Zapier's CEO Will Grade Your AI PM Skills Live

频道: Aakash Gupta
视频: https://www.youtube.com/watch?v=WbvJMnlB6wA
原文语言: en
统计: 共 110 轮 · Aakash 48 · Wade 46

[0:00]

Every PM is about to get graded on AI fluencies. Zapier, Shopify, and many others already are. If AI was important in 2026, it's just the standard expectation for every job.

Aakash： 每个 PM 都即将被按 AI 流利度（AI fluency）打分。Zapier、Shopify 还有很多公司已经在这么做了。

Wade： 如果说 AI 在 2026 年很重要，那它已经是每一份工作的标准要求了。

[0:10]

Wade Foster is the CEO and co-founder of Zapier. He wrote the viral AI fluency rubric that [music] arguably kicked off this trend. There is still something to be said about really understanding the customer problem, really having some intuition around how to go solve those problems. And if you combine those, you get excellence. But if you just let AI kind of just do everything, you're probably going to get mediocrity. You're going to get what people call AI slop.

Aakash： Wade Foster 是 Zapier 的 CEO 兼联合创始人。他写的那份爆火的 AI 流利度评分表，可以说正是这股风潮的起点。

Wade： 真正理解客户的问题、对怎么解决这些问题有自己的直觉，这些依然很有价值。把它们和 AI 结合起来，你就能做到卓越。但如果你干脆让 AI 包办一切，你大概率只会得到平庸的东西，也就是大家说的 AI slop（AI 糊弄出来的垃圾内容）。

[0:34] Aakash

What was capable last year that's now unacceptable? What does a PM that's transformative actually do differently on a Tuesday?

哪些去年还算「胜任（capable）」的做法，现在已经「不合格（unacceptable）」了？一个「变革型（transformative）」的 PM，在一个普普通通的周二，做的事情到底有什么不一样？

[0:42] Wade

The line we use at Zapier on slop. You can delegate the work, but you can't delegate the accountability.

关于 slop，我们在 Zapier 常说一句话：活可以委派出去，责任委派不出去。

[0:46]

Looking at Air Table [music] now, it's less than 10% of that amount for you at Zapier. You've been around since 2012. Where does Zapier's business sit now? One thing that's unique about Zapier is the valuation thing doesn't matter as much to us.

Aakash： 再看看 Airtable，现在只剩那个数的不到 10% 了。说回你们 Zapier，你们从 2012 年就在做了，Zapier 的业务现在处在什么位置？

Wade： Zapier 有一点很特别：估值这件事对我们来说没那么重要。

[1:00] Aakash

So if you saw me running a product team like this, how would you grade my output?

那如果你看到我这样带一个产品团队，你会怎么给我的产出打分？

[1:04]

So if I'm spotting things like are obvious to me and not to you, I'm like, "Ooh, that makes me nervous, right?" and and I'm like, 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. [music] Check out bundle.ac.com at 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 I've been trying to figure out what AI fluency actually means for PMs inside of companies how do product leaders how do CEOs measure you and I could think of no better person than Wade Wade published what I think was the first AI fluency rubric online many companies modeled it since Then Zapier has updated their thinking. How did they update their thinking? What actually matters day-to-day? How is Wade actually grading employees? In today's episode, he's not just going to walk through his rubric theoretically. We'll actually show you some AI behaviors. And Wade will give his take on is this transformative, is this unacceptable.

Wade： 所以如果我发现一些东西对我来说显而易见、对你却不是，我就会想：「哦，这让我有点紧张了，对吧？」然后我就……

Aakash： 在进入今天的节目之前，请花一秒钟确认你已经在 YouTube 上订阅了频道，并在 Apple 播客和 Spotify 上关注了本节目。如果你想用上我最喜欢的所有 AI 工具，我已经说服它们送你一整年的付费版。去 bundle.aakashg.com 看看，可以拿到一整年的 bolt.new、Airtable、Speechify、Descript、Magic Patterns、Linear、Dovetail、Arize 和 Mobbin。好，下面进入今天的节目。我一直在琢磨，AI 流利度对公司里的 PM 来说到底意味着什么？产品负责人、CEO 们是怎么衡量你的？我想不出比 Wade 更合适的人了。Wade 发布了我认为是网上第一份 AI 流利度评分表，此后很多公司都照着它来做。后来 Zapier 自己的想法也更新了。他们是怎么更新的？日常工作里真正要紧的是什么？Wade 实际上是怎么给员工打分的？在今天这期里，他不会只是从理论上把评分表讲一遍，我们会实际展示一些 AI 使用行为，由 Wade 来给出他的判断：这算变革型，还是不合格。

[2:28] Aakash

He'll also talk to you about how these expectations have changed over time so that you can meet the bar at your company or exceed it. Wade, thanks for being on the podcast.

他还会聊聊这些期望是怎么随时间变化的，好让你在自己公司达到、甚至超过这条标准线。Wade，谢谢你来上节目。

[2:38] Wade

Yeah, thanks for having me.

好的，谢谢邀请。

[2:39] Aakash

So, I have been dying to do this recording. I think we've rescheduled a few times because I think you are the world's best person to talk to us about AI fluency. How do you guys measure AI fluency as Zapier?

我真的特别期待录这一期。我们好像改了好几次时间，因为我觉得要跟我们聊 AI 流利度，你是全世界最合适的人。你们在 Zapier 是怎么衡量 AI 流利度的？

[2:52] Wade

Sure. might be most helpful to actually simply just show uh so we have uh an AI fluency rubric. This published March 31st, 2026 is our updated AI fluency bar uh at Zapier. So this is the V2. Uh and you can see here we basically break things down into four categories. You actually don't see there's a there you see three on the scape the screen here. Capable, adaptive, and transformative. We also have an unacceptable one. uh which is you know just simply the bar that we choose not to hire at. Obviously for any company you can tweak and tune these to fit your needs. Uh we felt it was really helpful for us to really name some of the skills and behaviors that uh help people stand out in terms of you know their their AI fluency. And so maybe you know this is a product podcast so let's maybe uh zoom down here to the product one and uh I I'll be the first to to share like this is the the V2 that we've shared externally but we've even updated some of these things internally since like this space moves fast and you do have to constantly be tuning and tweaking uh what you're doing and so you can see here you know some of the characteristics that make someone unacceptable capable adaptive and you know transformative uh at uh this work.

好的。也许最有帮助的，就是直接给你看。我们有一份 AI 流利度评分表。这份是 2026 年 3 月 31 日发布的，是 Zapier 更新后的 AI 流利度标准线，也就是 V2。你可以看到，我们基本上把它分成四个档位。其实屏幕上你只看得到三个：胜任（capable）、适应型（adaptive）和变革型（transformative）。我们还有一档是不合格（unacceptable），说白了就是我们决定不录用的那条线。当然，任何公司都可以按自己的需要去调整这些档位。我们觉得，把那些能让人在 AI 流利度上脱颖而出的技能和行为明确地点出来，对我们非常有帮助。既然这是一档产品播客，我们就往下拉到产品这一栏吧。我得先说一句，这是我们对外公开的 V2，但其实在那之后，我们内部又更新了其中一些内容，因为这个领域变化太快了，你确实得不停地调整、打磨你的做法。所以你可以在这里看到，在这项工作上，一个人被评为不合格、胜任、适应型或变革型，分别具备哪些特征。

[4:15] Wade

And so, you know, in 2026, you know, if you're using AI to, you know, do simple tasks, summarization, writing, looking up information, generating basic PRDs, things like that, um, you know, that's kind of table stakes, I think, for most companies at this point in time. And so, you know, we're usually looking for somebody that's operating at a a meaningfully higher higher higher level. And so, you know, we want folks that have like a really structured appro uh approach for generating specs, prototypes, things that they're going to reuse and refine across projects. Um, you know, we want people using it to really tap into user insights, quant being able to operate at a scale that maybe was not prior possible with sort of humans. Um, you know, you should be able to generate like tons of different working solutions very rapidly. uh you know so those are the types of things that end up being sort of table stakes for these roles. Uh you I think if you're really trying to start to to push the needle you're starting to stretch into other types of roles. I think PMs uh are you know well suited to become sort of like these generalist like super builders inside of an organization and so that means you can tap into whole new skills whether that's data analysis or you know marketing skill sets or sales skill sets like you know PMs are usually already decent at some of these things but now you can get to you know 80 90% as good and it makes you way more valuable sort of to to be able to sort of like work up and down uh you know sort of the the the life cycle of a project. Um you know I think you know a lot of this as you sort of move up the stack starts to be about systems building. Like this is the really critical piece is you're not just thinking about like how's this feature going to work. You're really thinking about how do I build this like pipeline of agents that takes in customer feedback generates specs starts to

所以在 2026 年，如果你用 AI 只是做些简单的任务，比如总结、写作、查信息、生成基础的 PRD（产品需求文档）之类，我觉得到了现在，对大多数公司来说这些都只是入场门槛。所以我们通常要找的，是在明显更高的水平上做事的人。我们希望大家有一套非常结构化的方法来生成 spec、原型这些东西，并且能在不同项目之间反复复用、持续打磨。我们希望大家用 AI 真正去挖掘用户洞察、做定量分析，在过去单靠人力做不到的规模上开展工作。你应该能非常快地生成大量不同的、真正能跑起来的方案。这些东西最后就成了这类岗位的入场门槛。我觉得，如果你真的想再往前推一把，你就会开始延伸到其他类型的岗位上。我认为 PM 很适合成为组织里那种通才型的超级建造者，这意味着你可以掌握全新的技能，不管是数据分析、营销技能还是销售技能。PM 通常本来就对其中一些还算在行，但现在你能达到 80%、90% 的水准，能在一个项目的整个生命周期里上上下下都顶得上，这会让你的价值大很多。我觉得越往上走，很多事就开始变成搭系统了。这才是真正关键的部分：你想的不只是这个功能要怎么运作，而是真正在想，我要怎么搭一条 agent 流水线，让它接收客户反馈、生成 spec、开始

[6:03] Wade

prototype solutions chipfall features. You know if you've heard of things like software factories like I think this is where PM start to have like a lot of values. you can start to think about like how do we uh build the system that builds the ultimate system. Uh and that's where you start to get into this like transformative C category where it's like wow we've actually built like a product factory here. Uh and we're really changing the way you know product works at a sort of uh fundamental level. And so you know these are the sort of four like stages or phases that we've sort of rubric out. Um but like I said you know I think you can adapt these to your company. You know not everyone is you know has the same standards as us. Maybe you have maybe you're not far far as far along and so you sort of have a different state or maybe you're even more advanced than us and you're saying hey actually for us like we need more people that are in these categories um to really push the envelope.

做方案原型、交付完整的功能。如果你听说过「软件工厂」这类东西，我觉得这正是 PM 开始变得大有价值的地方：你可以开始思考，我们要怎么搭建那个能造出终极系统的系统。这时候你就开始进入变革型这一档了，你会觉得：哇，我们居然真的在这儿搭起了一座产品工厂，我们正在从根本上改变产品工作的方式。所以这就是我们在评分表里梳理出的四个阶段。不过就像我说的，我觉得你可以按自己公司的情况去调整。不是每家公司的标准都和我们一样。也许你们还没走到那么远，所以处在不同的状态；也许你们比我们还超前，你会说：嘿，其实对我们来说，我们需要更多人处在这几档，真正去突破边界。

[6:53] Aakash

So what is the distribution that you want to see at Zapier? What should where should the ratings fall

那你希望在 Zapier 看到的分布是什么样的？评分应该……应该落在哪儿？

[6:59] Wade

for our product crew? Like we we're generally looking for folks that sort of have you know like you know sort of the bulk of their skill set is probably here. you know, it doesn't mean that there's they have places where there's gaps like all of us do. You know, there's so much new stuff happening every day that there's new techniques and tactics that people haven't used. And so there's definitely people that are sort of in this bucket. And then, you know, we usually want to see somebody that's like bringing like one example of something that they've done here. You know, most people are not hanging out at transformative every day. Uh in fact, like it's probably not actually helpful to be like here 100% of the time because if you are, you're not actually getting work done. you're constantly just like tweaking your systems and way of working and so you know you you're probably hitting this on occasion but I think kind of like for our PMorg like you know we're kind of hanging out in here like most of the time uh and occasionally running experiments to sort of like push us push the envelope and so that that's kind of where I'd say like our bell curve sort of you know uh sits around is like this is sort of the fat middle of the curve and we have you know some folks in here and some folks in here but you know not it's it's I wouldn't label it as like you know as a person you're this It's it's more like most of the time I'm this, but I have sometimes I've got some holes here and I've, you know, sometimes got some things up here. Like that that's what it looks like more often than not.

就我们的产品团队来说？我们一般希望大家的技能组合大部分落在这里（适应型）。这不代表他们没有短板，我们每个人都有。每天都有那么多新东西冒出来，总有一些新技巧、新打法是大家还没用过的，所以肯定有人处在这一档（胜任）。然后，我们通常希望一个人能拿出一个自己在这一档（变革型）做过的例子。大多数人不会每天都待在变革型。事实上，100% 的时间都待在这里可能反而没什么帮助，因为如果真是那样，你其实就没在干活了，你一直在折腾你的系统和工作方式。所以你大概只是偶尔能达到这一档。但我觉得就我们的 PM 团队来说，大部分时间我们都待在这一档（适应型），偶尔做些实验，推着自己去突破边界。所以我会说，我们的钟形曲线大致就落在这里：这是曲线最鼓的中间部分，有些人在这儿，也有些人在那儿。但我不会给一个人贴标签，说你就是这一档。更像是：大多数时候我在这一档，但有时候我在这儿有些短板，有时候又能在上面这里做出点东西。多数情况下就是这个样子。

[8:17] Aakash

That's an interesting way to frame it that maybe you don't actually want to be transformative because you wouldn't have time to do work. What does like a transformative do you guys do quarters or halves when you guys give performances reviews to folks?

这个说法很有意思：你也许其实并不想一直处在变革型，因为那样你就没时间干活了。那一个变革型的……你们给大家做绩效评估，是按季度还是按半年？

[8:30]

Uh I guess let's see our performance review is in halves. Yeah. So what does it like if you recall or if you can recall like a transformative half from PM what did that look like?

Wade： 呃，我想想，我们的绩效评估是按半年的。对。

Aakash： 那如果你还记得的话，某个 PM 表现达到变革型的那半年，是什么样子的？

[8:41]

Well you know I remember uh at the tail end of last year we had all these new uh you know uh models that came out like uh the the latest like Opus series, the latest GPT series, the latest Gemini series all came out and the reasoning power was like so much better than before. you sort of went from, you know, hey, these AIs are, you know, kind of writing code alongside engineers to over the winter break, basically everyone was like, I'm not writing any code anymore. Like, I'm very rarely writing code anymore. Here's a quick word from our sponsors. Building a lot of AI products lately. My job search OS has 16 different agents. My newsletter has a recommendation engine. And I kept running into the same problem. I'd ship something, it would work in my testing, and then I'd get messages from users saying, "It's hallucinating or picking the wrong tool." The issue wasn't the prompts or the tools. It was that I wasn't actually evaluating anything. I didn't have a way to see what my agent was actually doing step by step. Every tool call, every decision. That's where Arise comes in. Let me show you. I'm going to open Claude Code and install Arise with just one command. npx skills add Arise AI Arise Skills Skill. Yes.

