I let Codex control my browser so I don't have to

频道: How I AI
视频: https://www.youtube.com/watch?v=lk63Sl-LRKE
原文语言: en
统计: 共 78 轮

[0:00]

One of my favorite guests ever, Jessie Janes, said that AI was great for moms because moms' hands are always busy. We're always doing something with our hands, and we do not have time to talk. Now, in that episode, she was talking about the benefits of chatting with AI,

我最喜欢的嘉宾之一 Jessie Janes 说过，AI 对当妈的人特别友好，因为妈妈们的手永远是闲不下来的。手上总有活儿要干，但嘴是闲着的。在那期节目里，她讲的是跟 AI 用聊天的方式打交道有多香，

[0:15]

but in this episode, which I'm going to lovingly call Look, Ma, no hands, I'm going to talk about browser and computer use, the way that you can let AI navigate your computer and navigate websites to get work done on your behalf. Browser and computer use have gotten exceptionally good recently,

而这一期，我想亲切地叫它《看，妈，我不用手》。我要聊的是 browser use 和 computer use——也就是让 AI 直接操作你的电脑、浏览网页，替你把活儿干了。最近这一年，browser use 和 computer use 的能力提升得非常夸张，

[0:32]

especially in my experience with Codex and the GPT 5.6 models, and I'm going to show you a builder use case, a productivity use case, and a personal fun use case that I found super useful in using computer and browser use. Let's get to [music] it. This episode is brought to you by Runway, a new kind of creative platform that has everything you need to generate any image, video,

尤其是我自己用 Codex 和 GPT 5.6 系列模型的体验。我会给你演示三个场景：一个面向开发者的，一个提效类的，还有一个纯属个人娱乐的，这几个我用下来都特别顺手。咱们直接开始。本期节目由 Runway 赞助，这是一个全新形态的创作平台，你想生成的任何图片、视频

[0:57]

or piece of content you want, all in one place.

或者内容，在一个地方就能全部搞定。

[1:00]

[music]

[音乐]

[1:01]

With Runway, it's now possible to go from initial idea to a finished deliverable in a matter of minutes.

有了 Runway，从最初的一个想法到最终能交付的成品，几分钟就能走完。

[1:07]

[music]

[音乐]

[1:07]

From turning low-fidelity product shots into campaign-ready imagery all the way through putting together big brand films, Runway can help your team scale your creative ambitions while keeping your budgets and timelines from doing the same.

从把粗糙的产品样片变成能直接投放的营销素材，到制作大型品牌影片，Runway 能帮你的团队把创意野心放大，同时预算和排期还不用跟着一起膨胀。

[1:21]

[music]

[音乐]

[1:21]

Runway brings together the world's most advanced AI models, which is why enterprises like Microsoft, Robinhood, Amazon, and Adobe, along with studios like Lionsgate and Legendary, all use Runway to ship real work every day. Try it yourself at runwayml.com/howiai.

Runway 把全球最先进的 AI 模型整合到了一起，所以像 Microsoft、Robinhood、Amazon、Adobe 这样的大企业，还有 Lionsgate、Legendary 这样的影视公司，每天都在用 Runway 做真正的商业产出。你也可以去 runwayml.com/howiai 亲自试一试，

[1:43]

Promo code howiai. Okay, before we dive into how I use computer use, just a refresher for anybody who's new. In case you missed it, these LLMs, your AI chatbot, can control your computer and can control your browser. Now, there have been a class of AI native browsers that have come out in the last year. Thinking about Perplexity's Comet or ChatGPT's Atlas, but I like to use browser and computer use directly through Claude or

优惠码是 howiai。好，在深入讲我怎么用 computer use 之前，先给新朋友简单科普一下。可能你还不知道：这些 LLM，也就是你平时用的 AI 聊天机器人，是可以直接操控你的电脑和浏览器的。过去一年冒出来一批 AI 原生浏览器，比如 Perplexity 的 Comet、ChatGPT 的 Atlas，但我更习惯直接在 Claude 里用 browser use 和 computer use，

[2:15]

in this instance, Codex. Why have I used Codex, specifically the Codex desktop app or I guess we're calling it the ChatGPT desktop app these days? Well, it is just the best at controlling my computer. And if you didn't know, it can control the mouse, it can control the keyboard, it can do all sorts of things.

