so now we have three AI agents kind of working for us. Each of them is going to take some time. So let's go back over to Claude, which was working on our five diverse solutions. So the five concepts it created, it didn't format these very well, but let's just keep moving forward for the sake of time here because I know we're running out of time. So, in the homepage feed, a warm signal guard. Perfect. That's the type of thing I want to prototype. Um, a dedicated tab, the grow your work network like an inbox. Also cool. The user's firstderee network rendered as a visual map. Yep. Also very cool. So against these, what I would want do is like let's quickly just apply a framework to think about these. How innovative is this? How much is this really going to drive success? Does that surface get enough area? Will this drive our ultimate goal? So, let's just look at some of these three against that. Immediately, we know that the heat map is going to be a lot more innovative than the reconnect Q, but it's going to roughly rate the same on the other two variables. Warm surfacing is going to get a lot of volume. So, it's really going to drive the outcome, but it's not very innovative. a profile context layer. This just seems like it needs to be a part of most of what we're doing. So really, it can go in harmony with whatever we create between one and three. And then a homepage notification, this can

嗯，趁着那个 prompt 还在跑，我们可以做一件我特别喜欢的事——这是我很爱的一个工作流：先让它把 design system 定义出来，之后再喂给它更多信息。什么意思呢？就是一句「学习 LinkedIn 的 design system」，就这么简单。然后我们进到这里，把刚才生成的那张截图给它，把截图放进来，用这样一个 prompt 打过去。同样这套 prompt 我也原样丢给 magic patterns。所以现在等于有三个 AI agent 同时在给我们干活，每个都得跑一会儿。那我们先回到 Claude 这边，它刚才在做我们要的五个差异化方案。它生成的这五个概念，格式排得不太好看，但时间紧，我们先往下走，因为我知道时间快不够了。第一个：在 homepage feed 里做一个「温度信号」守门层。完美，这正是我想拿去做 prototype 的那类东西。第二个：一个独立的 tab，「拓展你的职场人脉」，做成收件箱那样。也挺酷。第三个：把用户的一度人脉渲染成一张可视化关系图。嗯，这个也很酷。针对这几个，我想做的是快速套一个框架来评估：这个有多创新？它对成功的推动力有多大？这个入口界面拿得到足够的曝光面吗？它能不能推动我们的终极目标？那我们就拿其中这三个对着这套标准过一遍。一眼就能看出来，热力图肯定比 reconnect 队列创新得多，但在另外两个维度上大致打平。「温度前置露出」的量会很大，所以它对结果的拉动很强，但不够创新。至于 profile 上下文层，这个感觉本来就该是我们大部分方案里的一部分，所以它其实能跟一号到三号里做出来的任何东西共存配合。然后 homepage 通知，这个也能……

[31:43]

also go in harmony. So four and five are surfaces for a feature. So really what we're thinking about is like one and three. And this is the most innovative. So really what I want to do is I want to do like a combo of concepts. Three through five is what we want to do. create the final prompt for lovable magic patterns cloud code. And so now we should have like a decent solution that we can work with. And I know we're like almost at time here. So what's going to happen next? We're going to take this prompt that comes out. Take like 30 seconds. We're going to paste it into the three. We're going to see which one we like the most. That sound good to you? Sounds great. Yes. Okay. While we're waiting for this, um I know of course you've shared the key tools you'd use if you're focusing on front-end design. If you also had to do a back-end design for the Sakash, could you talk a little bit about what tools you would use and how you would go about that? I am a sucker for an IDE plus either codeex, Gemini, or Claude model. So, we were looking at cursor plus cloud code. I usually what I'll do if I'm in magic patterns for instance magic patterns allows you to just like export or lovable has the same lovable you can also just publish to GitHub so what I do is with lovable I publish to GitHub and magic patterns I export I take it usually into I'm a claude fan over Codeex or Gemini for 90% of cases the only 10% I really use codeex for is very long running coding tasks like when I'm actually coding a backend so for the front end work, I tend to prefer the taste of Fable 5. So, I would be using Fable 5 for most of that. And then if I had some long backend tasks, I might use Codeex for it. One, because it's a little bit cheaper than Fable 5, and two, because I find it's much better. It's like a workhorse. Like, it'll work for like a whole day. Fable wants to work for like an hour. Lovely. Thank you. So, [snorts] let's go