Wade： 嗯，我记得去年年底，一大批新模型接连发布，最新的 Opus 系列、最新的 GPT 系列、最新的 Gemini 系列全都出来了，推理能力比以前强了太多。你基本上是从「嘿，这些 AI 差不多是在跟工程师一起写代码」，一下子变成了过完一个寒假，基本上所有人都在说：我已经不写代码了，我现在很少再亲手写代码了。

Aakash： 先插播一段赞助商的话。我最近在做很多 AI 产品：我的求职 OS 里有 16 个不同的 agent，我的 newsletter 有一个推荐引擎。我一直碰到同一个问题：我上线一个东西，自己测试时好好的，然后就收到用户消息说「它在产生幻觉」或者「它选错了工具」。问题不在 prompt，也不在工具，而是我根本没在做任何评估。我没办法一步一步地看到我的 agent 到底在做什么：每一次工具调用、每一个决策。这就是 Arize 派上用场的地方。我演示给你看。我打开 Claude Code，只用一条命令就装好 Arize：npx skills add Arize-ai/arize-skills --skill '*' --yes。

[9:57]

Now, Claude Code already knows how to instrument my agent. I tell it set up tracing to arise and it automatically analyzes my codebase, figures out where the LLM and tool calls are and adds instrumentation automatically. Now I can see everything, every trace, every span, every decision. And more importantly, I can evaluate it. That's the shift. Trace what's happening, evaluate where it fails, then fix it. This trace right here, my resume feedback agent was supposed to pull the company's text stack from the job posting, but instead it hallucinated that they use React when the posting said Python. I never would have caught that without seeing the trace. And here's the part that blew my mind. I asked Claude Code to look at these traces and tell me what I should be evaluating. It came back with four eval criteria I hadn't written. Things like picking the right tool and staying grounded in the input. I wrote the evals, ran them, and found that my agent was making the same kind of mistake about 12% of the time. Claude pushed a fix. I reran the evals and it dropped to under 2%. That whole loop, trace, evaluate, fix, took me about 20 minutes, and now it runs automatically. If you're building AI products and not evaluating them, you're shipping blind. Try Arise free at arise.com and get a year free, a $1,260 value with my bundle, Arise. Check it out. It's one of the top AI evals platforms used by all of the top AI teams for a reason. So, alongside of that, a a thing that we started to see a lot of folks doing both online, but even intern inside of Zapier was building um these like second brains. Uh, and you know, if you haven't done this, uh, you know, it's a pretty simple thing where you'll create like a, at least the way it started was like you create like a a bunch of markdown files that, you know, might say here's here's what I'm here's my goals. Here's here's the way I want you to work with me. Uh, you know, just

Aakash： 现在 Claude Code 已经知道怎么给我的 agent 埋点了。我跟它说「帮我把 tracing 接到 Arize」，它就会自动分析我的代码库，找出 LLM 调用和工具调用在哪儿，然后自动加上埋点。现在我什么都看得到：每一条 trace、每一个 span、每一个决策。更重要的是，我能对它做评估了。这就是转变所在：追踪发生了什么，评估哪里出了错，然后修复。看这条 trace：我的简历反馈 agent 本该从职位描述里提取这家公司的技术栈，可职位描述写的是 Python，它却凭空编出他们用的是 React。要不是看到这条 trace，我永远发现不了这个问题。接下来这部分真的让我大受震撼：我让 Claude Code 看看这些 trace，告诉我应该评估些什么。它给出了四条我自己没写过的 eval 标准，比如选对工具、紧扣输入内容不跑偏。我把这些 eval 写出来跑了一遍，发现我的 agent 大约有 12% 的时候都在犯同一类错误。Claude 推了一个修复，我重跑 eval，错误率降到了 2% 以下。追踪、评估、修复，这一整个循环花了我大概 20 分钟，而且现在它会自动运行。如果你在做 AI 产品却不做评估，那你就是在闭着眼睛发布。去 arize.com 免费试用 Arize，通过我的 bundle 还能免费用一年，价值 1,260 美元。Arize，去看看吧。它是顶尖的 AI eval 平台之一，所有顶尖 AI 团队都在用它，这是有原因的。

Wade： 所以，与此同时，我们开始看到很多人在搭建这种「第二大脑」，网上有，Zapier 内部也有。如果你还没搭过，其实挺简单的：你会创建一个……至少最开始的做法是，你创建一堆 markdown 文件，里面可能写着：这是我的目标，这是我希望你配合我工作的方式。嗯，就是

[11:53] Wade

have like a bunch of context. Um, you know, this is a project I'm working on, etc. And then, you know, you'd boot up a coding agent and the coding agent would have access to all that context uh to to go, you know, build projects with. And so maybe you have your D design system in there, maybe you have your company strategy in there, maybe you have, you know, coaching files in there where you're like, help me be better at this. Um, you know, things like that. And so, you know, one of our product leaders said, you know what, this second brain concept is pretty interesting, but uh it's limited if those are individual um skills. this should be a company brain. And so, you know, he took it upon himself to say, I'm gonna go build uh, you know, out the sort of V1 of a company brain. Uh, and, you know, I'm going to sort of put together like all this context that would help uh, anyone in the company, but he certainly started with a lot of the product leaders and product managers and said, "Hey, we're going to get everybody sort of using this, you know, second second brain or company brain to to go do the work." And I would say there was like you could sort of feel this like um phase shift happened where it was like not just him but like every product person in Zapier kind of got like a jetpack put on their back uh when he sort of launched this system. And so that's a great example of like this moment in time where it was like you know the models sort of had this like lurch forward in capabilities and you know product leader said okay with these new capabilities I can lurch forward the way we work as well too with it.

放上一大堆上下文。嗯，比如「这是我正在做的一个项目」之类的。然后你启动一个 coding agent，这个 coding agent 就能用上所有这些上下文，拿去做项目。所以里面可能放着你的设计系统，可能放着你们公司的战略，也可能放着一些教练类的文件，比如「帮我在这件事上做得更好」。嗯，诸如此类。后来我们有一位产品负责人说：你知道吗，第二大脑这个概念挺有意思，但如果这些只是个人层面的 skill，那就有局限了。这应该是一个公司大脑。于是他主动把这件事揽了下来，说：我来把公司大脑的 V1 搭出来，我要把所有能帮到公司里任何人的上下文都汇集起来。不过他当然是先从一大批产品负责人和产品经理开始的，他说：「嘿，我们要让所有人都用上这个第二大脑，或者说公司大脑，来干活。」我会说，你能明显感觉到发生了一次相变：不只是他本人，而是 Zapier 的每一个产品人，在他推出这套系统之后，都像是背上了一台喷气背包。所以这是个很好的例子，说明在那个时间点上，模型的能力猛地往前蹿了一大步，而这位产品负责人说：好，有了这些新能力，我也能让我们的工作方式跟着猛地往前蹿一大步。

[13:22] Aakash

All right. So, I want to try some PM artifacts. I've actually gathered some that I created this morning, and I want to put you on the spot. And what I really want to do, it's like the the final rating doesn't matter as much as how does a CEO look at PM work products, PM using AI or how might a particular PM who's in a company be being graded on this? So, let's start here. What I thought of was what if Zapier had an IDE harness for Claude code. [laughter] Who knows if this is a good idea at all. That's why I wrote a PRD on this and I want to walk you through this PRD as if I were a PM. And I'd love to hear your thought process as you're looking at this.

好。接下来我想拿几份 PM 的产出物来试试。其实我今天早上专门做了几份，想让你当场点评一下。我真正想看的是——最后打几分倒没那么重要，重要的是一个 CEO 会怎么看 PM 的工作成果、怎么看 PM 用 AI 这件事，或者说，一家公司里某个具体的 PM 在这方面会被怎么打分？那我们就从这儿开始。我想到的点子是：如果 Zapier 给 Claude Code 做一个 IDE harness 会怎么样？（笑）谁知道这到底是不是个好主意。所以我为它写了一份 PRD，我想以 PM 的身份带你过一遍这份 PRD。也很想听听你看的时候是怎么思考的。

[14:17] Aakash

Um, obviously I've used Chat GPT here and I've I've given it a few prompts. So I have iterated on it. It's not, you know,

嗯，显然我这里用了 ChatGPT，也给了它几轮 prompt，所以是迭代过的。它不是，你知道，

[14:24] Aakash

my first draft or anything,

我的第一稿什么的，

[14:26] Aakash

but uh I I've included the overview. I said it's important for automation workflows, and you we're uniquely positioned to solve this because of our thousands of integrations, right? And our deep knowledge about how workflows work. So, we combine cloud code with Zapier's automation platform. That's my pitch here.

不过我写了概述部分。我写的是：这对自动化工作流很重要，而我们在解决这个问题上有独一无二的优势，因为我们有成千上万个集成，对吧？还有我们对工作流运作方式的深刻理解。所以，我们把 Claude Code 和 Zapier 的自动化平台结合起来。这就是我这里的核心卖点。

[14:45] Aakash

I feel like the problem is that you might write prompts in cloud code, configure apps, deploy it. It's all fragmented. Cloud code can make mistakes with the API and so what I wanted to do was create a better environment for building these Zap year automations and I feel like the the target is you know software engineers founders and so specifically for an engineer as a developer using cloud code I want to build and test zapper in automations without constantly switching and so that's really the problem we're solving we have this proposed solution where users can interact with cloud code a workflow might be that they have this workflow and Claude could easily help them build it. We could test it. I've come up with some of the key features that we have here and I won't go through too much more of the detail, but it'll have AI slop detection.

我觉得问题在于：你可能在 Claude Code 里写 prompt、配置应用、再部署，整个过程是割裂的。Claude Code 在调 API 时可能出错。所以我想做的，是为搭建这些 Zapier 自动化打造一个更好的环境。我觉得目标用户是软件工程师、创始人。具体到工程师就是：「作为一名使用 Claude Code 的开发者，我想搭建和测试 Zapier 自动化，而不用来回切换。」这才是我们真正要解决的问题。我们提出的方案是，用户可以和 Claude Code 交互，一个使用场景可能是：他们有这么一个工作流，Claude 就能轻松帮他们搭出来，我们还能对它做测试。我列出了一些关键功能，细节就不再多展开了，不过它会带 AI slop 检测。

[15:40] Aakash

It I've come up with a little bit of an MVP. Some metrics that I think you know leadership might care about is the number of workflows successfully deployed through the IDE as sort of our northstar metric.

我还构思了一个简单的 MVP。至于管理层可能会关心的指标，我觉得可以把通过这个 IDE 成功部署的工作流数量，作为我们的北极星指标。

[15:52] Aakash

Um, of course developers might prefer cloud code directly but I don't think there's a huge build cost and sort of my vision is cloud writes or automation. Zapier helps make sure it actually works. So I wrote this with AI. What do you think of my work?

嗯，当然，开发者可能更愿意直接用 Claude Code，但我觉得这个东西的开发成本不算高。我的愿景大概是：Claude 来写自动化，Zapier 负责确保它真的能跑通。这份东西是我用 AI 写的。你觉得我做得怎么样？

[16:08] Wade

Okay. So, uh, you know, it's interesting to think about how you would start to go, uh, evaluate this because, you know, part of your prompt to me was, uh, you know, how does this grade an AI fluency, but there's also a part of this that's just how do you grade the overall like product thinking that's sort of happening here. And so, the first thing I notice is one, I I think this is actually a pretty interesting idea. uh you know I think you know you're sort of like like good product intuition you know I would say is like you know pretty pretty savvy here uh you know recognizes that like there's this new way of building automations or this new way of building there's some value in sort of having like the zap year uh you know connective library hooked up to claude or claude code uh to help go build and then um you know one thing I didn't see you say here but maybe maybe it was in it I I I suspect it's in here uh I couldn't read quite fast as you were going through this was um there's a lot of value in sort of deploying these things and running them on Zapier because Zapier can run them in a uh deterministic manner which you know allows you to get better costs you know better accuracy all that sort of thing so there's a lot of stuff in here that I like now where I think you know a a sort of like AI pill AI maximalist like product manager could go further would be in a couple dimensions I think one um why stop at the PRD uh you know uh there's an opportunity to go build out you know the first prototype of what this thing looks like.

好。嗯，琢磨一下该怎么着手评估这份东西，其实挺有意思的。因为你给我出的题里，一部分是「它在 AI 流利度上该打几分」，但还有一部分就是：这里面体现出来的整体产品思维该怎么打分。所以我注意到的第一点是：我觉得这其实是个挺有意思的点子。你的产品直觉不错，我会说在这一点上相当有眼光——你意识到现在出现了一种构建自动化的新方式，或者说一种新的构建方式；把 Zapier 的连接库接到 Claude 或 Claude Code 上来帮忙构建，是有一定价值的。然后，有一点我没看到你提，不过也许里面写了——我猜应该写了，你翻得太快我没来得及看——那就是：把这些东西部署到 Zapier 上、在 Zapier 上运行，价值很大，因为 Zapier 能以确定性的方式运行它们，这能带来更好的成本、更高的准确率，诸如此类。所以这里面有不少东西我挺喜欢的。那么，我觉得一个彻底 AI 化、信奉 AI 最大化的产品经理还能再往前走的地方，体现在几个维度上。我觉得第一个是：为什么止步于 PRD？你完全有机会把这个东西的第一版原型做出来。

[17:39] Wade

Uh and so let's go fire up you know your favorite you know coding agent of choice and start to like see what you can prototype and build. And now the point here is not for product managers to build production level code. The point is that uh you know seeing is believing and it's a lot easier now for a product manager to uh bring some of these ideas to life. Uh and so you know uh there's that sort of you know a picture you know says a thousand words well you know a prototype is you know that times 10 right so I think that would be one dimension I wish I this could be uh you know stronger and you know an AI pill person would sort of you know probably do something impressive there. I think the second dimension where I would try and strengthen this and this is you know I think you could do this sort of in a you know non-AI way but certainly I would be using AI to help me which is uh you know let's go start to figure out um two things like where is our customer evidence behind this uh so however you sort of can find customer evidence uh I would try and find it so I would look through you know the Zap year um you know like gong recordings or Zindesk tickets. Um, you know, I would go sift through like, you know, subreddits or uh LinkedIn or, you know, X and sort of find like, hey, where are people like complaining about stuff like this or expressing a desire for a thing like this to try and figure out like, am I getting the shape of this product right?