或者像这次一样，在 Codex 里用。为什么我用 Codex，具体说是 Codex 桌面 App——或者现在大家好像都叫它 ChatGPT 桌面 App 了？因为它操控我电脑的能力就是最强的。你可能不知道，它能控制鼠标、能控制键盘，基本什么都能干。

[2:35]

So, imagine you have a virtual coworker and that coworker can put its hands on your keyboard and get stuff done on your computer. That is what browser and computer use can do. Now, this sounds really scary, but in fact, it has been one of the most life-changing features of AI. Certainly professionally changing, but it has also reduced so much personal toil and personal digital toil that I had to tell you about a couple of the use cases that I use

你可以想象成：你有一个虚拟同事，这个同事能把手放到你的键盘上，直接在你电脑上帮你干活。这就是 browser use 和 computer use 能做的事。听起来是有点吓人，但说实话，这是 AI 里最能改变生活的功能之一。工作上肯定是革命性的，同时它还帮我省掉了大量个人生活里那些琐碎的数字杂事，所以我必须来跟你分享几个我常用的场景，

[3:04]

because I literally use this every minute of every day. Now, invoking browser or computer use is pretty simple. First, you have to have those desktop apps installed on your computer. So, whether those are the ChatGPT desktop app or the Claude desktop app, you need those installed.

因为我真的是每天每分钟都在用。启动 browser use 或者 computer use 其实很简单。首先，你得在电脑上装好对应的桌面 App——不管是 ChatGPT 桌面 App 还是 Claude 桌面 App，都得先装上。

[3:22]

If you want to use the browser, you need to have the browser extensions installed inside Chrome. I don't know if they work for these other browsers. I just use Chrome and I know they have browser extensions. So, you can just search for, for example,

如果你想用浏览器这部分能力，还得在 Chrome 里装好浏览器扩展。别的浏览器行不行我不清楚，我自己就用 Chrome，我知道 Chrome 是有扩展的。你直接搜就行，比如搜

[3:36]

Codex browser extension or Claude browser extension. Make sure they're they're the official ones and then install those and then these two things together, the desktop app plus the Chrome extension, can really allow your computer plus AI to do a lot of work on your behalf. Now, it sounds really simple. Let your AI use your computer.

Codex browser extension 或者 Claude browser extension。注意认准官方的那个，装好之后，这两件东西配合起来——桌面 App 加 Chrome 扩展——就能让你的电脑和 AI 一起替你干不少活了。听起来特别简单，就是让 AI 用你的电脑嘛。

[3:56]

But, the trick, and this is the how I AI sort of mantra, is you have to find a good use case for browser use. And having used browser and computer use for a while, I think I've come up with a couple unique and interesting use cases that maybe will inspire you and give you some ideas about how you can use it in your work and life. We're going to start with a use case out there for all the coders and all the vibe coders and all

但关键在于——这也是 How I AI 一直在讲的那句话——你得给 browser use 找到一个真正合适的场景。browser use 和 computer use 我用了一段时间，摸索出了几个还挺独特、挺有意思的用法，说不定能给你一些启发，让你想到怎么把它用到自己的工作和生活里。第一个场景是给所有写代码的人准备的，不管你是 vibe coding 还是

[4:22]

the professional software engineers, which is using browser use alongside a coding agent to validate changes. Now, this is very simple, but one that I use quite a bit. So, what I'm showing here is the Chat PRD onboarding flow. It's just a step-by-step where you fill out your profile. Let's say I was working on some enhancements on this feature and really wanted someone, and by someone I mean that all those features work really well. Now, I