……也能共存配合。所以四号和五号其实是承载某个功能的入口界面。那我们真正要在里面挑的，是一号和三号。而这个是最创新的。所以我真正想做的是把几个概念组合起来，三号到五号组合起来，这就是我们要的。「生成给 lovable、magic patterns、Claude Code 用的最终 prompt。」这样我们应该就能拿到一个还算像样、可以往下推的方案了。我知道我们基本已经到点了。那接下来会发生什么？我们把跑出来的这个 prompt 拿走，花个 30 秒，粘到那三个工具里，看看我们最喜欢哪个出来的效果。这样可以吗？——太好了，可以。——好。趁等结果的功夫，我知道你已经讲了做前端设计会用的关键工具。如果 Sakash 这个项目你还得做后端设计，能不能稍微讲讲你会用什么工具、怎么下手？——我对「IDE 加上 Codex、Gemini 或者 Claude 模型」这套组合毫无抵抗力。刚才我们看的就是 cursor 加 Claude Code。我一般的做法是，比如我人在 magic patterns 里，magic patterns 支持你直接导出；lovable 也一样，lovable 可以直接 publish 到 GitHub。所以 lovable 我就 publish 到 GitHub，magic patterns 我就导出，然后拿进来。90% 的情况我是 Claude 党，胜过 Codex 和 Gemini。剩下那 10% 我真正会用 Codex 的场景，是超长时间跑的编码任务，比如我真的在写后端的时候。所以前端这块，我更偏爱 Fable 5 的审美，大部分我都会用 Fable 5。然后如果有长时间的后端任务，我可能会上 Codex：一是它比 Fable 5 便宜一点，二是我觉得它在这块强得多，它就像一头老黄牛——能连着干一整天，而 Fable 大概只想干一个小时。——太好了，谢谢。那么，[笑] 我们进到……

[33:42]

into this final prototyping prompt. I'm excited to like finally get there. Yes, we got it. So, I agree with this prompt. So now I'm going to go into magic patterns. Yes. It had a question about whether it should scaffold the system. I said yes. This did too. I said yes. And then I hit it with the prompts. Now let's go look at cloud code. So if you recall with claude code, we didn't uh we actually gave it a little more autonomy. And the reason we did that because we can use fable to do it. And so cloud code came up with its own different concepts. It looks like it came up with five different concepts that it's working on and it is developing these HTMLs for them. As I can see it, although I'm not positive, it looks like the first four concepts might be ready. So, let's go into concept one cuz that's most likely to be fully ready. And let's go see here. So, this would be warm threads. So, if we open this in browser theoretically, nice. I'm loving this. Yeah, it did a pretty good job of keeping true to the design system and of uh pulling in the right context. Yep. So, I'm pretty happy with this. Like what I'm saying pretty happy is like this is in the design system. It's a real feature that will work. Now, we're going to go evaluate the other five concepts and decide which concept we like the most and we can iterate further on it. So, we can then open the next one. Reconnect in Q. Nice. worth reaching out to this week. Write with context exactly the problem we were talking about. Home moments. Okay, the moments one, I understand what it's doing here. I feel like this is almost like a P2 or P3 feature, right? Like those other ones, they're going to be hitting much more common surfaces and they're going to be happening more often for more people. So, I think I like those a little bit better. Reconnecting with we Okay, this is just a summary of the concepts that it created. We don't need that. All right, here's what a draft