那就打开你最顺手的 coding agent，开始看看你能做出什么原型、搭出什么东西。这里的重点不是让产品经理去写生产级代码。重点在于眼见为实，而现在产品经理要把一些想法变成看得见摸得着的东西，比以前容易太多了。有句话叫一图胜千言，那原型差不多就是再乘以 10，对吧？所以我觉得这是一个维度，我希望这份东西在这方面能更强一些，一个彻底 AI 化的人大概会在这里拿出点让人眼前一亮的东西。第二个我会想办法加强的维度——这个其实不用 AI 也能做，但我肯定会用 AI 来帮我——就是去搞清楚两件事，比如：这背后的客户证据在哪儿？所以不管通过什么途径能找到客户证据，我都会去找。我会去翻 Zapier 的 Gong 通话录音、Zendesk 工单。我会去各个 subreddit、LinkedIn 或者 X 上筛一遍，找找看：大家在哪儿抱怨这类问题，或者在哪儿表达过想要这种东西，以此来判断：我对这个产品形态的把握对不对？

[19:08] Wade

Um, now like I said, you could do a lot of this stuff without AI, but AI is probably going to make it a lot easier for you to sift through a mountain of data sources, um, that, uh, you know, you probably just wouldn't have been able to do in, you know, a short period of time in the past. Um, so I think that is, you know, a second thing you could do to sort of like really beef up uh, the evidence um, that like, hey, am I in the right direction here? Uh, like I think this is a good idea. Are our customers going to think it's a right idea? and like what is the pain point that they really care about? Um you know you posed the question like hey Claude could maybe just do some of this. So there is a little bit of a question around like well what is like where is the unique value that Zapier brings to the table on top of it and I think you know sifting through a lot of those customer details will start to like give build strengthen that intuition around like ah there is something here that you know is uniquely zapiers to own uh and provide value on top of. Um, so those are the two sort of like obvious things that I would use just like would encourage somebody to strengthen. And I don't think that either of those things takes that much more time. Uh in the past there would have been things where it was like either a product would a PM would have just basically well I can't build a prototype so I I just simply can't do it or with the customer evidence one it's like well I can do that but you know there's a question of like before I go spend you know like a couple weeks doing it I kind of want to get some buy in that this is like even directionally interesting whereas nowadays you know that buyin step can happen way later you can just go so much farther and so by the time you're sort of asking you know me or somebody else in the company to be like, "Hey, let's go put some oomph behind it." You know,

嗯，就像我说的，这里面很多事不用 AI 也能做，但有了 AI，你去筛查堆积如山的数据源大概会容易得多——换作过去，你大概没法在很短的时间里做完这些。所以我觉得这是你可以做的第二件事，用来真正把证据做扎实，确认：嘿，我的方向对不对？我觉得这是个好主意，可我们的客户会不会也觉得这个主意对？他们真正在意的痛点是什么？你自己也提出了一个问题，比如「Claude 说不定自己就能做掉其中一部分」。所以这里就有个疑问：在这之上，Zapier 能拿出来的独特价值到底在哪里？我觉得把大量这类客户细节筛一遍，就会开始帮你建立、强化那种直觉，比如：啊，这里确实有一块是只有 Zapier 能拿下、能在上面创造价值的。嗯，所以这就是两件比较显而易见的事，我会鼓励别人去加强。而且我不认为这两件事哪一件会多花很多时间。放在过去，情况要么是 PM 基本上会说：「我不会做原型，所以这事我压根干不了。」要么就客户证据这件事来说：「这个我能做，但问题是，在我花上几周去做之前，我得先争取到一些支持（buy-in），确认这件事哪怕只是在方向上值得做。」而现在，争取支持这一步可以往后放很多，你自己就能先往前走得远得多。这样等到你来找我，或者找公司里的其他人说「嘿，咱们给这件事加把劲吧」的时候，你知道，

[20:49]

you've built like a mountain of evidence behind it that makes it, you know, a lot more likely that I'm going to say, "Yes, let's go for it." Or maybe you've already figured out yourself like actually this is just a really bad idea. Like I I I thought I thought my intuition was good, but you know, I put some effort into it and I just can't find something that I like get super excited about. And so you can shelf it. Uh and so that's where I think like, you know, an AI powered PM just has these like superpowers that you didn't have in the past. Did you guys realize you were getting a master class in how to write good PM documents? It's not just AI fluency. I think Wade just brought up a couple really, really important points around it's not just the PRD. And what I tried to do, I tried to engineer this document so you would rate it as like a capable in 2024, but almost like an adoptive or unacceptable in 2026. Do you feel like that's fair?

Wade： 你已经为它攒下了堆积如山的证据，这让我说出「好，就这么干」的可能性大得多。又或者，你自己已经想明白了：其实这就是个很糟糕的主意。比如「我原以为自己的直觉挺准，但花了些功夫之后，我就是找不到能让我特别兴奋的东西」。那你就可以把它先搁置起来。所以我觉得，这就是 AI 加持的 PM 身上那些过去没有的超能力。

Aakash： 大家有没有意识到，你们刚刚上了一堂「怎么写好 PM 文档」的大师课？这不只是 AI 流利度的事。我觉得 Wade 刚才提了几个非常非常重要的点，核心是：不能只停在 PRD。我当时的想法是，刻意把这份文档设计成：放在 2024 年你会给它打「胜任」，但到了 2026 年，差不多就是「适应型」或者「不合格」了。你觉得这个评价公平吗？

[21:40] Wade

I think so. Yeah. Like I think if someone came to me with this like you know, in 2026, I would sort of feel like I it feels a little oldfashioned, like as sort of the answer that I would have. I'd be like, yeah, uh, you know, some good product intuition, some some good thinking, but like gosh, you really could have gone a lot further. Like, you you really and and I would almost want to I'll tell you what my instinct would be. My instinct would be like, just give me the doc and I'll start go doing some of these things. Like, I'll literally just take your PRD and like, you know, throw it into cursor and be like, hey, bu build this, you know, for me. and see what happens. And you know that feels like you know I think you know the way that I would sort of here's an interesting like rule of thumb like you know you probably want to be in your discipline you probably want to be farther ahead than your CEO is. And you know I I would say you know my AI fluency is like probably you know adaptive in most places. But you know what it means for me to be AI fluent in the PM job or the marketing job or the engineering job, like I should be behind you. Uh, and so if there's things that I know that you don't know, like that that's probably not a great signal. Uh, you should be further ahead than me because you're you wake up and breathe this stuff all day every day. And so if I'm spotting things that, you know, like are obvious to me and not to you, I'm like, "Oo, the the that makes me that makes me nervous, right?" and and I'm like, hm, h how do we go fix that?

我觉得公平。对。我觉得如果 2026 年有人拿着这份东西来找我，我会觉得它有点过时了，这大概就是我会给的回答。我会说：嗯，有不错的产品直觉，也有一些不错的思考，但天哪，你本来可以走得远得多。你真的……我甚至会想——我跟你说说我的本能反应会是什么吧。我的本能反应会是：把文档直接给我，我自己开始去做其中一些事。我真的会直接拿你的 PRD 扔进 Cursor，说「嘿，帮我把这个做出来」，然后看看会发生什么。这就引出一个挺有意思的经验法则：在你自己的专业领域里，你大概应该比你的 CEO 走得更靠前。我会说，我的 AI 流利度在大多数方面大概是「适应型」。但说到在 PM、市场或者工程这些岗位上做到 AI 流利，我应该落在你后面才对。所以如果有些东西我知道而你不知道，那大概不是个好信号。你应该比我走得更靠前，因为你每天一睁眼就泡在这些东西里，天天如此、整天如此。所以如果我发现有些东西对我来说显而易见、对你却不是，我就会想：「哦，这让我有点紧张，对吧？」然后我会想：嗯，我们该怎么把这个问题解决掉？

[23:12] Aakash

Okay, so I think we're aligned there. And one other thing I would just pull out as an insight to triple click for everybody who is listening is it's not easy to just separate AI fluency and good product thinking. They almost come together in a sense in this era. Like to do good product thinking, you really need to have gone and looked at the gong calls. You need to have gone and looked at the Reddit comments and now AI can help filter half of that for you. And so you really need to put those two skills together.

好，那我觉得我们在这点上是一致的。还有一点，我想单独拎出来，给所有在听的人深挖一下：AI 流利度和好的产品思维，其实很难简单地分开。在这个时代，它们某种意义上几乎是合在一起的。要做好产品思考，你真的得去看过 Gong 的通话录音，得去看过 Reddit 上的评论，而现在 AI 能帮你筛掉其中一半。所以你真的需要把这两项技能结合起来。

[23:47] Wade

Yes. Like we were talking before the call about AI slap a little bit. And so there is a version of this where you could do all the things that I said which is like oh just take this PRD slap it into a coding agent have it build the first prototype. You could sort of sort of have AI do all of those steps, but out the output you could come away would be just like pretty mediocre and unexciting. And so this is where the human like a lot of people call it taste judgment like whatever you know whatever term we're calling it these days really does matter. And so you know there is still something to be said about like really understanding the customer problem really having some intuition around how to go solve those problems. And if you combine those, you get excellence. But if you just let AI kind of just do everything, you're probably going to get kind of mediocrity. You're going to get what people call AI slop. You're going to look at it and go like, h, this is not like I'm not excited about this.

对。我们录制之前也稍微聊过一点 AI slop。所以有一种做法是，你把我刚才说的事全做了：直接拿这份 PRD，往 coding agent 里一扔，让它做出第一版原型。你可以让 AI 把这些步骤全包了，但你最后拿到的产出可能就是相当平庸、毫无亮点。所以这就是人真正要紧的地方——很多人管它叫品味、判断力，不管我们现在用什么词来称呼它。所以，真正理解客户的问题、对怎么解决这些问题真正有一些直觉，这些依然很有价值。把这些结合起来，你就能做到卓越。但如果你就放手让 AI 把所有事都干了，你大概只会得到平庸的东西，得到大家所说的 AI slop。你看着它会想：嗯，这不是……我对这个提不起兴趣。

[24:43] Aakash

Speaking of slop, I want to know what your take is on this framework from Dan Hawinmire over at Fair. If you don't have good judgment and you use AI, you can become a slob cannon. But if you have good judgment and you use AI, you can become a turbo brain. Is this the right way to think about it?

说到 slop，我想听听你怎么看 Faire 的 Dan Hockenmaier 提出的这个框架：如果你判断力不行又用 AI，你可能会变成一门「slop 大炮」（slop cannon）；但如果你判断力好、又用 AI，你就能变成一颗「涡轮大脑」（turbo brain）。这么看待这件事对吗？

[24:59] Wade

Uh, I mean, it's definitely funny and evocative. I I think it's a reasonable framework, honestly. Uh, I can't think of why this is not I can't think of why I would sort of have an an objective to to this. Like, I I can think of folks that I've worked with who are using AI and have ex exceptional judgment and just watching them work, you're just like, "Wow, this is this is incredible to like watch this person work." You're it's almost like you're you're witnessing a new species being invented. Uh, and then I've also seen folks that are using AI a lot but do not have good judgment. And you know, it's there's this old saying like kind of like in the social media wars around like uh disinformation where it's like uh you know, a lie uh can go around the two the world two times before the truth gets out of bed. And I feel like the folks without good judgment who use AI are little like that where they can just generate so much stuff that like you're just trying to like, you know, it takes so much just to kind of like swat it out and be like, uh, like get, you know, that's not that's not what I'm looking for. Not what I'm looking for. And so it actually ends up being in some ways more challenging to work with those folks than maybe even like the bottom left corner where it's like doesn't have good gener doesn't use AI. Well, it's like, well, at least at least you're quiet, right?

呃，我是说，这个说法确实很好笑，也很有画面感。说实话，我觉得这是个挺合理的框架。呃，我想不出它有哪里不……想不出我有什么理由要反对它。比如，我能想到一些共事过的人，他们在用 AI，判断力也非常出色，光是看着他们干活，你就会觉得：「哇，看这个人干活真是太不可思议了。」简直就像在见证一个新物种被发明出来。呃，然后我也见过一些人，AI 用得很多，但判断力不行。你知道，有句老话，有点像社交媒体上围绕虚假信息的那些论战里常说的：真相还没起床，谎言已经绕着世界跑了两圈。我觉得那些判断力不行、又在用 AI 的人就有点像这样：他们能生成海量的东西，你就只能拼命……你知道，光是把这些东西一个个拍掉就得花好大力气，还得说：呃，这不是我要的，不是我要的。所以在某些方面，跟这些人合作反而更费劲，可能比左下角那一格还费劲——就是没有好的……呃，不用 AI 的那种。那种人吧，好歹……好歹你还挺安静的，对吧？

[26:13]

[laughter]

（笑）

[26:14] Wade

Like uh so uh you know it's it's an interesting framework too. I'd probably have to think a little bit more about some of the pros and cons with it. But on the surface I'm like yeah I I feel this

所以，呃，你知道，这也是个挺有意思的框架。它的利弊我可能还得再多想想。但乍一看，我的感觉是：对，我有同感

[26:25]

on my own.

在我自己……

[26:26]

There's a reason it probably went pseudo viral. [laughter]

它能小火一把，大概不是没有原因的。（笑）

[26:28] Aakash

On my own team I find that like when somebody sends me something that might fall in the slop category. They're basically

在我自己的团队里，我发现，当有人发给我一个可能算得上 slop 的东西时，他们基本上

[26:38] Aakash

they didn't exercise the judgment on their own and they're passing that responsibility to me. Like they're like saying like, "Oh, this might be good. Can you judge it?" Right? And that's like kind of the opposite of what you want to do when you're moving up a corporate hierarchy, right? You want to give people filtered information that is quality versus

就是自己没动用判断力，把这份责任甩给了我。他们等于是在说：「哦，这个可能还行，你帮我判断一下？」对吧？而这差不多正好跟你想在公司里往上走时该做的事反着来，对吧？你应该给别人经过筛选、有质量的信息，而不是

[26:57] Aakash

what do you think of this?