专业的软件工程师，就是把 browser use 和 coding agent 配合起来，用它来验证改动。这个特别简单，但我用得非常多。我现在屏幕上展示的是 Chat PRD 的 onboarding 流程，就是一步一步填资料建档案。假设我刚对这个功能做了一些优化，我特别想找个人——这里说的“人”其实就是 AI——来确认这些功能都跑得没问题。当然我

[4:54]

could write what AI loves, a smoke test. I can definitely test things on the back end. I can definitely write tests for the front end. But, what I really want to know is is it usable? Can we use it? And so, what I might do here is say, "Use browser." And there are three ways you can invoke kind of computer use in Codex. You can do @browser, that will pull up use of this side browser window. There is @chrome,

可以写一个 AI 最爱写的 smoke test。后端我肯定能测，前端我也能写测试。但我真正想知道的是：它好用吗？真的能用吗？所以我在这儿可能就直接说一句 use browser。在 Codex 里调用 computer use 有三种方式：可以用 @browser，它会拉起旁边这个内置浏览器窗口；还有 @chrome，

[5:22]

which will use the Chrome extension to manage Chrome. And then there's @computer, which lets it manage the whole computer, just just so you know. But, I'm going to say use browser cuz I'm already logged into my localhost over here on the browser. I'm going to say, "To test the open onboarding flow for usability and mobile responsiveness.

它会通过 Chrome 扩展来操作你的 Chrome；再有就是 @computer，这个能让它接管整台电脑，你先知道有这么回事。不过我这次说 use browser，因为我在旁边这个浏览器里已经登录了本地的 localhost。我会说：请测试一下这个 onboarding 流程的易用性和移动端响应式表现，

[5:45]

Please screenshot as you go and make a Google spreadsheet of any issues you identify. Okay, so I'm going to click submit and let that run. And what that's going to do is spin up the browser skill, which is embedded in Codex, and you'll actually see this little cursor start to go over to the browser side and walk through the onboarding flow. Why I think this is useful is this is how I would test if this is working. You know, I would go

过程中随时截图，并把发现的问题整理到一个 Google 表格里。好，我点提交，让它跑起来。接下来它会调起 Codex 内置的 browser skill，你马上就能看到那个小光标跑到浏览器那边，一步一步走完整个 onboarding 流程。我觉得这个用法有价值，是因为这本来就是我自己测的方式：我会一路

[6:20]

through, I would click, I would test all the drop-downs, I would change the viewport sizes. And what's really nice about browser use in this instance is all of that can happen automatically. I can just sit back and watch it go. And I will tell you, it's really fascinating to watch these things go. If you have looked at that little glowy cursor and watched it bop around your website,

点过去，把所有下拉框都点一遍，再改改视口尺寸。browser use 在这种场景下最爽的一点，就是这些全都能自动完成，我往后一靠看着它跑就行。而且我跟你说，看它自己跑真的挺上头的。如果你看过那个发光的小光标在你网站上到处蹦跶、

[6:43]

click things, test things, you know the magic feeling of this particular AI feature. Now, what I love about using browser use specifically for QA purposes is it is exhaustive or it is more exhaustive and organized than I as a human would be. I don't know about you all, but when I am QAing things, I like to do the happy path QA where it's like,

点这点那、测这测那，你就懂那种神奇的感觉了。我特别喜欢把 browser use 用在 QA 上，是因为它比我这个人类要更穷举、更有条理。不知道你们怎么样，反正我自己做 QA 的时候，特别容易只走 happy path，就是那种：好，

[7:06]

yes, I fill out all the form fields, everything's required, I click next, I don't get tired, I don't go back. But you can see here, my little glowy Codex browser icon, it's going to be exhaustive. It's going to test signing up with a team, it's going to test signing up as an individual, it's going to fill out form fields, it's going to try for error states. It's going to do all these things that you wouldn't probably have the stamina or