……最终这个 prototyping prompt。我很期待终于走到这一步。好，拿到了。这个 prompt 我认可。那我现在进 magic patterns。好。它问要不要先把系统骨架（scaffold）搭起来，我说要。这个也问了，我也说要。然后把 prompt 打过去。现在我们去看 Claude Code。你还记得的话，Claude Code 这边我们其实给了它更大的自主权，之所以这么做，是因为我们可以用 Fable 来跑。所以 Claude Code 自己想出了另一批不同的概念。看起来它自己搞了五个概念在做，正在给每个概念写 HTML。我看着——虽然不能百分百确定——前四个概念可能已经好了。那我们先看概念一，因为它最可能已经完全跑完。我们来看看。这个叫 Warm Threads（温度会话）。理论上我们在浏览器里打开它……漂亮，我很喜欢这个。——是的，它在忠实还原 design system、把正确的上下文拉进来这两件事上都做得挺不错。——嗯。所以我对这个挺满意。我说「挺满意」的意思是：它在 design system 之内，而且是一个真的能跑通的功能。接下来我们要去看其余几个概念，决定最喜欢哪一个，然后在那个上面继续迭代。那我们打开下一个：Reconnect Queue（重连队列）。不错。「这周值得联系的人」。「带上下文起草」——正是我们刚才在聊的那个问题。Home Moments（首页时刻）。好，moments 这个我明白它在干什么。我感觉这差不多是个 P2 或 P3 的功能，对吧？前面那几个触达的是更常见的入口界面，发生得更频繁、覆盖的人更多。所以我更喜欢那几个一点。「和……重新连接」。好，这个只是它生成的概念汇总，不需要。好，接下来看看一份草稿……

[35:47]

will look like. I think this is good. So, I like the draft. It's actually using the context like you were talking about in a strong way. So, that's great. And then the last one, constellation. It may not be ready yet. Seems like it's writing it, but it looks good enough. Mhm. Hm. So, this is so interesting, right? When we were talking about this in chat, I thought the constellation would be like the killer feature. Now, I see it. I'm like, huh, [laughter] maybe not. So actually based on everything I'm seeing right warm threads is really really really powerful and reconnect Q is super super powerful and then drafts first. So those three to me seem like the ones that based on my product sense and product taste those would be the most powerful. And I know we're running out of time now. So let me just summarize what I'd do from here. I'd put these all together. I'd keep it at front end and then I would user test this a little bit with myself with my own team. Then I'd take it to a system like user testing or user voice where you can literally pay people to interact with a front end and I'd get their feedback. Once I got the positive signal from them, I'd go into the back end and I'd start to create this as a real live prototype, connect it into our GitHub, pull from our actual codebase, our actual design system, and then from there, I think we might be ready to do like a 0.1% AB test or something. That sounds great. Um, well, thank you so much. I know we are at time, so I really appreciate you going through the flow, Kash, and for taking all the curveball questions. Um, how are you feeling about it? We good. I wish we could have seen Magic Patterns and Lovable finish up. And so I'm glad that you pushed me on the time. Like the big challenge here is always getting through everything in 30 minutes. But I think if we imagine a world where um we are surfacing up warm connections to you in