你觉得这个怎么样？

[26:59] Wade

Yeah. I I think you know there's there's almost like some etiquette that I think the workplace is trying to iron out around this right now. Uh you I talked to a lot of folks that are sort of dealing with this like you know the rise of slop internally and um you know I think some of the things we are starting to do is like be really honest upfront around like hey what's what level of effort has been put into this yet? And so, you know, when I share a document, I'll say like, "Hey, um, you know, obviously I have used AI to help draft this. I have done a quick skim through." And so, like, I've removed sort of like the most obviously bad stuff inside of here, but, you know, maybe I haven't given it like a thorough once over, but I want to share anyway because I think speed is really important now, and I think I've gotten this I've gotten some ideas out there and I want you to take a look at it, right? So, that might be one thing I will do. Um, or I might say, you know what, I think this is excellent. like I have really poor, you know, so I I stand by every statement, every sentence like has been like impeccably reviewed by me. And so we we kind of have like these stages of like when we're sharing these things to help the reader understand like you know what's the level of like effort that's sort of been put into this. Um, and you know, I do think it's okay at times to just share, you know, what might be pure AI slot, but I think you got to like set the expectation to the reader to go like, okay, how much, you know, uh, how much should I like like what's the quality level that I should expect out of this? And I think that helps. So, like I I don't know. I and I'm not sure that we totally perfected this to be honest, but the the place that is really painful is when you sort of do try and pass it off as like really good, but then the the reader can just like quickly tell like you you're just you

对。我觉得现在职场上几乎正在围绕这件事磨合出一套礼仪。呃，我跟很多人聊过，他们都在应对这个问题，就是公司内部的 slop 越来越多。呃，我觉得我们开始做的一件事，就是一上来先坦白说清楚：嘿，这东西到目前为止花了多少功夫？所以我分享文档的时候会说：「嘿，呃，你们也看得出来，这份我用了 AI 帮忙起草，我自己快速扫了一遍。」也就是说，最明显的烂东西我已经删掉了，但可能还没从头到尾仔细过一遍；不过我还是想先发出来，因为我觉得现在速度真的很重要，我已经把一些想法摆出来了，想请你看一看，对吧？这是我可能会做的一种。呃，或者我也可能会说：你知道吗，我觉得这份东西非常出色，我真的仔细推敲过，所以每一个论断我都认，每一句话都经过我一丝不苟的审阅。所以我们分享这些东西的时候，大概有这么几个档次，好让读的人明白这份东西到底花了多少功夫。呃，而且我确实觉得，有时候直接分享一份可能纯属 AI slop 的东西也没问题，但你得先给读的人定好预期，让他知道：好，我该……我对这份东西该期待什么样的质量水平？我觉得这样会有帮助。所以，嗯，我也说不好。说实话，我不确定我们已经把这件事做到完美了，但真正让人难受的，是你想把它包装成很好的东西拿出来，结果读的人一眼就能看出来：你就是

[28:40]

literally put no effort into this and now you're making me do the task and it's like you should have just asked me to just do the job anyway versus try and pass it off as your own. So, one thing that really surprised me so far into this podcast that I definitely wasn't expecting coming in was it sounds like it's okay to use some AI. It's okay to have a little bit of slop if you're going for speed. It's okay to not always be transformative with AI as long as you're upfront and transparent about it.

Wade： 压根没花一点功夫，现在还要我来替你干活。那你当初还不如直接让我来干这活，而不是想把它当成你自己的成果交上来。

Aakash： 这期播客录到现在，有一点真的让我挺意外，是我来之前完全没料到的：听起来，用点 AI 是可以的；如果你追求速度，有一点 slop 也可以；用 AI 不必每次都做到「变革型」——只要你事先说清楚、保持透明就行。

[29:09] Wade

Yeah, I think so. Like I think these are these are tools, right? Uh and you know, a lot of it depends on the context. you know, if you're putting out an external product for your customers, like there's a different standard for that versus like, hey, I got this one quick and dirty task over here and like the level of quality required for that is maybe different. Uh, and so, you know, a lot of it comes down to what's on Dan's right hand side has good judgment. And so, you know, judgment is like, you know, kind of the ultimate in a lot of ways. is like one of the main critical skills right now in the AI era is like if you got you got to turn your brain on. Uh and if your brain is on, you're probably going to, you know, mostly find yourself in like a good camp. But if you sort of turn turn your brain off, you're probably going to find yourself struggling.

对，我觉得是这样。我觉得这些说到底就是工具，对吧？呃，很多事要看具体场景。如果你是要给客户推出一款对外的产品，那标准就不一样；如果只是，嘿，我这儿有个又快又糙的小活儿，那它要求的质量水平可能就不同。呃，所以很多时候归根结底要看 Dan 那张图右半边那一栏：有没有好的判断力。所以你看，判断力在很多方面可以说是终极能力，是 AI 时代眼下最关键的核心技能之一——你得把脑子打开。呃，脑子开着，你多半会发现自己基本落在好的那一边；可要是你把脑子关了，你多半会发现自己很吃力。

[29:55] Aakash

Awesome. So, let's go review another work product that I've prepared. I think we can safely say the last one, we would probably not give that a very high rating in 2026. Let's go check out another one

太好了。那我们再来看一份我准备的工作成果。我想可以放心地说，上一份放在 2026 年，我们大概不会给它打很高的评级。我们来看看另一份

[30:08] Aakash

that I worked on before this. Uh, and I'm nervous now that it might get a low rating, but that's the point, right? So, let's go see what rating I get. I actually decided to use the same feature so that we get it makes it easier for us to really compare. And so, it's again, it's like a Zapier harness for Claude code. I predicted we should probably start with a prototype. So, I did come up with a prototype here. Here's what I have for my next layer. I actually did start with a prototype and I tried to create some Zapier branding even uh your agent has the connections it just can't reach them is kind of my pitch right since we hold authenticated access to 8,000 plus apps I think even more now what it would it look like if a coding agent could actually use them and so the whole idea is we just cla install Zapier Zapier connect we could approve a Slack message, we'd start to get this exposed area. We could even send out things like Jira delete issue. And all of that happens so incredibly fast. And the whole point is that it found six connections. Three were selected. It looked at 14 custom Jira fields. It showed us which tools across our cap, which ones were preapproved. It gave a preview of the payload. There was this thing that I built in here. It happened really fast where the juror delete was refused.

是我在这之前做的。呃，现在我有点紧张，怕它评级也不高，不过这正是重点，对吧？那我们来看看我能拿到什么评级。我特意用了同一个功能，这样更方便我们真正做对比。所以还是那个东西：给 Claude Code 做的一个 Zapier harness。我预判到我们大概应该从原型做起，所以这次我确实做了个原型。这是我为下一个层级准备的东西。我真的是从原型开始的，甚至还试着加了点 Zapier 的品牌元素。我的 pitch 大概是：「你的 agent 手里有这些连接，只是够不着它们」，对吧？既然我们握有 8,000 多个应用的授权访问——我觉得现在还不止——那如果 coding agent 真能用上它们，会是什么样子？所以整个思路就是：我们只要 cla install Zapier、Zapier connect，就能批准一条 Slack 消息，接着就开始出现这块暴露出来的区域。我们甚至能发出像 Jira delete issue 这样的操作。而这一切都快得不可思议。关键在于：它找到了 6 个连接，选中了其中 3 个；它查看了 14 个 Jira 自定义字段；它给我们列出了有哪些工具、其中哪些是预先批准过的；它还给出了 payload 的预览。我在里面还做了一个设计，刚才一闪而过：那个 Jira 删除操作被拒绝了。

[31:40] Aakash

So what did we do then and why? Well, it was a destructive action and so it was flagged. And so that's one of the things that our harness would be able to add on that like maybe a normal cloud code wouldn't because we have a lot of enterprise customers. We could even replay it in Zapier history. And so the whole idea here is that this harness would be much more powerful. And this is where I started my PRD. So I'm going to keep going and just explain a little bit more the PRD behind it too and then let's get your feedback. So now if I show the PRD, this is going to be a markdown file instead of just a chat window this time. So we just went through the prototype. Basically the whole idea is you just install and connect once. You just saw that it has a policy file you just saw where it denied an action. The loop we're talk we're really talking about it unlocking is let's say there's some sentry issue. It can go read it. It can go edit it. It can go create the jurama. It can go post a slack message. All of that can happen using Zapier's connections and Zapier's policies. And so somebody who is super technical can build it or somebody who is not technical. We already went through what this is deliberately proving. And so we since we talked about this hypothesis, I won't talk about it too much, but I've written it in kind of like a slightly tighter way this time.

那我们当时做了什么，为什么这么做？因为那是个破坏性操作，所以被标记了出来。这就是我们的 harness 能额外加上、而普通 Claude Code 可能做不到的东西之一，因为我们有很多企业客户。我们甚至可以在 Zapier 的历史记录里回放它。所以整个想法是，这个 harness 会强大得多。我的 PRD 就是从这里开始写的。我接着往下讲，把背后的 PRD 也再多解释一点，然后听听你的反馈。现在我打开 PRD，这次它是一个 markdown 文件，而不只是一个聊天窗口。刚才我们过了一遍原型。基本思路就是：你只需要安装、连接一次。你刚才也看到了，它有一个策略文件（policy file），也看到了它拒绝执行某个操作的那一幕。我们真正想让它解锁的闭环是这样的：比如 Sentry 上出了个 issue，它可以去读、可以去改、可以去 Jira 建单、可以去发一条 Slack 消息，这些全都可以借助 Zapier 的连接和 Zapier 的策略来完成。所以不管是技术很强的人，还是不懂技术的人，都能把它搭起来。这个项目刻意要验证什么，我们前面已经讲过了。既然这个假设已经聊过，我就不多说了，不过这次我把它写得更紧凑了一点。

[33:02] Aakash

And I tried to use a little more evidence. So I did predict that we needed to use some evidence. So I've added in like some gong information like what I from 47 calls from February to July and I found that agent tooling came up unprompted in 34 of these 47 calls. This is just an observation. On Zenesk, I saw we had 312 tickets tagged MCP or AI agent. In our community Discord, over 1,800 posts mentioning the MCP. And our design partners, we had nine interviews where they had built an MCP server. Now, what did all this information show? What it showed is the bottleneck is permission, not connection. It also showed that too many tools make agents worse and users already know it and that there is a real tale. So my hypothesis with the little six design partner pilot that I ran was that devs will let an agent perform rights to production systems if the gate is good. I wasn't sure if this was going to be true and I did find evidence to support this. Five of six enabled at least one preapproved right. I went through several other hypotheses and ultimately looked at it across our strategic fit and I came up with this idea that we do a claude plugin and this plugin I feel is really a harness not a connector and I've gone through a little bit more detail on that including the metrics the risk and what I think I want to get your permission for in this meeting is that we go through 10 design partners now we do a legibility bake off. We start to look at security. So, if you approve this, then we'll plan on kind of going forward with a Q4 beta on this project. What do you think of my latest work artifact?

我也试着多用了一些证据。我确实预判到我们需要拿出点证据，所以我加了一些 Gong 的信息：我看了 2 月到 7 月的 47 通通话，发现其中有 34 通，客户在没人提起的情况下主动聊到了 agent 工具。这只是一个观察。在 Zendesk 上，我看到有 312 张工单被打上了 MCP 或 AI agent 的标签。在我们的社区 Discord 里，有 1,800 多条帖子提到了 MCP。至于我们的设计合作伙伴（design partner，早期共建客户），我们做了 9 次访谈，他们都自己搭过 MCP server。那这些信息说明了什么？说明瓶颈在权限，不在连接。它还说明，工具太多会让 agent 变差，而且用户已经知道这一点；另外，长尾是真实存在的。所以，我跑了一个 6 个设计合作伙伴参与的小试点，假设是：只要闸门设计得好，开发者就愿意让 agent 对生产系统执行写操作。我原本不确定这个假设成不成立，结果确实找到了支持它的证据：6 家里有 5 家至少开启了一项预先批准的写操作。我还过了其他几个假设，最后从战略契合度的角度通盘看了一遍，得出的想法是：我们做一个 Claude plugin，而我觉得这个 plugin 本质上是个 harness，不是 connector。这部分我写得更细一些，包括指标和风险。我想在这次会上请你批准的是：我们现在去跑 10 个设计合作伙伴，做一轮 legibility bake-off（可读性比拼），开始研究安全问题。如果你批准，我们就计划推进这个项目，在 Q4 推出 beta。你觉得我最新的这份工作成果怎么样？

[34:56] Wade

Well, uh I mean, based on the critique I gave you in the last one, you basically nailed the two things I called out, which was this should be a prototype and where's the evidence? So, like obviously right off the bat, like this is so much stronger. Um, in fact, it makes me even wonder where I'm like, is Akos actually running this live transcript into a coding agent and it's live fixing this stuff, which I have seen people do uh at Zapier and that is always like feels magical where it's literally you're just taking like what the person is telling you and you're literally feeding it right back to them. And so when they see it come up, they're like, "Oh, great. That's amazing." And so, uh, I don't know if that's what you did or if you just anticipated that that would be my feedback and, you know, guess it on the way. Either way, like that is the magic of like, you know, good product judgment. So, you know, I think you you sort of immediately cleared those things up. The next thing uh that is honestly where I would push you is the ask. You say, "Hey, I I want to prove this. I want to ship, you know, uh I want to get these design partners and I want to get to a Q4 beta. [snorts] This is where I'm like, why are we waiting for Q4? Like, you've already gotten all this far. Like, can you do it faster? Uh and I think this is one again one of the things that's pretty magical about these AI tools is you can go from zero to design partner exceptionally fast. And so, you know, if we're talking Q4, that's uh October. And it's like, well, shoot, you got you got five weeks.

嗯，呃，我是说，对照我在上一份里给你的批评，我点出的两件事你基本都做到位了：一是这个应该做成原型，二是证据在哪儿？所以很明显，一上来这份就强太多了。呃，事实上，这甚至让我怀疑：Aakash 是不是真的把这段实时转写喂进了某个 coding agent，让它现场把这些东西改好？我在 Zapier 见过有人这么干，每次都感觉很神奇——你就是把对方正在跟你说的话，原样直接喂回给对方。所以对方一看到它冒出来，就会说：「哦，太好了，太厉害了。」所以，呃，我不知道你是这么干的，还是你只是预判到我会给这个反馈，在来的路上就猜到了。不管哪种，这正是好的产品判断力的神奇之处。所以我觉得你一下子就把那些问题解决了。接下来，说实话，我会逼你一把的地方是你提的请求。你说：「嘿，我想验证这个，我想发布，呃，我想拿下这些设计合作伙伴，想在 Q4 推出 beta。」（笑）我这时候就会想：为什么要等到 Q4？你都已经走了这么远了，能不能更快一点？呃，我觉得这又是这些 AI 工具相当神奇的一点：从零到拿下设计合作伙伴，你可以快得出奇。所以你看，如果我们说的是 Q4，那就是，呃，10 月。那就是说，好家伙，还有五周呢。

[36:27] Wade

Like, you know, what are we doing for five weeks? you might be able to get there a lot faster than that. Um, and so that would be the main thing I would start to push on is like how do we get these iteration cycles to be going a lot lot faster, but by and large like this is much better than before.

这五周我们要干什么？你可能远用不了这么久就能做到。呃，所以这会是我开始重点追问的地方：怎么让这些迭代周期转得快得多、再快得多。但总的来说，这份比之前好太多了。

[36:42] Aakash

Awesome. And I'll just note out a couple things or like what would you give the final rating on this?

太好了。我想点出几件事……或者说，你最终会给这份打什么评级？

[36:47] Wade

I I think this is like like this this feels to me like solidly adaptive. uh like I I would be very happy to have like this caliber of PM work sort of across Zapier at this moment in time. I do think there are you know like the thing if you were actually like live listening to my feedback and feeding it back in like that's an example where I've started to see some people doing it but I don't think that that's a tactic that's like widely used yet. That's starting to look pretty transformative. So if you were doing that like there I'd say like nicely done. Uh but by and large this is like very solidly adoptive I would feel like.