表单该填的我都填了，必填项都填了，点下一步，我不会累，也不会往回退。但你看这儿，我这个发亮的 Codex 浏览器图标，它是真的会穷举：它会测以团队身份注册，也会测以个人身份注册，会去填各种表单字段，还会故意去触发报错状态。这些事情，换成你自己来，多半是没那个耐心、也没那个

[7:31]

organization to do if you were like me a little late in the afternoon after coding and asked to sort of manually test your web app. And so what I really like about this is its ability to be exhaustive. Now, if I were being a little bit more organized,

条理去一条条做的——尤其像我这样，下午写完一堆代码之后被要求手动测一遍 web app，谁受得了。所以我最欣赏的就是它这种穷举的能力。当然，如果我再规整一点，

[7:46]

I would probably have it come up with a suite of test cases based on the code, then go through those test cases kind of specifically over time, so I knew it was being exhaustive. That being said, you know, side tip for some of these newer models, I actually found that under prompting these models get you a much better effect than over prompting them.

我大概会先让它基于代码整理出一整套测试用例，然后按着这些用例一条条跑，这样我心里更有底，知道它确实覆盖全了。话说回来，这里插一句关于新一代模型的小心得：我发现对这些模型来说，prompt 写得少反而比写得多效果更好。

[8:06]

And so because they're so intelligent, you can literally just tell them QA the onboarding flow, and it will do a much better job than saying QA these 25 things in the onboarding flow. So, a little tip, under prompt your frontier models, let them use their that intelligence to come up with their own plans, and do what it wants to do. Now,

因为它们本身足够聪明，你就直接说一句“QA 一下 onboarding 流程”，效果会比你列“把 onboarding 里这 25 项都测一遍”好得多。所以给你个小 tip：对前沿模型要少写 prompt，让它自己动脑子定计划，想怎么干就怎么干。好，

[8:28]

we're going to watch this a little bit. And what I like about what I've told it to do is it also is being asked to look at responsive design. And even seeing here, I see an issue with the um right hand of the browser. I'm curious if that's actually an artifact of how this responsive browser inside CodeX works,

咱们再看一会儿。我刚才那条指令里我比较满意的一点，是我还让它顺便看了响应式设计。现在我已经能看到浏览器右侧那块好像有点问题。我挺好奇这到底是 Codex 里这个响应式浏览器本身渲染出来的假象，

[8:50]

or if it's an issue with my responsive design. We will see. But you can also see that it's changing the dimensions of the viewport, so the size of the monitor to match laptop size. It's going to show matching a phone size, all those things.

还是我响应式设计真的有 bug。待会儿就知道了。另外你也能看到，它在改视口的尺寸，也就是把显示器尺寸调成笔记本那么大，接下来还会调成手机大小，这些它都会试一遍。

[9:05]

Again, things that I could technically do manually that I really don't want to do. I'll let this run, and then I'm going to show you one other use case as a builder that I think is a little bit more unique that you might find inspirational. Okay, so something interesting here. It did find an issue.

再强调一次，这些事我理论上都能手动做，只是我是真的不想做。我让它先跑着，然后再给你看另一个面向开发者的场景，那个我觉得更独特一些，说不定能给你点灵感。哎，这儿有个有意思的：它还真发现了一个问题。

[9:21]

Now, I'm a little embarrassed because this onboarding flow has been here for a while. Um it surfaced one blocking issue, which is if you don't select anything on this page, you can still click continue, but it won't move forward because one of these fields is required, but there's no validation code in here. Now, again, the reason why this bug exists is because I always tested this flow as a human and I always click

说起来有点不好意思，因为这个 onboarding 流程已经上线挺久了。它发现了一个 blocking 级别的问题：如果你在这一页什么都不选，照样能点“继续”，但页面不会往下走——因为其中有个字段是必填的，可代码里压根没写校验。这个 bug 之所以一直没被发现，就是因为我每次都是以真人的身份来测，而我每次都会去点

[9:49]

the required buttons. So, again, these benefits of having AI test your flow is it will test failure states as well as success states. It actually bubble up edge case bugs that you might not have thought of. And so, it's going to continue through and actually work and go to the next screen. I'm curious if it will spin up mobile testing next.