……长什么样。我觉得这个不错，草稿这块我喜欢。它确实像你说的那样，把上下文用得很到位，这很棒。然后最后一个：Constellation（星座图）。它可能还没跑完，看起来还在写，但已经够看了。嗯……哈。这就太有意思了，对吧？我们之前在聊天里讨论的时候，我以为 constellation 会是那个杀手级功能。现在真看到了，我心想：呃，[笑] 也许不是。所以基于我现在看到的一切，Warm Threads 真的非常非常强，Reconnect Queue 也超级强，然后是 Drafts。按我的 product sense 和 product taste，这三个在我看来是最有力量的。我知道我们时间到了，那我简单总结一下接下来我会怎么做。我会把这些整合到一起，先只停在前端，然后我会自己、连同我自己的团队先做一轮用户测试。之后我会把它放到 UserTesting 或 UserVoice 这类平台上——在那儿你可以直接付钱找人来跟一个前端交互——收集他们的反馈。一旦从他们那儿拿到正向信号，我就会去做后端，把它做成一个真正能跑的活 prototype，接进我们的 GitHub，从我们真实的代码库、真实的 design system 里拉东西。到那时候我觉得我们可能就可以做个 0.1% 的 AB test 之类的了。——听起来很棒。嗯，非常感谢你。我知道时间到了，真的很感谢你把整个流程走了一遍，Kash，也感谢你接下了所有的刁钻问题。嗯，你自己感觉怎么样？还行吗？——我有点遗憾没能看到 Magic Patterns 和 Lovable 跑完。所以我很高兴你一直在时间上推我，这里最大的挑战永远是怎么在 30 分钟里把所有东西走完。但我觉得，如果我们设想这样一个世界：我们把有温度的人脉推到你的……

[37:40]

your home feed and in your grow your network tab with in context messages, we will for sure make a significant dent in leading metrics like how many messages are sent and responded to and hopefully lagging metrics like retention. Yeah, it was uh very nice to see Fable just pull both the the importance of the direction that we had given it in terms of warmth and cold outreach as well as stick to the design system so well and the fact that you established a harness right out of the gate which then was able to pull in the design system I think was quite crucial to that. Yeah. So I think overall I'd give myself like a 8 out of 10. I think you rightly pushed me and had to push me on let's get to the actual prototyping and how do you prototype well and so I was doing a too deep a job on the user and problem. So I think if I were to do it over I would go even a little bit faster on user and problem so that I could just have a little bit more time and we could see the output. you were doing a great job from a traditional product sense standpoint, right? Like I would have loved to see that level of depth and thoughtfulness in a traditional product sense interview. What's more critical in a prototyping interview is you get that out of the way quickly so as to then showcase your skills with the prototyping part. And the trick is having attained the depth of thinking in a shorter time span so as to have your prototype be deep and thoughtful not just surface level. Right? Like that is the key trade-off and challenge that you face that if you don't spend sufficient time and thoughtfulness in the first part then your prototype is not going to be great. Right? The interviewer is not there just to see your process. They're also assessing the outcome that you land at. And I think you were able to accomplish both of them quite well.

……首页 feed 里，推到「拓展人脉」那个 tab 里，还配上带上下文的消息，那我们一定能在先行指标上——比如发出去多少条消息、多少条得到回复——打出明显的增量，也希望能带动滞后指标，比如留存。——对，看到 Fable 既抓住了我们给的方向重点（也就是「温度」相对「冷启动式外联」的区别），又把 design system 守得这么好，感觉非常棒。而且你一上来就先立了一个 harness，正是这个 harness 之后才能把 design system 拉进来，我觉得这一步相当关键。——嗯。所以总体我给自己打 8 分（满分 10 分）吧。我觉得你推我推得对，你不得不推我说「我们赶紧进到真正的 prototyping，以及怎么把 prototype 做好」，所以我在用户和问题那一段挖得太深了。如果重来一次，我在用户和问题上会再快一点，这样就能多留点时间，我们也能看到最终产出。——从传统的 product sense 角度看，你做得非常好，对吧？这种深度和思考的细致程度，我在一场传统的 product sense 面试里会非常乐意看到。但在 prototyping 面试里，更关键的是你得快速把这部分过掉，好腾出时间展示你在 prototyping 部分的本事。而诀窍就在于：在更短的时间里达到同样的思考深度，从而让你的 prototype 也是有深度、有想法的，而不只是浮在表面。对吧？这就是你面对的核心取舍和挑战——如果你在前半段没有投入足够的时间和思考，你的 prototype 就不会出彩。对吧？面试官不是只来看你的过程的，他们同时也在评估你最后落到的结果。而我觉得这两点你都完成得相当好。

[39:34]