我觉得这份……给我的感觉是稳稳的「适应型」。呃，就眼下这个阶段来说，如果整个 Zapier 的 PM 工作都能达到这个水准，我会非常满意。我确实觉得，如果你真的是在现场听我的反馈、再把它喂回去，那就是一个例子：我已经开始看到有人这么做了，但我不觉得这是一种已经被广泛使用的打法。那就开始相当接近「变革型」了。所以如果你真是这么做的，那我会说：干得漂亮。呃，但总的来说，我觉得这份是非常扎实的「适应型」。

[37:22] Aakash

Okay. I was going for capable. I didn't make it. The only thing I did uh live was I added in

好吧。我本来是冲着「胜任」去的，没做到。我现场唯一做的一件事，就是加进去了

[37:28] Aakash

like really clearly what I got from Gong. And so you can even see that [laughter] that prompt that came in just a few minutes before.

就是非常清楚地写明我从 Gong 拿到了什么。所以你们甚至能看到（笑）几分钟前刚输进去的那条 prompt。

[37:37] Wade

There you go.

瞧，就是这个。

[37:38]

Um to say like pretend should pretend looked at Gong call. Well, that that's what the dead giveaway was was like the it listed out Gong and Zindesk and I was like those are the tools I cited like [laughter] I was like I I think this is an AI like hyper listening to me because you can kind of you can kind of tell where it's like AI just has that hyper specificity to it uh a lot of the times.

Aakash： 呃，就是跟它说，假装……要假装看过 Gong 的通话录音。

Wade： 嗯，这就是露馅的地方：它把 Gong 和 Zendesk 都列了出来，我一看，这不正是我刚才举的那几个工具嘛（笑）。我当时就想，我觉得这是个 AI 在一字不漏地听我说话，因为你多少能看得出来，AI 很多时候就是带着那种过分具体的劲儿。

[38:03] Aakash

So the first one was somewhere between adop unacceptable and adoptive. This one hopefully is somewhere between adoptive and capable. Let's see where this last one lands where you can kind of see my agent title kind of gives it away of what I'm trying to achieve. And then what we're seeing is if Wade actually feels that which I would say for my last one, he didn't give it an outright capable. So that shows you guys that the bar for capable. Like even though I built this inside a pretty advanced claude code harness, I think the writing was quite a bit better than the last one. There was pretty good evidence. There was a pretty decent prototype. There's still a pretty high bar for capable. So, let's see. Can I hit transformative? So, what I did here is I created two images to show um what my agent stack looks like. So, if I go into here, it looks like it's this is my agent stack. And so the whole idea is that I have this five layer sort of architecture where I have these skills for our team's recurring work like PRD, prep, eval, recall, ingest, impact sizing, launch checklist, weekly review and 37 more. Then we have this orchestration layer where sub aents fan out in parallel. So there's engineering, design, exec, legal, UX, trust and safety, and skeptic, which argues everything shouldn't ship because we know like we don't want feature bloat.

所以第一份大概介于不合格和适应型之间，这一份希望能落在适应型和胜任之间。来看看最后这一份会落在哪一档——你们从我这个 agent 的标题大概就能看出我想冲什么了。然后就看 Wade 是不是真这么觉得——我得说，上一份他就没有干脆地给到胜任。这也让大家看到了胜任的门槛：哪怕我是在一个相当高级的 Claude Code harness 里做的，文字我觉得也比上一份好了不少，证据挺扎实，原型也相当像样，胜任的门槛依然相当高。那来看看，我能不能冲到变革型？我这次做了两张图，展示我的 agent 栈长什么样。点进来看，这就是我的 agent 栈。整体思路是一个五层架构：有一组 skill，专门处理团队反复要做的工作，比如 PRD、prep、eval、recall、ingest、影响测算、上线检查清单、每周复盘，另外还有 37 个。然后是编排层，subagent 在这里并行地分头展开：有工程、设计、高管、法务、UX、信任与安全，还有 skeptic（唱反调的怀疑派），它专门主张什么都不该上线，因为我们都知道，我们不想要功能膨胀。

[39:37] Aakash

Layer three is the tools in MCP. So of course we use Zenesk and Gong which was that was the only update I made

第三层是工具和 MCP。当然，我们用了 Zendesk 和 Gong，这是我唯一改动的地方，

[39:45] Aakash

once I heard it from you. Reddit, I don't know if you guys use amplitude and linear, but those are the assumptions I made in terms of analytics and ticketing. Figma for design, although maybe you guys are all into cloud design now, GitHub and Chrome. And then I think this is like the most important thing is that there's this deterministic enforcement in this environment. Um, and so what happens is it will post any tool use, it will reject a memory file if the evidence has no source. And so everything that we put into memory will have really good sourcing and it will stop the session from ending with our brain being dirty. So the brain is like constantly updated and that feeds into this memory layer where we have all these sourcing of everything we're learning in product across the our product team. We have this ingestion

是听你说完之后才加的。还有 Reddit。我不知道你们用不用 Amplitude 和 Linear，不过在数据分析和工单这两块，我就是这么假设的。设计用 Figma——虽然说不定你们现在都在用 Claude Design 了——还有 GitHub 和 Chrome。然后，我觉得最重要的一点是，这个环境里有一层确定性的强制约束。具体来说，在每次工具调用之后（PostToolUse），只要证据没有出处，它就会拒收这份记忆文件。所以我们放进记忆里的每样东西都会有扎实的出处；而且只要大脑还处在「脏」状态，它就不让会话结束。所以这个大脑会一直保持更新，并且喂进记忆层——整个产品团队在产品上学到的所有东西，出处都记录在那里。我们有这样一个导入环节，

[40:40] Aakash

and automatically our agents are creating these are the hypotheses that are coming let's say out of zenesk plus gong plus reic. These are the decisions that are coming out of let's say linear tickets and GitHub PRs. This is the research that we found from Gong and Reddit. These are what the stakeholders are saying from meeting notes. So this is my agent setup. This is part one and there's one more thing I want to show you before you give me my rating. So the second part is like the auto pipeline that I've created around it. So I basically have a a loop running nightly. So we take our raw customer signals. Then we have these auto agents that run on a nightly cron job. So I call this the triage swarm. They're going to ddup and cluster by job. They're going to of course populate our source folder which hopefully understand now why that's so important. They're going to tag things as observations or interpretation or a hypothesis. They're going to match things versus open hypotheses and flag contradictions. So overnight, for example, we might get like nine clusters of in insights from all of our signal signals in maybe two of those are new. Like maybe something came up on a Reddit forum. Maybe one even contradicts a bet made in March. All of that the triage swarm would just identify for us. Then what happens next is using the memory layer I would come in as the PM and I would say do we do we promote this right are there independent sources is it just one person on Reddit we don't care but is Reddit and our amplitude data showing it maybe then it's relevant so in four to five minutes I look at that the system itself is going to help me document the hypotheses and decisions and then I think this is the part that hopefully makes it a little more powerful than what I've shown before is we have this auto build system. So each step is going to need the layer above, but first it starts with a quick

然后我们的 agent 会自动生成这些内容：比如，这些是从 Zendesk 加 Gong 加 Reddit 里得出的假设；这些是从 Linear 工单和 GitHub PR 里得出的决策；这是我们从 Gong 和 Reddit 里找到的研究；这些是会议纪要里各方干系人的说法。这就是我的 agent 配置，这是第一部分。在你给我评级之前，我还想再给你看一样东西。第二部分是我围绕它搭的一条自动流水线。基本上，我有一个每晚都会跑的循环。我们把原始的客户信号拿进来，然后有一批自动 agent，靠每晚的 cron 定时任务运行，我管它叫分诊 swarm（一群并行跑的 agent）。它们会先去重，再按 job（用户想完成的任务）聚类。当然，它们也会往我们的 source 文件夹里填内容——希望大家现在明白这为什么那么重要了。它们会给内容打上标签：是观察、解读，还是假设。它们还会拿这些内容去比对待验证的假设，并把矛盾标出来。比如说，一夜之间，我们可能从所有信号里得到九簇洞察，其中也许有两簇是新的：可能是 Reddit 论坛上冒出了什么东西，甚至可能有一簇跟我们三月份下的某个赌注相矛盾。这些分诊 swarm 都会替我们识别出来。接下来，我作为 PM 会借助记忆层介入，判断：这条要不要往上提？有没有相互独立的来源？如果只是 Reddit 上某一个人说的，我们不在乎；但如果 Reddit 和我们的 Amplitude 数据都显示出来了，那可能就值得关注了。所以我花四五分钟过一遍，系统本身就会帮我把假设和决策记录下来。然后，我觉得接下来这部分应该能让它比我之前展示的更强一点：我们有一套自动构建系统。每一步都要依赖上面那一层，不过第一步是先快速写一份

[42:42] Aakash

two-page PRD. Then it does a PRD review panel. Then it creates the prototype. Then it creates the eval. And then it creates the code first draft. We get that with the prototype link, the user tests, the draft PR, and the linear tickets. And that's when we have the next human gate. So human gate one was do we care about this build chain? Human gate two is do we merge it? And if not, and once it's done, we put it back into the memory. And there's some hooks that run it all. But the idea is that we can go from, let's say, a 10-day development process to a two-day development process. So, if you saw me running a product team like this, how would you grade my output, Wade?

两页的 PRD。接着过一轮 PRD 评审小组，然后做原型，然后写 eval，再然后产出代码初稿。最后我们拿到的是原型链接、用户测试、草稿 PR，还有 Linear 工单。到这一步，就来到了下一道人工关卡（human gate）。第一道人工关卡是：这条构建链值不值得我们关心？第二道人工关卡是：要不要合并？如果不合并——或者整个流程跑完之后——我们都会把结果放回记忆里。整套东西由一些 hook 来驱动。核心想法是，我们可以把比如说 10 天的开发流程，压缩到两天。所以 Wade，如果你看到我这样带一个产品团队，你会给我的产出评什么级？

[43:22] Wade

This is definitely starting to like get into the transformative territory. Like, this is just not how, you know, uh, product and software was built, you know, three years ago. like not even close. Uh it is a totally different way of working. Um you know it looks a lot more like a factory, right? You're you know you're sort of saying hey my job as a PM or an engineer or whatever your role is in this system is not to do each of these individual tasks. You've set up agents who are handling these tasks and now your job is to actually audit the system. Like when the agent fails, okay, why did the agent fail at that task? You know, if you think about like a factory, right? Um if a factory produces a failed widget, what does the factory person do? Do they like try and like individually fix the widgets? No, not really. They just sort of discard that batch and they go back and look at the machine and say like why did the machine generate these failed widgets? Like what's what's going on? And they fix that. And so, you know, this is where the job is kind of shifting to. It's like oh you know instead of having human machines that are sort of like stamping out you know each step of the process we're replacing those and now the job of you know the humans is to sort of audit the overall system and say hey is this working or not. It's a very different way of running a software and product team. Uh now the question that I think all of us is trying to figure out is what is the right way to run this factory and that's where I was like listening intently to see like hey how are you designing your system here because I think it's it like I talked to a lot of folks building these types of things and I don't think there is a consensus yet on like what is the exact best way to run these loops.

这绝对已经开始进入变革型的范畴了。三年前，产品和软件根本不是这么做出来的，差得远了。这是一种完全不同的工作方式。它看起来更像一座工厂，对吧？你等于是在说：我作为 PM、工程师，或者不管我在这个系统里是什么角色，我的工作都不是去亲手完成每一项具体任务。你已经让 agent 来处理这些任务了，现在你的工作其实是审查这个系统。比如 agent 失败了，好，那它为什么在这个任务上失败？想想一座工厂，对吧？如果工厂生产出一个残次零件，工厂里的人会怎么做？他们会一个个去修这些零件吗？不太会。他们会直接把那一批报废，然后回头去检查机器：为什么这台机器会生产出这些残次品？到底出了什么问题？然后把它修好。所以，工作的重心正在往这边转移。也就是说，以前是人像机器一样，把流程里的每一步一件件冲压出来；现在我们把这些换掉了，人的工作变成了审查整个系统，判断它到底行不行。这是一种非常不同的软件和产品团队运作方式。现在，我觉得我们所有人都在琢磨的问题是：这座工厂到底该怎么运转才对。所以我刚才一直在很认真地听，想看看你这套系统是怎么设计的。因为我跟很多在搭这类东西的人聊过，我觉得在「这些循环究竟怎么跑最好」这件事上，目前还没有形成共识。

[45:05] Wade

Um for example the thing I was trying to figure out is like where are you putting those human gates and why are you deciding to put those human gates in those places versus other places? Um, maybe this is the correct place to put them. Maybe not. Maybe you want to actually take the the first human gate out and say actually I I want to kick off coding agents for everything and I'm only and then I'm going to review all the stuff that sort of comes through. Maybe you don't want to do that because you think actually like a lot of these things are burning lots of tokens and we're spending a lot of money on things that could be wasteful and so it's worth it to have a human gate there even though it might slow the overall system down. Um but those are things that like I think a lot of teams are trying to figure out is like hey is this the right way to to do this and candidly a lot of people aren't to this yet. Um but you you know it is becoming more and more common so you do see it. Uh but I think the exact design of this this this loop is what I see a lot of teams trying to like scratching their heads on trying to figure out like how do they get it get it working um as best as possible. So like this is starting to tick into the the the transformative the bucket.

比如，我刚才一直想弄明白的是：你把这些人工关卡放在哪儿？为什么决定放在这些位置，而不是别的地方？也许放在这儿是对的，也许不是。也许你其实想把第一道人工关卡拿掉，直接说：我要对所有事情都启动 coding agent，然后我只——然后我再去审查所有跑出来的东西。也许你不想这么做，因为你觉得很多这类东西在烧大量的 token，我们在可能纯属浪费的事情上花了很多钱，所以哪怕会拖慢整个系统，在那里设一道人工关卡也是值得的。但我觉得这些正是很多团队在摸索的问题：这样做到底对不对？而且坦白说，很多人还没走到这一步。不过它正变得越来越普遍，所以你确实能看到。但我觉得，这个循环具体该怎么设计，是我看到很多团队正挠头琢磨的事——怎样才能让它尽可能跑顺。所以说，这已经开始迈进变革型那一档了。

[46:07] Aakash

Awesome. So, if a PM really wants to get that rating, although you said at the beginning that it you don't actually need to, right? Like maybe if you're in the transformative, you're too focused on the AI tools, but let's say somebody's manager has told them, you know, you should you should be going for that transformative rating.

太好了。那如果一个 PM 真的想拿到这个评级——虽然你一开始说过其实没必要，对吧？比如说，如果你一直处在变革型，可能就是太专注于 AI 工具本身了——但假设某个人的经理跟他说：你应该冲一冲变革型这个评级。

[46:25] Aakash

If in 2026, they start to build a system that operates like this that actually works, they are getting to that transformative level.