那些该点的必选项。所以说，让 AI 来测流程的好处就在这儿：它会把失败路径和成功路径都测一遍，能真的挖出你压根没想到的边缘 case。接下来它会继续往下走，进到下一屏。我挺好奇它接下来会不会自己开始跑移动端测试。

[10:12]

Okay, you can see now it's also switched to mobile testing. Again, this is something I would probably give a cursory glance to. Like I would make the browser smaller and then kind of click through stuff and make sure it works. But what I really love about browser use is again, it's going to be just a lot more exhaustive. It's going to look at things like accessibility. It's going to look at overflow. It's going to look at

好，你看，它现在已经切到移动端测试了。这种事我自己一般也就是草草扫一眼——把浏览器窗口拉小，随便点一点，确认能用就行了。但 browser use 让我特别喜欢的一点是，它做得彻底得多：它会看 accessibility，会看内容溢出，还会看

[10:33]

touch targets. All that kind of stuff. And just has a much more organized sense of what makes for good mobile testing. Okay, so it wrapped up. It found 11 items including one high severity navigation blocker, the one we talked about. And now, this is one of my favorite things that I do with AI is I just have it make me a Google Sheet or a Google Document. And it's going to make me one which is this prioritized. And then what I can do is use that Google

触控区域大小，诸如此类。它对“什么才算像样的移动端测试”有一套非常有条理的认知。好，跑完了。它一共找出 11 个问题，其中包括我们刚才说的那个高严重级别的导航阻塞。接下来是我最爱让 AI 干的事之一：直接让它给我生成一份 Google Sheet 或者 Google Doc。它会做一份排好优先级的表，然后我就可以拿这份 Google

[10:59]

Sheet as a tracker to kick off a bunch of fixes and track the status of them with Sub Agents. So, again, this is a very comprehensive QA that was done. I get this beautiful list of issues that came out. I can get those fixed ASAP. And it ends up in a sheet. Okay, so it wrapped up and it did everything I asked it to. It tested it in desktop and mobile. It gave me these beautiful screenshots telling me where to work and

Sheet 当 tracker，用 sub agents 批量启动修复、跟踪进度。所以这就是一次相当完整的 QA：我拿到一份清清楚楚的问题清单，可以马上安排修，最后全都沉淀进一张表里。好，它跑完了，我交代的事一件没落——桌面端和移动端都测了，给了我这些漂亮的截图，标明该从哪儿下手，

[11:25]

it even made a spreadsheet. Let's see if the spreadsheet is actually if I can pull it up. So, I love this flow, this workflow, which is to have Codex make you a spreadsheet of issues that then it's going to go fix. And so, it has this nice summary. It found 11 issues,

甚至还做了一张表格。我们看看这表格能不能打开。我真的很喜欢这个工作流：让 Codex 给你生成一份问题清单表格，然后它再照着这张表一个个去修。它还写了段挺不错的总结：一共 11 个问题，

[11:41]

one that was high severity. It listed out all the issues. Again, this is really useful for agents because it has all the information it needs to reproduce the issue, the viewport, what the issue was, what the steps are to fix it, and then how to remediate it. And then, look at this. It put all the screenshots here in the spreadsheet, so I could reference them and even compare them to fixes. So, this is just like a very practical use case of browser use

其中一个是高严重级别。它把所有问题都列了出来。这对 agent 来说特别有用，因为复现问题需要的信息全在里面：viewport、问题是什么、复现步骤、以及该怎么修。再看这儿——它把所有截图都塞进表格里了，我可以随时对照，甚至拿来跟修完之后的效果做对比。所以这是 browser use 一个非常实在的用法，

[12:09]

that I think is really nice. But, what I want to show you is the same thing, but a slightly more fun use case. So, I'm going to go in here. I'm going to go to the core of Chat PRD, and I'm going to show you a use case that my husband came up with. I have to give him credit, my husband EJ Lawless.