So would this be a passing interview, do you think? Oh, this would definitely be. Okay. So, let me go through. I have my notes right here. So, I would say overall uh you you probably underrated yourself a little bit. I'd put you between an 8.5 and a nine. A very clear and confidently uh sorry, a very clear and confident pass for me. and it was genuinely just great to watch you through and tackle all the curveball questions. Let's actually break this down into the different buckets that you use because the framework that you use Akash is a great one for approaching any prototyping interviews. Um so I'm going to just use that. So on user I think you uh probably put you at again an 8.5 to 9. You segmented really well uh everything from network engagers, job searchers, etc. and putting numbers on each of the buckets uh was super helpful. I think that is generally a strong instinct and the way you were able to connect it back to the mission again really anchors you in the user being very critical to the company. It was crisp, it was correct. What would have been great to see was probably if you would have done a little bit of um a mid exploration on how do you keep the professional profile bucket after I redirected you, right? Um the second part is problem uh in your framework. And this is where I'd probably put a an 8 out of 10 because I had to push you a little bit to to get to the core underlying problem of hurdle, cognitive load, etc.

——那你觉得，这算是一场能通过的面试吗？——哦，这绝对能过。——好。——那我来过一遍，我笔记就在手边。我会说，总体上你可能把自己评低了一点，我会给你 8.5 到 9 分之间。对我来说，这是非常清楚、非常有把握的一个 pass。而且看着你一路走下来、把所有刁钻问题都接住，真的挺过瘾。我们不如就按你用的那几个模块来拆开讲，因为 Akash 你用的这套框架，拿去应对任何 prototyping 面试都很好用，那我就直接用它。

先说用户（user）。这块我大概同样给你 8.5 到 9 分。你的分层做得非常好——从「人脉活跃者」到「求职者」等等，而且给每一层都配上了数字，这特别有帮助。我觉得这总体上是很强的直觉；而且你能把它一路连回公司的使命，这就把「用户」牢牢锚定成对公司至关重要的东西。表达干净利落，判断也对。那还差点什么？大概是在我把你往回带之后，你可以再补一小段探索，讲讲「职业档案（professional profile）」那一层要怎么保住。

第二块是你框架里的问题（problem）。这里我大概给 8 分，因为我得推你一把，你才走到那个底层的核心问题——门槛、认知负荷等等——而且是更……

[41:10]

faster and uh the live product audit I would say though was the highlight of the the section and that's probably better than 95% of the candidates who going to theorize the the entire experience rather than actually opening up the app. I think you found the empty catch-up state, the bizarre notification coupling. All of those were very real problems that the interviewer is going to be like, "Ah, yes, I see why this doesn't work today." And potentially then leads you very clearly into laying out potential solutions and ideas. And that is exactly what you did, right? Uh then on the solution, as I said, you uh you landed pretty well. Once the concepts rendered, you were able to generate very clear five directional ideas. And of course, we'll talk about that, but I'd probably again put you in the 8.5 to 9 bucket. To me, the game changer here would have been more specificity and more opinion on the solution itself. But given the time constraint, uh I like I'm not going to ding you against it. If you were in a product sense interview where you had a lot more time, that is where the taste and opinion piece plays a much bigger role. Then on the PRD, I think you hit the non-negotiables pretty cleanly. You had a very clear, sharp set of goals, a set of non-goals. You had a very clear leading and lagging metric split, which was pretty awesome. So, I think you you did a pretty phenomenal job here. Again, a solid 8 to 8.5. Um, on the prompt, I like even though your prompt was generated using AI, I don't think that that comes naturally to a lot of people. So, despite the fact that Fable or Sonnet generated your prompt, I'm still going to give you a 9 to 9.5. And the part that you were able to talk about the good parts or like what constitutes a good prompt really stood out to me. Right. So it's not that you did not write the prompt well, but if you were given the time and if you had the just head space to write the prompt,