那在 2026 年，如果他们开始搭一套像这样运转、而且真正跑得通的系统，就算达到变革型的水平了？

[46:33] Wade

Yes. Well, and the thing that I would call out is at any at at a small company, like if you're the only PM at the company, then yeah, you're probably building out a system like this because who else will do it? You have to be the one that does this. But if you're getting to company of Zapier size where we have, you know, quite a few PMs working here, not everyone's job is to sort of come up with this thing, buddes the system. You probably have a team that's saying, "Okay, I'm going to sort of go build this." uh or maybe we got a couple teams doing a few variants of it and then the where the rest of the PMs come in is like okay now help me figure out how to run this system right and so you know you've got a handful of folks that are sort of in the transforming part where they're like okay I got to go sort of re redo all this stuff well in the meantime you know we've kind of got some PMs that are sort of running the old standard playbook uh and you know occasionally sort of coming in and being like the user testers for like this new way of working uh and you can sort of parallel process like uh you know this this new way of working. Uh and so you know it it sort of just depends a little on like how your organization is set up.

对。不过我想特别指出的一点是：在小公司，比如你是公司里唯一的 PM，那没错，你大概就得自己搭一套这样的系统，因为还有谁会做呢？只能是你来做。但如果公司到了 Zapier 这样的规模，我们这儿有相当多的 PM，那就不是每个人的工作都是去琢磨、去搭这套系统。你们大概会有一个团队说：「好，我来搭这个。」或者可能有几个团队在做它的几个变体。然后其余 PM 的切入点是：好，现在帮我想想这套系统该怎么运转。所以你会有少数几个人处在「变革」的那一部分，他们会说：好，我得把这些东西全部重做一遍。与此同时，我们也有一些 PM 还在按老一套标准打法做事，偶尔过来当一下这套新工作方式的用户测试员。这样一来，你就能并行推进这套新的工作方式。所以说，这多少要看你的组织是怎么搭建的。

[47:38]

But that's where I look at it. I'm like I don't need all of my PMS to sort of drop everything and rebuild this system. I don't want you know I think we have like something like 30 or 40 PMs in Zapper. I don't need 30 systems of this loop running. Like I actually want like one sort of way of doing this stuff. or maybe maybe I want a couple like depending on there's a few different jobs inside of Zapier and so they might need to work like slightly differently um for these different jobs. Um but but the goal is not like everyone runs their own mini version of this. There is a sort of like Zap Zapier product factory uh that does a lot of this. All right. So now you guys have gotten the road map. You understand as a PM how to rate on AI fluency. I want to see a little bit more of Zapier's internal sauce. Can you show us a little bit of your internal agent dashboards and how you guys are starting to work towards these ways of working?

Wade： 但我就是这么看的。我不需要我所有的 PM 都放下手头的一切去重建这套系统。我不希望——我记得我们 Zapier 大概有 30 到 40 个 PM——我不需要 30 套这样的循环同时在跑。我其实希望只有一种做法，或者也许要几种，因为 Zapier 内部有几类不同的工作，针对这些不同的工作，运转方式可能得稍有不同。但目标不是每个人各跑一个自己的迷你版本，而是会有一个类似 Zapier 产品工厂的东西，来承担这里面的很多工作。

Aakash： 好。现在大家已经拿到路线图了，也明白了作为 PM，AI 流利度是怎么评级的。我想再多看一点 Zapier 内部的独门秘方。你能给我们展示一下你们内部的 agent 仪表盘，以及你们是怎么一步步朝这些工作方式迈进的吗？

[48:32] Wade

Yeah, maybe let me just show my some of my like individual use cases. Uh this is what I do the most of with uh agents today. So these are not like Zapier production systems. These are just like weighed personal agents. Um, and so I've, you know, built out like a little, uh, dashboard here that I can sort of show off some of the, um, you know, simple agents that I have set up, uh, that sort of, you know, kind of span, uh, the gamut of, you know, simple use cases to maybe more sophisticated. Um, I have I I bucketed them into like I I had my agent build out this this demo page for us. Uh, but I've kind of bucketed them into a few things. I've got my robot uh, staff. So, these are like actual agents that are helping me run my like day-to-day systems. I have this like CEO CRM. This is sort of a lot closer to like kind of the the production agent that we showed off just now. Uh I have this AI that argues back. This is kind of like a simple thing that you can do and you probably should do. You know, it's it's akin to the skeptic that you sort of set up in your, you know, uh uh uh product loop that you have. And then I have just a bunch of executable skills that I do for like you know real-time assistant stuff that's helping me with like a whole variety of cases. Um maybe what would be most interesting is we could sort of poke into the robot staff first. What do you think?

好，也许我就展示一下我自己的一些个人用例吧。这是我目前用 agent 做得最多的事。所以这些不是 Zapier 的生产系统，只是我 Wade 个人的 agent。我在这儿搭了一个小仪表盘，可以展示一下我配置的一些简单 agent，从简单的用例到可能更复杂一些的都有。我把它们分成了——这个演示页面是我让我的 agent 给我们搭的——总之我大致把它们分成了几类。一类是我的机器人员工（robot staff），这些是真正在帮我运转日常事务的 agent。还有这个 CEO CRM，它跟我们刚才展示的那种生产级 agent 要接近得多。还有这个会跟我唱反调的 AI，这是个简单的东西，你可以做，而且大概也应该做。它有点像你在自己那套产品循环里设的那个 skeptic。然后还有一堆可执行的 skill，处理各种实时助理类的事务，帮我应对五花八门的场景。也许最有意思的是，我们先看看机器人员工。你觉得呢？

[49:50] Aakash

Yeah, let's do it.

好啊，来吧。

[49:51] Wade

So the robot staff. So you know here's five of the more commonly things that I have set up. Um you know the one that I always sort of suggest for folks who are just getting into this. The place to start morning brief. It's kind of a classic. Uh it's uh you know pretty simplistic to set up. you hook up an agent to your calendar, your inbox, your Slack, all that other sort of stuff, and then just say, "Hey, I want you to go build out um a daily brief for me every morning." Very simple place to start. You know, especially if you're not used to building agents or not used to sort of connecting it to tools, it is a very easy one to do. Uh I have mine run at, you know, at 6 a.m. every morning. You can actually see uh how it runs uh inside of Zapier. So, the nice thing is I have this running uh on Zapier. And uh what is handy about running these workflows on Zapier versus you know having uh you know cloud code or codecs or any of these tool run them is that Zap year is going to run them deterministically. So you can see it's you know kicking off at 6 a.m. every day. It's fetching these calendar events. This is not an agent running this. This is code. Uh so you can actually poke over here and see there's code that's running a lot of this stuff.

那就先看机器人幕僚团（robot staff）。这是我搭好的五个比较常用的 agent。嗯，我总推荐给刚入门的人的那个——起点就是晨间简报（morning brief）。它算是个经典款，搭起来相当简单：把一个 agent 接上你的日历、收件箱、Slack 之类的东西，然后跟它说：「嘿，我要你每天早上给我整理一份每日简报。」起步非常简单。尤其如果你还不习惯搭 agent，或者不习惯把它接到各种工具上，这个做起来非常容易。我的设成每天早上 6 点跑。你们其实能看到它在 Zapier 里是怎么跑的。好处在于我是把它放在 Zapier 上跑的。把这些工作流放在 Zapier 上跑，而不是交给 Claude Code、Codex 这类工具去跑，方便之处在于 Zapier 会以确定性的方式运行它们。你们看，它每天早上 6 点启动，去拉取这些日历事件。跑这一步的不是 agent，是代码。你们可以点开这里看看，这里面很多步骤都是代码在跑。

[50:57] Wade

And the reason you want to run these things deterministically is that it's reliable. It runs the same way every single time. time and it's way lower cost these steps. None of this is burning tokens. You only burn tokens right here on this stage where it's trying to summarize and like actually draft uh the daily gig. But then when it posts to Slack, again, that's just a sort of straight up uh API call that it's doing. So

之所以要确定性地跑这些东西，是因为它可靠：每一次都以完全相同的方式运行。而且这些步骤的成本低得多，全都不烧 token。只有在这一步——它要做总结、真正起草那份每日简报的时候——才会烧 token。之后发到 Slack 那一步，同样只是一个直截了当的 API 调用。所以

[51:22] Aakash

devil's advocate here just so I we can understand how people should engineer this because I think this is actually one of the most important agents people need to get, right? Couldn't we this miss out on some context because it just has Google calendar?

我在这儿唱个反调，好让大家明白该怎么设计这个东西——因为我觉得这其实是大家最需要搭好的 agent 之一，对吧？它只接了 Google 日历，会不会漏掉一些上下文？

[51:36] Wade

So, uh, a lot of what's it's actually doing here is uh it's in the code step, but it's like hitting um my to-do list. Uh, it's hitting um inside of calendar, it also has uh a lot of times there's like documents added to it. So, it's clicking through those links, fetching those documents that are

嗯，它在这里做的很多事其实都在代码那一步里，比如去调取我的待办清单；在日历里面，很多时候日程上还附了文档，所以它会点进那些链接，去抓取那些文档，就是那些

[51:55] Aakash

So, it has that power. Got it.

所以它有这个能力。明白了。

[51:57] Wade

Exactly. Uh, it has access to um what else do I give do I have it uh have it on? Uh I mean I think calendar in like your your to-do list is probably like the two biggest things you want to give it access to. Um but it can sort of click through links and click through attached assets and things like that and then sort of summarize those. Um so that's where it's you know you know having these sort of like quasi deterministic uh agents that can also sort of fall back to using of more like agentic loop more like inference are really powerful. You get you know determinism where you care about it but you get the power of like an agent reasoning through the task where you need it most.

没错。它能访问……我还给它接了什么来着？我觉得日历和你的待办清单，大概是你最该给它开权限的两样东西。但它也能点进链接、点开附件之类的东西，然后做个总结。所以这就是为什么这种半确定性的 agent 特别强大——必要时它还能退回去用更像 agentic loop、更偏推理（inference）的方式。在你在乎的地方，你拿到确定性；在最需要的地方，你又能用上 agent 边推理边把任务做完的能力。

[52:36] Aakash

And I think the probably the best thing that I've seen with Zapier is you can actually prompt an AI agent to create these workflows for you. So you don't actually need to code it yourself.

我觉得我在 Zapier 身上看到的最棒的一点，大概是你其实可以直接给 AI agent 下 prompt，让它替你把这些工作流搭出来，所以根本不用自己写代码。

[52:45] Wade

Exactly. In fact, these that uh are built inside of I use Cursor and so I'm mostly just talking to Cursor and have them build them, but then I'm sort of deploying it to to Zapier to run. Uh and so that way I can close my laptop at night. Uh I can have it run deterministically. I don't have to you know be doing a lot of the um you know kind of clunkiness that a lot of these uh you know uh agent products make you do right now because everything's running locally and it's running you know you're burning lots of tokens um to do it.

没错。事实上，这些是在……我用的是 Cursor，所以我基本上就是跟 Cursor 对话，让它把这些搭出来，然后再部署到 Zapier 上去跑。这样一来，我晚上可以合上笔记本，让它确定性地跑着。我不用去忍受现在很多 agent 产品逼你做的那些笨拙操作——因为一切都在本地跑，而且跑起来要烧掉大量 token。

[53:12] Aakash

Yeah. It's like a managed automation in the cloud essentially.

对，本质上就是一个云端托管的自动化。

[53:17] Wade

You got it. You got it. What else do I have? This is so, you know, as much as I tell people to start with the morning brief um because I think it's a simple easy place to start, I actually think it's only modestly valuable. My favorite is the scribe, which is uh I have this set up at the end of my day. You can also run it after meetings, but uh you know, in my case, I'm having it run at 5:00 p.m. Uh another way to think about this is this is the daily recap. the daily recap way more valuable. So, this thing loops over all of my outstanding emails, all of the outstanding to like save for later messages I have in Slack. And the biggie loops over all of my granola transcripts and it sort of sucks out and says, "Here's all the things you need to complete uh based on today, and I'm going to go ahead and draft uh anything that you need for me." So, I'll actually draft the follow-up emails. Like, if you know, you and I were talking and you said, "Wade, can you intro me to soand so?" Okay, great. The granola catches that and says, "Okay, uh, we're going to go ahead and draft a follow-up email for this person." Or, you know, this outstanding action item is in your to-do list. Do you want me to go ahead and take a stab at authoring these things? And so the scribe catches all of that stuff from the day and helps me like do the end of day cleanup that used to take me gosh it probably took me two hours at the end of every day to just like you know get to most of the the stuff that just kind of piles up. I mean PMs know this like PMs just like you just aggregate to-dos all day every day. Um some are really important some are just like things you want to do to be a good teammate and you know it's like how do you fit those things in? And in the past, you would just kind of you just kind of muscle through them. And so the scribe helps me muscle through those way faster. I would say my end of day recap work is now

你说对了，你说对了。我还有什么来着？虽然我总跟人说从晨间简报起步——因为我觉得它简单、好上手——但我其实觉得它的价值一般。我最喜欢的是书记员（the scribe），我把它设在每天收工的时候跑。你也可以在开完会后跑，不过我是让它下午 5 点跑。换个角度想，这就是每日复盘（daily recap），而每日复盘的价值要大得多。这东西会把我所有还没处理的邮件过一遍，把我在 Slack 里标了「稍后处理」（save for later）的所有消息也过一遍。重头戏是，它会把我所有的 Granola 会议转录过一遍，从里面提炼出来，然后说：「根据今天的情况，这些是你需要完成的所有事项，凡是你需要的，我都先帮你起草好。」所以它真的会把跟进邮件都起草好。比如你我聊天时你说：「Wade，能帮我引荐一下某某吗？」好，没问题。Granola 会捕捉到这句话，然后它会说：「好，我们来给这个人起草一封跟进邮件。」或者：你待办清单里有这么一条还没完成的事项，要不要我先试着帮你写起来？所以书记员会把一天里的这些事全都兜住，帮我做收工前的清理。这活以前要花我——天哪，大概每天收工都得花两个小时，才能把堆起来的大部分杂事处理掉。PM 都懂这个，PM 就是一天到晚不停地攒待办。有些真的很重要，有些只是为了做个好队友而想做的事，问题是你怎么把这些事塞进日程里？以前你就只能硬着头皮一件件啃过去。现在书记员帮我啃得快多了。我会说，我每天收工复盘的活现在

[55:06] Wade

probably 15 minutes instead of two hours. So it is a big like quality of life improvement for me. Um at the end of the day,

大概只要 15 分钟，而不是两个小时。所以对我来说，这是生活质量的一次大提升。嗯，一天下来，

[55:13] Aakash

I was wondering how you managed to keep up with my emails about a podcast as a CEO. There you guys go. [laughter]

我还一直纳闷，你一个 CEO 怎么还跟得上我那些约播客的邮件。这下大家知道答案了。（笑）

[55:19] Wade

There you go. You know, here's a fun one. So you know, you probably run a, you know, weekly meeting as a PM. uh and you know how do you set the agenda? What's the important things to talk about? Well, um I have a human in the loop myself helping with the agenda but by and large I'm using AI to help me generate all the agenda items. Now, uh I do this for our exec meeting. I do this for our board meetings. Um but basically it again it loops over all my granola meetings from the week. It loops over Slack. It loops over email. Loops over all my AI sessions. So like my chat sessions that I have with uh any of my AI tools and it says here are the most important things that are still outstanding that you need to talk about and raise. And so I still audit and sort of approve or disprove these things but I find that this makes sure our exec agenda is way more relevant uh than uh it is in in the past and also uh it generally has um better judgment on what are the topics we should be talking about. I think one of the toughest things when you're, you know, in charge of coming up a weekly agenda is people sort of bring their pet items, but are the pet items actually the most important things? Maybe, maybe not. But if you have it looping over, you know, all the metrics from the last week, the AI is going to say, you know what, your conversion rate for this was really crap last week. You need to talk about this.