我觉得挺不错的。不过我想给你看的是同样的玩法，但场景更好玩一点。我这就进去，打开 ChatPRD 的核心功能，给你演示一个我老公想出来的用法。这个功劳得记在他头上——我老公 EJ Lawless。

[12:25]

And he uses Browser Use to impersonate specific personas of his applications and then give feedback as the persona. So, let me just give an example of what this might look like. And I'm going to pull up Model Log and just speak my prompt. I want you to go through the core of Chat PRD's chat and document creation mode as two different personas.

他会用 Browser Use 去扮演他产品里某个具体的 persona，然后以那个 persona 的口吻给反馈。我举个例子你就明白大概是什么样。我打开 Monologue，直接把 prompt 说出来：“我要你以两个不同的 persona，走一遍 ChatPRD 核心的对话和文档创建模式。”

[12:49]

Actually, three different personas. I want you to do one as a product manager trying to quickly create PRD for a feature that they just got out of a meeting um about. The second thing I want you to do is impersonate a engineer who is taking that already created PRD and trying to make a technical spec out of it and eventually a prototype. And the third thing that I want you to do is as a team leader, try to figure out how

不对，是三个 persona。第一个，你扮演一个产品经理，刚开完会，要赶紧给某个功能写一份 PRD。第二个，你扮演一个工程师，拿着这份已经写好的 PRD，把它变成技术方案，最后再做出一个原型。第三个，你扮演团队负责人，想搞清楚

[13:18]

your team is using Chat PRD. I'm going to give you no more instructions than just that. I want you to go through those three persona use cases and flows, and then I want you to create a doc providing research-style critique about the overall experience, including friction points, places of delight, and improvements you think we can make to serve those three personas better.

自己团队到底是怎么用 ChatPRD 的。我就给你这么多指示，别的一概不说。你把这三个 persona 的用例和流程各走一遍，然后写一份文档，用做用户研究的口吻点评整体体验，包括哪里有摩擦、哪里让人惊喜，以及你觉得我们可以怎么改进，才能更好地服务这三类用户。

[13:41]

Okay, so I got that prompt in. I'm going to click enter. So, just to like reiterate what this is, if you have personas and you want to test against those personas. Now, like I'm not a big synthetic user girl, but I do kind of like this framework. We used to always do this with our prompting where we're like, "You are a genius product manager." or "You are the best software engineer in the world." Now, I'm using that, but I'm using browser use and

好，prompt 输进去了，我点回车。再解释一下这是在干嘛：如果你手上有 persona，又想针对这些 persona 做测试，就可以这么玩。我本人其实不太吃 synthetic user 那一套，但这个框架我还挺喜欢的。我们以前写 prompt 不也老这么干嘛——“你是一个天才产品经理”“你是全世界最厉害的软件工程师”。现在我还是这一套，只不过配上了 browser use，

[14:06]

saying, "You are a persona type. Use my application as that persona, and then give me research as you go through the browser experience about what you like and don't like." So, I'm going to let that run a little bit, and then I will show you what we come up with when it's done. As you can see here, it's going down the PM path. It just put a fake prompt in that says, "Hey, I just got out of a customer design meeting about a

变成“你是某某类型的 persona，用这个身份去用我的产品，然后一边走完浏览器里的完整体验，一边像做研究一样告诉我你喜欢什么、不喜欢什么”。我让它先跑一会儿，跑完了再给你看结果。你看这儿，它在走 PM 这条路径。它编了个假 prompt，写的是：“我刚开完一个客户设计会，是关于

[14:30]

decision digest feature in ChatPRD. Here are the problems. Here's like kind of what I want. Here are some open questions." And now, it's watching ChatPRD create the document. I'm really curious if it will actually analyze the quality of the document itself or not.