……更快一点走到。不过我要说，现场做的 live product audit（真机产品体检）是这一段的亮点，这一点大概比 95% 的候选人都强——那些人只会对整个体验做理论推演，而不是真的把 App 打开。我觉得你找到的那些点——空的「catch up」状态、那个诡异的通知耦合——全都是非常真实的问题，面试官看到会想：「啊，对，我明白它今天为什么不成立了。」而且这些还能非常顺畅地把你带进后面列出候选方案和想法的环节，你做的正是这个，对吧。

然后是解决方案（solution）。像我刚说的，你落得挺稳。概念渲染出来之后，你能给出五个方向非常清晰的想法。当然我们待会儿还会聊到，但这块我大概还是给你 8.5 到 9 分。对我来说，这里真正能拉开差距的，是对方案本身有更多具体细节、更强的观点。但考虑到时间限制，我不会拿这个扣你分。如果是在时间宽裕得多的 product sense 面试里，那「品味和观点」这一块的权重就会大得多。

再来是 PRD。我觉得那些不可妥协项你都干净利落地打到了：你有一组非常清晰、锋利的目标，还有一组非目标（non-goals）；你还把先行指标和滞后指标清清楚楚地分开了，这个相当漂亮。所以我觉得你这块做得相当出色，同样是稳稳的 8 到 8.5 分。

嗯，接下来是 prompt。虽然你的 prompt 是用 AI 生成的，但我觉得这个做法对很多人来说并不是自然就能想到的。所以尽管你的 prompt 是 Fable 或者 Sonnet 写的，我还是要给你 9 到 9.5 分。而且你能讲清楚它好在哪、什么才算一个好 prompt，这一点让我印象特别深。对吧，所以不是说你 prompt 写得不行，而是如果给你时间、给你足够的脑力空间去写这个 prompt，……

[43:19]

that makes me confident that you would be able to write a great prompt. Right? And watching you use AI to write the prompt was sort of the cherry on on top. Now on paradle which is one of the very unique aspects of your framework that is the part most people are going to underestimate or not leverage well that in a very short burst of time you leveraged magic patterns. You leveraged Bolt Cloud Code and you quickly talked about codeex and a couple of other tools and the way you were walking through it, it made me confident that you had a good idea of understanding what to use when and even how to use it, right? You didn't use cloud code for let's say idea X, you used it for specifically first building the harness and then jumping into blah. you used a very specific tool for generating your idea graph and so on and so forth. And I think that working in parallel with a pretty broad set of tools shows the AI nativeness that most interviewers uh are sorry most interviewees and interviewers are looking to solve for. And then finally, I think back end. Um, we didn't get to spend a whole lot of time on the back end, but again, talking about codeex and how you would go about completing the rest of the journey to me was a confident enough signal. I'd give you an eight on 10. If you could have gotten a little bit more into the backend design, if we had time, I'd probably given you a 9 to 9.5, but again, very solid job on that. So overall, I think it's a it's a solid 8.5 to 9. Very very solid score for uh an inproduct evolution of an idea. To close this off, what I would say is most people underestimate how hard it is to iterate on an existing and mature product than come up with a new idea. It's pretty easy to say, okay, build product X for a completely new set of users.

……我有信心你能写出一个很棒的 prompt。对吧？而看着你用 AI 来写 prompt，算是锦上添花。

接下来是并行（parallel）这一块，这是你框架里非常独特的一环，也是大多数人会低估、或者用不好的部分。在很短的一阵时间里，你同时调动了 magic patterns，调动了 Bolt、Claude Code，还顺带讲了 Codex 和另外几个工具。而你走这一遍的方式，让我确信你很清楚什么时候该用什么，甚至清楚该怎么用，对吧？你没有拿 Claude Code 去做比方说想法 X，你是专门用它先搭 harness，然后再跳到下一步；你用一个非常特定的工具去生成你的想法图谱，诸如此类。我觉得这种「同时并行操作一套相当宽的工具」的能力，正体现出大多数面试官——抱歉，是大多数面试者和面试官——想要考察的那种 AI 原生度。