可不是嘛。再来一个好玩的。作为 PM，你大概会主持一个周会，那议程怎么定？哪些是要紧、该拿出来聊的事？嗯，定议程这件事，我自己会作为 human in the loop（人工把关）参与，但大体上，所有议程条目我都是用 AI 帮我生成的。我们的高管会我这么做，董事会我也这么做。基本上，它同样会把我这一周所有的 Granola 会议记录过一遍，把 Slack 过一遍，把邮件过一遍，把我所有的 AI 会话也过一遍——也就是我跟各种 AI 工具的聊天记录——然后它会说：这些是仍然悬而未决、你需要拿出来讨论的最重要的事。我仍然会审一遍，决定批准还是否掉，但我发现这样能确保我们高管会的议程比过去切题得多，而且在「我们该讨论哪些议题」这件事上，它的判断通常也更好。我觉得，当你负责定每周议程时，最难的事情之一是：大家都会带着各自的心头好议题来，可这些心头好真的是最重要的事吗？也许是，也许不是。但如果你让它把上周所有的指标过一遍，AI 就会说：你知道吗，你们这块上周的转化率烂透了，你们得讨论一下这个。

[56:41] Wade

Don't let people get away with not raising this thing. And so the most important topics tend to be um better if you have one of these agents running over all your stuff every week is and pointed at the right data sources uh and you give it some ability to sort of uh call out like these topics need to be addressed. Um so those are some of the like more fun uh you know items in my robot staff. Um maybe the next one that's worth talking about is this prospector one. But with that, uh, maybe let's go show off the CEO CRM. Uh, because the CEO CRM is one of my, uh, more fun little projects. So, everyone says, hey, you know, don't, there's that debate online around like, should you vibe code your CRM or not? And, and I am very firmly in the camp of don't vibe code your CRM. Uh, but yet here I am. I have vive coded a CRM.

别让大家就这么躲过去、不把这事摆上桌。所以，如果你每周都有这么一个 agent 把你所有的东西过一遍，对准正确的数据源，再给它一点能力去点名「这些议题必须处理」，那么最重要的议题往往会挑得更好。这些就是我机器人幕僚团里比较好玩的几个。接下来也许值得聊聊这个拓客员（prospector）。不过说到这儿，要不我们去秀一下 CEO CRM 吧，因为 CEO CRM 是我比较好玩的小项目之一。大家都说，嘿，别……网上一直有个争论：你到底该不该 vibe code 一个自己的 CRM？我非常坚定地站在「别 vibe code 你的 CRM」这一边。可结果呢，我自己却 vibe code 了一个 CRM。

[57:32] Aakash

Oh, man.

好家伙。

[57:33] Wade

Uh, now look, this is super basic, right? This is not anywhere close to what you get from like HubSpot or Salesforce or any of that stuff out of the box. What this really is is an agent that handles a very set of specific workflows that I personally care about. Now, you're looking at a demo version of it, but this is what my real sort of CEO CRM looks like. Here's the deal. I don't actually look at this all that often because I actually started by building it as a CRM to look at but mostly now it runs as agent and I only showing the UI because it's hard to demo agents without showing you something. So what this is is it kind of, you know, it has access to like here's all the enterprise accounts we have. Here's how much ARR is going on in them. You know, it kind of flags what needs attention for me based on a whole bunch of different signals like, you know, is there a renewal coming up? Is there like usage dropage happening in the account? Like what's what's going on in these things? Of course, I can poke into, you know, the account and start to take a look at certain stuff where it'll show me like here are the apps that they're using, you know, here's a little bit about the account, here's the people on it, etc.

好吧，听着，这东西超级基础，对吧？跟 HubSpot、Salesforce 那类开箱即用的东西比，差得远了。它本质上是一个 agent，专门处理我个人在乎的一组非常具体的工作流。你们现在看到的是演示版，不过我真正在用的 CEO CRM 大致就长这样。是这样的：其实我并不常看这个界面。我一开始是把它当成一个拿来看的 CRM 去搭的，但现在它主要是作为 agent 在跑；我之所以展示这个 UI，只是因为不给你们看点东西的话，agent 很难演示。它大概是这样：它能访问我们所有的企业客户，每个客户身上有多少 ARR（年度经常性收入）。它会根据一大堆不同的信号，帮我标出哪些需要关注：比如是不是快到续约了？这个客户的用量是不是在下滑？这些客户那边到底是什么情况？当然，我也可以点进某个客户去看具体情况，它会告诉我：这是他们在用的应用，这是这个客户的一些基本情况，这是对接的人，等等。

[58:36]

But what is really all this context is leading up to is I really just want to be talking to these customers more. And so who should I be talking to and why should I be talking to to them? And so what this agent really does for me now is every Saturday night it runs. And then when I wake up in the morning, I've got 10 to 20 emails drafted in my inbox for enterprise customers that I want to stay in touch with. and it tells me, here's why I think you should talk to this person this week. You know, it's got a renewal coming up. Usage has gone down or usage has gone up in a lot of ways, and here's what I think you should go talk to them about. So, it suggests an email. Now, you're seeing it in the CRM, but what this actually does for me is it's literally a draft in my email inbox. And so, all I have to do is just like take a look and say like, "Yep, that's someone I want to go talk to." And just fire off all these emails. If I make some quick edits, sometimes I will, but by and large, I'm just using this to do what would otherwise be like a pretty cumbersome workflow uh for me every week. And you know, it helps me keep up with our most important customers uh and build like a personal relationship with them. Uh which in the past would have been really tough for me to have to have done because I just, you know, you have a CSM team and the CSM team does all that other stuff. But does, you know, our CSMs are great, but what people really love hearing from is they love hearing from the CEO. And so now I can, you know, build a much stronger relationship with some of those most important customers. Uh, and it's in your voice, which is really cool. Like I think that's given our conversation about slop earlier. It's all lowercase.

Wade： 但所有这些上下文最终指向的，其实就是我想多跟这些客户聊聊。那我该跟谁聊？为什么要跟他们聊？所以这个 agent 现在真正替我做的事是：每周六晚上跑一次。等我早上醒来，收件箱里已经起草好了 10 到 20 封邮件，发给那些我想保持联系的企业客户。它还会告诉我：这就是我认为你这周该跟这个人聊一聊的原因——比如快续约了；用量下降了，或者在很多方面用量上涨了；以及我觉得你该跟他们聊些什么。所以它会给出一封建议的邮件。你们现在是在 CRM 里看到它，但它实际为我做的，是在我的邮箱里放一封真正的草稿。所以我要做的，就是看一眼，然后说：「嗯，这个人我确实想去聊聊。」再把这些邮件一口气发出去。有时候我会快速改几笔，但大体上，我就是用它来完成一个否则每周都会相当繁琐的工作流。它帮我跟上我们最重要的那些客户，并跟他们建立起私人关系。这在以前对我来说真的很难做到，因为——你知道的，公司有客户成功经理（CSM）团队，那些事都由他们来做。我们的客户成功经理都很棒，但大家真正喜欢的，是收到 CEO 本人的消息。所以现在我能跟其中一些最重要的客户建立起牢固得多的关系。

Aakash： 而且是用你的口吻写的，这点真的很酷。结合咱们前面聊的 slop，我觉得这点尤其值得一提。通篇都是小写。

[1:00:10] Aakash

It's written how you would write it. I think that's a key component that people probably need to iterate on their agent setup if they're going to want to have drafts created. make sure that it's writing to the right level.

写法就跟你本人写的一样。我觉得这是个关键环节：大家如果想让 agent 帮忙起草，大概都得反复打磨自己的 agent 配置，确保它写出来的东西拿捏到位。

[1:00:22] Wade

I'll tell you what, funny enough, that was the that has been the hardest part of this thing. Like building the CRM, getting all the agentic logic, all that stuff is is so easy compared to just trying to like bend these agents to to write like a human. Uh they just are not very good at that. And so, uh even still, it's doing okay here. I there's still things I think it could do better, but it is it is really tough. And I do actually have I don't have it in my demo, but I do have like an agent uh that is running analytics on this stuff too uh that's sort of helping me say like oh you know this type of writing or these types of subject lines tend to be working better than than these like maybe maybe do more of that uh versus the other. So uh you know it just gets a lot easier to do these types of things that I think you know we all this is the stuff that I think we all want to be doing. It's like man I really want to be spending more time with my customers. I really want to be, you know, trying to figure out how to write better emails. I really want to be doing this stuff. But, you know, if you're in a product job, you know, there's a million things that you want to be doing, but you just can't do them well enough as one person. And so, this is where I get so excited about what these AI and agents can do for you is you can sort of have a standard in your head of like, well, this is what great looks like. How do I use these tools to help me be great even when I'm just a human? like I can't like I just simply can't be that great all the time. So that's where I get really excited about a lot of this stuff.

跟你说吧，说来好笑，这恰恰是整个东西里最难的部分。搭 CRM、把整套 agentic 逻辑弄好，这些跟「硬掰着这些 agent 像人一样写东西」比起来，简直太容易了。它们在这方面就是不太行。所以就算到现在，它在这儿也只是做得还行。我觉得还是有些地方它可以做得更好，但这真的很难。而且我其实还有——演示里没放——一个专门对这些做数据分析的 agent，它会帮我看出：哦，这种写法、这类邮件标题的效果往往比那些好，那也许就多用这种、少用那种。所以做这类事情就容易多了——我觉得这些正是我们都想做的事。就像是：天哪，我真的很想多花点时间跟客户待在一起；我真的很想琢磨怎么把邮件写得更好；我真的很想做这些事。但如果你是做产品的，你想做的事有一百万件，可你一个人就是没法把它们都做得足够好。所以这就是 AI 和 agent 能为你做的事里最让我兴奋的地方：你脑子里可以有一个标准——嗯，优秀就该是这个样子。那我怎么用这些工具，让我即便只是个凡人，也能做到优秀？毕竟我没法、我就是不可能时时刻刻都那么优秀。所以这就是这些东西里让我特别兴奋的地方。

[1:01:54] Aakash

Amazing. That covers it for our show and tell segment you guys got to see and you personally only rated yourself as adoptive, right?

太精彩了。我们的实操展示环节就到这里，大家都看到了。而你本人只给自己打了个适应型，对吧？

[1:02:04] Wade

Yeah. I mean honestly I think like I'm adoptive in a lot of places. Um, you know, it I feel like I'm just like a super generalist with a lot of these things. And I have folks on the team who are way better at many of these things. And, you know, when I'm pushing into transformative, honestly, I'm just stealing from them. I'm just seeing like, what are you doing? How are you figuring this thing out? And I'm just taking the tactics and just applying them to my own workflows and my own, you know, uh, things that I'm doing. But, you know, the stuff that we're doing as a company and as a team, um, I there's people who are far smarter and figured out far more uh than me on not just how to like use it themselves, but to build it build systems out that work for like all of our customers and all of our employee base. Uh, so yeah, I I feel like I'm, you know, you know, capable or adaptive at times. Uh, and you know, the places I'm transformative is like not that many.

是啊。说实话，我觉得我在很多方面是「适应型」。呃，我感觉自己在很多这类事情上就是个超级通才，团队里有人在很多方面都比我强得多。我往「变革型」推进的时候，说实话，就是在偷师他们。我就是去看：你在做什么？你是怎么把这个东西琢磨明白的？然后把这些打法拿过来，套用到我自己的工作流、我自己在做的事情上。但说到我们作为公司、作为团队在做的事，呃，有些人比我聪明得多，琢磨明白的东西也比我多得多——不只是自己怎么用，还有怎么搭出能服务我们所有客户、全体员工的系统。所以，是的，我觉得我有时候是「胜任」，有时候是「适应型」。至于我能算「变革型」的地方，真没多少。

[1:02:59] Aakash

So interesting. I think like one thing I'm taking away is like the curve I had in mind like oh maybe 20 30% of people can land in transformative. It's not like that like that you should have a very high bar for that final rating.

太有意思了。我觉得我的一个收获是：我原本脑子里的那条分布曲线是，大概 20%、30% 的人能落到「变革型」。其实不是这样——最后这一档的门槛应该定得非常高。

[1:03:12] Wade

Yeah. I I think there's not that many people who are truly transformative in the world at this stuff. Like it literally is like to me transformative is like you are literally an end of one. You're the first of its kind doing this new thing. Uh and you're figuring it out. Um, and so I that bar is just exceptionally high uh in my opinion.

是的。我觉得在这件事上，世界上真正称得上「变革型」的人没那么多。对我来说，「变革型」真的就意味着你是 n of one（独一份，n=1）：你是第一个做这件新事的人，而且正在把它摸索出来。所以在我看来，这个门槛就是极其高的。

[1:03:30] Aakash

So if I'm a product leader, should I only be handing out on a PM team of 30 to 40 like maybe one transformative rating a half

所以如果我是产品负责人，在一个 30 到 40 人的 PM 团队里，是不是每半年大概只该给出一个「变革型」评级？

[1:03:36] Wade

maybe? Yeah. Uh, you know, it's it's it's a low it's a low percentage probably.

也许吧？是的。呃，你知道，这个比例……这个比例应该挺低的。

[1:03:43] Aakash

Okay, so that covers AI fluency. Hopefully you all got a really good deep dive into that. We have the CEO and co-founder. So we get to ask him the tough questions. And the tough question on everybody's mind. Air Table like you guys huge valuation 2021. Although I think in 2021 the $5 billion mark that came for you guys was actually through like some sort of secondary share sale to Sequoia or something. But it was a $5 billion mark that got labeled onto Zapier. People were looking at Air Table now it's like it's less than 10% of that amount for you at Zapier. You've been around since 2012. you had this amazing high in 2021. Where does Zapier's business sit now? Are you at 10% of your value? Have you grown since 2021? Talk to us about the numbers and what you can share.

好，AI 流利度这部分就到这里。希望大家对它有了一次很深入的了解。今天我们请到的是 CEO 兼联合创始人，所以可以问他一些尖锐的问题。而大家心里都在想的那个尖锐问题是：Airtable 跟你们一样，2021 年估值非常高。不过我记得 2021 年你们那个 50 亿美元的估值，其实是通过某种老股转让——卖给 Sequoia 之类的——得来的。但不管怎样，Zapier 身上是被打上了 50 亿美元的估值。大家再看 Airtable，现在好像连当年那个数的 10% 都不到了。至于你们 Zapier，2012 年就开始做了，2021 年到了那个惊人的高点。Zapier 的业务现在处在什么位置？你们现在是不是只剩当年价值的 10%？2021 年以来有没有增长？跟我们聊聊数字，能分享多少就分享多少。

[1:04:33] Wade

Yeah, I mean, we've grown since 2021. Uh, you know, we don't One thing that's unique about Zapier is because we don't we haven't really raised primary capital since 2012, like the the valuation thing doesn't matter as much to us. like valuations matter really only at one moment in time and it's when there's a transaction that happens and our goal has always been to just build incredible automation products for our customers and we kind of let the rest of that stuff take care of itself and so for the last you know couple years our driving force is like hey how do we wield these incredible AI tools to make it a lot easier for folks to build and run automations uh in this new way and so that's where we sort of put the lion share of more focus.