ChatPRD 里的 decision digest 功能。这些是我们遇到的问题，这是我大概想要的东西，这里还有几个待定问题。” 现在它正看着 ChatPRD 生成文档。我特别好奇它到底会不会去分析文档本身的质量。

[14:46]

You know, it's giving me good positive feedback as we go, but we're going to watch this. We're going to like speed through some of the the browser use. I will say one downside of browser use is it's kind of slow. So, honestly, when this is running, I typically open up something else on the side and like work on something else as the browser is is running its flow.

目前它一路给的反馈都挺正面的，我们接着往下看。这段 browser use 我们快进一下。说实话 browser use 有个缺点就是慢。所以它跑的时候，我一般会在旁边开着别的东西，一边等它跑一边干点别的活。

[15:06]

But, it's going to go through and give me an analysis as that first persona, and then we're going to see it work through the rest of them. This episode is brought to you by HyperAgent, the platform [music] for deploying always-on agents that actually run your business.

不过它会一路走完，以第一个 persona 的身份给我一份分析，然后我们再看它把剩下的走完。本期节目由 HyperAgent 赞助——一个部署 always-on agent、真正帮你把业务跑起来的平台。

[15:19]

With HyperAgent, you build agents in the cloud and deploy them where your work already happens, [music] like Slack, Telegram, or email. An agent will scan your inbox and draft replies to vendor follow-ups. Another monitors competitors and spins up rich ad kits and landing pages. [music] A third notices a deal going cold in Salesforce and writes the save email with full account context. These aren't chatbots waiting for a perfect prompt.

在 HyperAgent 上，你在云端搭 agent，然后把它部署到你日常工作真正发生的地方，比如 Slack、Telegram 或者邮箱。一个 agent 可以扫你的收件箱，替你起草给供应商的跟进回复；另一个盯着竞品动向，自动生成完整的广告素材包和落地页；还有一个能发现 Salesforce 里某个正在变冷的商机，结合完整的客户背景写一封挽回邮件。这些可不是等你憋出一条完美 prompt 才肯动的聊天机器人。

[15:45]

They're proactive, [music] learning your preferences, retaining your playbooks, and getting better with every run. One user built four agents to run an outbound sales pipeline, prospecting, outreach, follow-ups, CRM updates, all in a single afternoon. No local setup,

它们是主动的：会学习你的偏好、记住你的打法，每跑一次都比上一次更好。有个用户一个下午就搭了四个 agent，把整条对外销售流程跑通了——找线索、触达、跟进、更新 CRM，全都包了。不用在本地配环境，

[16:01]

no VPS bills, no fragile permissions on your laptop, just powerful agents with full control over skills, tools, and guardrails. Hyperagent was built by the team behind Airtable, and How I AI listeners get a $1,000 in free inference to start building. Claim yours at hyperagent.com/howiai.

不用买 VPS，也不用在自己电脑上折腾那些一碰就崩的权限，就是一批强大的 agent，技能、工具、护栏全在你掌控之中。HyperAgent 由 Airtable 的原班人马打造，How I AI 的听众可以领 1000 美元的免费推理额度来上手。去 hyperagent.com/howiai 领取。

[16:24]

Okay, so I love this internal kind of like chain of thought that you get with browser use. So, what it tried to do was do the handoff from PM to engineering. And what you see here is that the engineer handoff exposed a structural break, which is a real issue in Chaptivity, which is you can create one document, but you can't like at mention or reference it in another thread. It's something that we're working on or thinking through how we want to

好，browser use 这种把内部思考过程摊开给你看的感觉，我真的很喜欢。它刚才想做的是从 PM 交接到工程师这一步。你看这儿写着，工程师这次接手暴露出一个结构性断点——这在 ChatPRD 里确实是个真实存在的问题：你能创建一个文档，但没法在另一个会话里 @ 提及或者引用它。这个我们正在做，也在琢磨到底要怎么