最后是后端（back end）。嗯，我们没能在后端上花太多时间，但同样，你讲 Codex、讲你会怎么走完剩下那段旅程，对我来说已经是足够有把握的信号了。我给你 8 分（满分 10 分）。如果时间够、你能再往后端设计里多讲一些，我大概会给到 9 到 9.5 分，但同样，这块完成得非常扎实。

所以总体上，我认为这是稳稳的 8.5 到 9 分。对于「在一个已有产品里演进一个想法」这件事来说，这是非常非常扎实的分数。最后收个尾我想说：大多数人都低估了在一个成熟的存量产品上做迭代有多难——比想出一个全新的点子难多了。说一句「好，给一批全新的用户造一个产品 X」，那是相当容易的。

[45:28]

But by asking you to evolve an existing product and its features, there are already constraints and assumptions in place that are often not surfaced or taken into consideration. Um, and that makes it much more harder to play into. So kudos on on cracking that well. Thank you. So let me put together for you guys what we just showed in a visual. So I wrote this piece for you all and this is kind of what we teach in the cohort Anka and I about the vibe coding interview. This is normally for paid subscribers. So you guys are getting a little a bit of alpha here in that we talk about this UPS PPPB framework and that's what I showed today. As you can see you have to adjust this for the time. Sometimes backend won't get 10 to 15 minutes. Sometimes these need to get less time. So think about your own interviewer and when your interviewer gives you a solution like uncut basically gave me the solution be willing to move forward. I think that was one of the things that this interview is still evolving. There's no clear time frame of spend this amount this amount. As you can see the way this interview evolved by the way I hadn't seen the question before and I didn't agree on any of the curve balls. It was all live. You have to handle it as it comes for you and you need to look at the timer. So I would say practice your timing. My even mine could have been better.

但当面试官让你去演进一个已有的产品和它的功能时，其实已经存在一堆约束和假设了，而这些东西往往没被摆到台面上，也没人特意提出来。所以这种题反而更难接得住。你能处理得这么好，值得点赞。谢谢。那我把我们刚才演示的内容给大家用一张图整理一下。我为大家写了这篇东西，这基本上就是我和 Anka 在 cohort 里讲 vibe coding 面试时教的内容。这些正常只对付费订阅者开放，所以你们今天算是拿到了一点内部料——我们讲的这个 UPS PPPB 框架，也就是我今天演示的那套。你能看到，这套框架你得按时间去调整。有时候后端拿不到 10 到 15 分钟，有时候某些环节就得压缩时间。所以要想清楚你自己的面试官是什么风格：当你的面试官像刚才那样，基本上直接把解法递给你的时候，你要愿意顺着往下走。我觉得有一点是，这种面试本身还在演化中，并没有一个明确的时间分配标准，说这一段该花多少、那一段该花多少。你也看到了这场面试是怎么一路演化的——顺便说一句，这道题我事先没见过，那些临时抛过来的变化球我也没有提前对过，全是现场的。你必须见招拆招，同时还得盯着计时器。所以我会说，练一练你的时间把控。连我自己这次都还可以做得更好。

[46:49]

Practice your confidence with different workflows, different tools. That was one of the things like most about what I did. and then practice how you explain crisply these are the problems, these are the solutions, this is the PRD. Once you put that all together, you'll be ready to succeed at this interview. Thank you again, Akash. All right, guys. See you in the next episode. See you.

多用不同的工作流、不同的工具去练，把手感和自信练出来。这也是我今天做的事情里我自己最满意的一点。然后练一练怎么把话说得干净利落：这些是问题，这些是解法，这就是 PRD。当你把这一整套都串起来，你就已经做好准备拿下这场面试了。再次感谢你，Akash。好了各位，我们下期节目见。拜拜。