是的，我们 2021 年以来是有增长的。呃，我们没有……Zapier 有一点比较特别：因为我们从 2012 年以后基本就没再做过新股融资，所以估值这件事对我们没那么重要。估值其实只在一个时间点上真正有意义，就是发生交易的那一刻。而我们的目标一直都是给客户做出特别棒的自动化产品，其余的事就让它顺其自然。所以过去这几年，我们的核心驱动力就是：怎么用好这些厉害的 AI 工具，让大家用这种新方式搭建和运行自动化变得容易得多。这就是我们把绝大部分精力投入的地方。

[1:05:17] Aakash

So if you look at the no code space in general, people are calling it like the no code apocalypse and I considered you guys the leaders of that space. So help me reconcile you guys are growing but there was this no code apocalypse. Did you adapt in a particular way that product people can learn or is that apocalypse overblown?

那放眼整个无代码领域，大家都管现在叫「无代码末日」，而我一直把你们看作这个领域的领头羊。所以帮我把这两件事对上：你们在增长，可同时又有这场无代码末日。你们是用了某种特别的方式去适应、有值得产品人学习的地方，还是说这场末日被夸大了？

[1:05:37] Wade

Well I think one of the things we've come to believe is that the new no code is code. you know, one of the demos I was showing the daily brief, that's code. Like that's not no code. Like that's code under this the hood that's running that workflow. Uh and so I think a lot of what the, you know, these no code products have to do is sort of re uh think sort of like the underlying architecture of how these tools work. And so, you know, in a lot of these places, we're trying to expose our capabilities, you know, our trigger infrastructure, our action infrastructure, our runtime environments, our governance layer to coding agents and allowing them to go be the thing that goes and builds on behalf of our customers because, you know, there's so many people who are non-engineers who are using these tools to write code even though they're never looking at them. Like, I actually don't know what the code underneath my uh daily brief is actually doing. like I I poked it open and showed you just to show you, but I I've never actually looked at it and I don't really care because the daily brief works for me. Uh and so I think that's the key thing that a lot of these um you know kind of like the SAS incumbents have to sort of figure out is like hey how do you expose these capabilities to these tools so you can generate value on top of them. You know Benny off is talking about this with Salesforce. Salesforce is headless now like that's the value of Salesforce.

嗯，我们逐渐形成的一个信念是：新的无代码，就是代码。你看，我刚才演示的那个每日简报（daily brief），那就是代码。它不是无代码，是底层的代码在跑那个工作流。所以我觉得，这些无代码产品很多都得重新……呃，重新思考这些工具运作的底层架构。所以在很多地方，我们都在尝试把自己的能力——触发器基础设施、动作基础设施、运行时环境、治理层——开放给 coding agent，让它们替我们的客户去做搭建。因为现在有太多非工程师在用这些工具写代码，尽管他们从来不看那些代码。比如我其实并不知道我那个每日简报底下的代码到底在干什么。我刚才点开给你们看，只是为了展示一下，但我自己从来没真正看过，也不太在乎，因为这个每日简报对我来说好用就行。所以我觉得，很多 SaaS 老牌玩家要想明白的关键就在这里：你怎么把这些能力开放给这些工具，好让大家在上面创造价值。你看 Benioff 就在拿 Salesforce 讲这件事：Salesforce 现在已经无头化（headless）了，这才是 Salesforce 的价值所在。

[1:06:55] Wade

It's not, you know, the UI that you log into and all that other stuff. It's the APIs and the MCP servers and all that stuff under the hood that exposes those capabilities to a coding agent.

不是你登录进去的那个 UI 之类的东西，而是底层那些 API、MCP server 这一整套，把这些能力开放给 coding agent。

[1:07:05] Aakash

Yeah. I think it was a year and a half ago, we had Brian on your CTO and co-founder talking about the Zapier MCP. Just talk to us a little bit about that. Should all SAS be going headless? Should every SAS be creating an MCP and how did that contribute to the growth of your business?

对。我记得大概一年半以前，我们请过你们的 CTO 兼联合创始人 Bryan 来聊 Zapier MCP。跟我们稍微讲讲这个吧。所有 SaaS 都应该走向无头化吗？每家 SaaS 都该做一个 MCP 吗？这对你们的业务增长起了什么作用？

[1:07:20] Wade

Well, I I think so. You know, I think uh one of the things that we've long stood for is that no one really wants to show up to work and log into your SAS product and like sit there and like do data entry to that tool. You know, all the SAS that we've had over the last decade, it was still a human doing work and it's like I really want software that actually does the work for me. And so, you know, Zapier, that's always what we've been had this like workflow automation tool that does work for you. Um, but the one place where a human had to do work to use Zapier was building out this automation. You had to sit down and you'd have to like log in and you have to click through a bunch of things. Heck, you'd have to think about what do I want to automate. Thinking about what to automate is actually harder in some places than building it. And building was hard enough. And the magic now is if you can expose all the context and the intelligence and the tools for building, you know, whether you're a tool like Zapier or you're a CRM like Salesforce or, you know, instant these other tools, the user of your software now is the agent. The agent is one that comes in and does all that work. And so that frees up the human for the things that are distinctly human, the judgment parts of the job, the like thinking through like, well, is this actually correct or not? um you know those are the places where I think the human now gets like elevated in in its work and if you're thinking of your software as still oh I have a human user like the human is going to use my software to do this that and the other I think that's a little backwards at least for B2B I think in TOC it's a little different like consumers behave a little differently uh in you know the way they want to use software and the way they do it but at work it's like I want to delegate this stuff to somebody else to go use it for me uh and increasingly that's a coding

嗯，我觉得是的。我们长期以来坚持的一个观点是：没有人真的想上班以后登录你的 SaaS 产品，坐在那儿往里面录数据。过去十年我们用过的那些 SaaS，干活的仍然是人；而我真正想要的，是能替我把活干掉的软件。Zapier 一直以来就是这样——一个替你干活的工作流自动化工具。嗯，但用 Zapier 时唯一还得人来干活的地方，就是把自动化搭出来。你得坐下来、登录、点一大堆东西。说白了，你甚至得先想：我到底要自动化什么？在有些场景下，想清楚要自动化什么，其实比搭建还难。而搭建本身就已经够难了。现在的神奇之处在于：如果你能把所有上下文、智能和搭建工具都开放出来——不管你是 Zapier 这样的工具、Salesforce 这样的 CRM，还是其他这类工具——你的软件的用户现在就是 agent。是 agent 进来把这些活全干了。这样人就被解放出来，去做那些真正属于人的事：工作中需要判断的部分，比如琢磨「这到底对不对」。嗯，我觉得在这些地方，人的工作被抬升了一个层次。如果你还把自己的软件想成「我有一个人类用户，这个人会用我的软件做这做那」，我觉得这有点想反了，至少在 B2B 是这样。B2C 我觉得会有点不同，消费者想怎么用软件、实际怎么用，行为上会有些差别。但在工作中，我想要的是把这些事委派给别人，让对方替我去用。呃，而越来越多时候，这个「别人」就是一个 coding

[1:09:03] Wade

agent so if your software isn't headless or cannot be run headless, I think you're going to find yourself, you know, sort of swimming against a a cur um, you know, the current a little bit.

agent。所以如果你的软件不是无头的，或者没法以无头方式运行，我觉得你会发现自己有点像是在逆着……呃，逆着水流游泳。

[1:09:13] Aakash

H, so this is making me think my PRD might not have been aligned with your strategy [laughter] because it was focused on a human user, not an agent user. Very interesting. I'm going to have to think about that a little bit more. I want to talk about kind of this new breed of competitors. So, we had the Lindy CEO on. He talked about how co-work actually did have a pretty big negative impact on their business and they had to reinvent himself. We also um I'm recording with Yan the CEO of N8N as well. They've had a pretty amazing trajectory 7 million to 40 million AR 40 million to 100 million ARR just raised at a pretty big valuation. When you think about these new competitors, how have you thought about counterpositioning against the Lindy's, the co-works, the NAMs, and how do you create that unique space for Zapier, which has been around for so long?

嗯，这让我觉得，我那份 PRD 可能跟你们的战略没对齐（笑），因为它是围绕人类用户、而不是 agent 用户来写的。很有意思，这点我得再好好想想。我想聊聊这批新型竞争对手。我们之前请过 Lindy 的 CEO 上节目，他讲到 Cowork 确实给他们的业务带来了相当大的负面冲击，他们不得不重新打造自己。另外，呃，我也在跟 n8n 的 CEO Jan 录一期节目。他们的增长轨迹相当惊人：ARR 从 700 万到 4000 万，又从 4000 万到 1 亿，刚刚还以相当高的估值融了一轮。面对这些新竞争对手，你是怎么考虑针对 Lindy、Cowork、n8n 这类玩家做反定位（counter-positioning）的？又怎么为已经存在了这么久的 Zapier 开辟出独特的位置？

[1:10:08] Wade

Yeah. Well, I think what is new for us is that, you know, to your point, we've been doing this for 15 years, and for most of those 15 years, um, the space was not very crowded. there wasn't a lot of man lot of automation players. Now on one hand that's nice like you know hey we're sort of the only game in town it landed it allowed us to build this incredible integration library that still nobody has matched. You know we have over 9,000 different tools that we hook into. Uh you know folks kept saying oh it's easier and easier to write integration code and yet no one has has matched what we've done here. Uh and so that's the nice part. The flip side though is that now um there's sort of like a a good new good news bad news story to this. The good news is the demand for automation skyhigh like it's never been higher. So the markets are so so so big these days compared to you know maybe our first decade of existence. Now with with big markets also comes new approaches, new ways of doing this stuff. And so there's a lot more competitors. And so I think where this, you know, really forces you to put your product head on is you now have to make strategic choices. You have to say, what are we going to really focus on and what are we going to be good at and really make um, you know, these smart choices for. I mean, heck, like let's use OpenAI and Anthropic as like a great example where, you know, I think Open AI like kind of was the early dogs in in the race. Uh, and they were chasing all sorts of different things.

是的。我觉得对我们来说新鲜的是——就像你说的，我们做这件事已经 15 年了，而这 15 年里的大部分时间，这个赛道都不怎么拥挤，没有多少做自动化的玩家。一方面这挺好的，我们差不多是独此一家，这让我们搭起了一个至今没人比得上的集成库。我们接入了 9,000 多个不同的工具。大家一直说写集成代码越来越容易了，可还是没人追上我们在这上面做到的程度。这是好的一面。但另一面是，现在这件事有点像是一个好消息、坏消息并存的故事。好消息是，对自动化的需求高到天上去了，从来没这么高过。跟我们创立后的头十年相比，现在的市场大得不得了。而市场一大，也会冒出新的思路、新的做法，所以竞争对手也多了很多。我觉得这真正逼着你拿出产品思维的地方在于：你现在必须做战略选择。你得说清楚：我们到底要专注什么、要在什么上做到出色，并为此做出明智的取舍。说真的，拿 OpenAI 和 Anthropic 来举例就特别好。我觉得 OpenAI 算是这场竞赛里最早领跑的，呃，他们当时各种方向都在追。

[1:11:36] Wade

Anthropic says, you know what, we're going to focus on code specifically. We're going to be great at code. And you look at what that did for them. Like that allowed them to sort of stand out and now it's forced open AI to be smarter about their strategy. And so you're seeing them get a lot more crisp about who are we for and where do we want to win, etc. And so I think that's you know when you have these markets that are so big and you're have many players that are going to be successful you have to figure out like okay what is our unique value prop and for what set of customers can we go build around and I think that's what you know ourselves Lindy Naden any of us like have to go answer those questions and I think for Zapier increasingly the answer is you know we're going to help you run these automations we're going to help you run them deterministically so you're not burning tokens you don't have to keep your laptop open um and but you still can have a power of an agent that's building them and helping you manage them. So, the setup is really easy and when they break and when they fail in unpredictable ways, the agent can help you recover from that stuff. And Sam is going to be really really good at that.

Anthropic 则说：你知道吗，我们就专门做代码，我们要在代码上做到最好。你看这给他们带来了什么——让他们脱颖而出，现在反过来逼着 OpenAI 在战略上更聪明。所以你能看到，OpenAI 在「我们为谁服务、要在哪里赢」这些问题上清晰了很多。所以我觉得，当市场这么大、会有很多玩家都能成功的时候，你就得想清楚：好，我们独特的价值主张是什么？我们能围绕哪一类客户去做？我觉得这正是我们自己、Lindy、n8n，我们当中任何一家都得回答的问题。对 Zapier 来说，答案越来越明确：我们会帮你运行这些自动化，而且以确定性的方式运行，这样你就不用一直烧 token，也不用一直开着笔记本电脑。但你依然拥有 agent 的能力：由它来搭建这些自动化、帮你管理它们。所以搭建起来非常容易；而当它们出故障、以意想不到的方式失败时，agent 也能帮你从中恢复。而我们会在这方面做得非常非常好。

[1:12:37]

Uh, and that's where we've been growing is is like helping folks with those types of capabilities. We could talk for another hour and a half, but I need to let you get on to the day job of managing Zapier to growth and success. Wade, thank you so much for being on and teaching us about AI fluency today.

Wade： 呃，我们一直在增长的地方，就是帮大家获得这类能力。

Aakash： 我们再聊一个半小时都聊不完，但我得放你回去干正事了——带着 Zapier 继续增长、走向成功。Wade，非常感谢你来做客，今天教了我们这么多关于 AI 流利度的东西。

[1:12:54] Wade

Thanks for having me.

谢谢邀请。

[1:12:55] Aakash

All right, everyone. If you want more episodes like this with killer CEOs, comment below and we'll see you in the next episode. 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 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 [music] 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 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, new, air tableable, 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.

好了，各位。如果你想看更多这种和顶级 CEO 对谈的节目，就在下面留言，我们下期见。希望你从今天这期节目里学到的和我一样多。如果你能做一件完全免费、又能帮到这个节目的事，那就是：确认一下你在 Apple 和 Spotify 播客上关注了我们，确认你在这些平台上留了评分和评论，确认你订阅了我们的 YouTube 频道，再给这个视频点个赞、留个言，然后分享给你的朋友。我们一直在努力把播客做得越来越好。做了 2 年，我们觉得已经做出了点相当不错的东西。所以告诉我们怎样能做得更好、还应该采访谁，我们会尽全力做出最好的节目。最后，别忘了我的 bundle 套餐优惠：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。价值 27,000 美元的东西，只要 150 美元。感兴趣的话去 bundle.aakashg.com 看看。迫不及待想尽快和你分享下一期节目。