[16:48]

reference. It works in a specific way, but obviously the agent here, the Codex agent, thought, reasoned, um, that this feature would work a different way. And so, it's really interesting when you combine browser use with sort of this AI chatbot,

做引用。它现在是按某种特定方式运作的，但很明显这个 agent，也就是 Codex agent，推理下来以为这个功能应该是另一种运作方式。所以把 browser use 和这种 AI 对话结合起来，真的挺有意思的，

[17:07]

how you start to inhabit the mindset of a persona or a type of user, and actually identify really obvious, but kind of like hard to get your head wrapped around friction points where someone with fresh eyes, again, these are agent eyes. I'm sorry. I don't know a better metaphor. Um, might have a challenge. And so, we went back to the document said I'm the engineer picking up this PRD, turn it into an implementation-ready technical spec,

你会开始真正代入某个 persona 或者某类用户的心态，然后发现那些其实特别明显、但你自己怎么都绕不过弯来的摩擦点——就是那种“新鲜眼睛”才看得见的东西。当然这次是 agent 的眼睛，不好意思，我实在想不出更好的比喻了。总之我们回到那份文档，跟它说：“我是接手这份 PRD 的工程师，把它变成一份可以直接开工的技术方案。”

[17:32]

we'll see what the experience is and if the agent is happy. Okay, it called out something really interesting here, which I don't know if it's an artifact of this running locally and something weird being happening locally, but it did call out that the loading screen was super slow and opaque and the interface didn't really make it clear what was happening when it was writing. I think we hit an error, but this is really good feedback

看看体验如何，看看这个 agent 满不满意。好，它这儿指出一个很有意思的点——我不确定是不是因为在本地跑才出的状况，本地有时候会有点奇怪——它指出加载页面特别慢，而且一片空白，界面在生成内容的时候完全没让人搞清楚到底在发生什么。我估计我们这儿是报了个错，但这个反馈对我来说特别有价值，

[17:55]

for me because again, I want to make sure even if we have a failure state, it's going to make sense, {quote} {unquote}, to the user. So, just to take a step back, I really love this use case of coming up with a persona that kind of inhabits the mind and, in this world, the body of your user and then have them going through your website using browser use and giving you a report of what works and what doesn't.

因为我希望哪怕是失败状态，对用户来说也得是“说得通”的。所以退一步看，我真的很喜欢这个用法：设定一个 persona，让它钻进你用户的脑子里——在这个场景下甚至可以说是钻进用户的身体里——然后用 browser use 走一遍你的网站，最后给你一份报告，告诉你哪儿好用、哪儿不好用。

[18:21]

Okay, so those are my building use cases. I'm going to zip through a couple other use cases that I think are kind of fun. So, my favorite one, I've talked about this a little bit briefly before, which is browser use for LinkedIn. LinkedIn, close your eyes, close your ears, do not pay attention, but I like to log into LinkedIn here in my sidebar and then I like to say, "Please use browser to go through my recent unread

好，做产品这块的用例就讲到这儿。接下来我快速过几个我觉得挺好玩的其他用法。我最喜欢的一个，之前也简单提过，就是拿 browser use 对付 LinkedIn。LinkedIn 你把眼睛闭上、耳朵捂上，千万别听——我喜欢在侧边栏里登录 LinkedIn，然后说：“请用浏览器帮我过一遍最近的未读

[18:49]

messages and reply to any that are super critical." Can't type. For Chat PRD or the podcast, leave me notes here on the others about how I should reply. If this is just a friendly thank you note, reply in a friendly way. So, this is just like the easiest way to go through the pile of LinkedIn requests I get. I love you all. Thank you for connecting with me on LinkedIn.

消息，把特别要紧的都回掉。” 打字都打不利索了。“跟 ChatPRD 或者播客有关的优先处理，其余的就在这儿给我留个备注，说明我该怎么回。如果只是普通的道谢消息，客气地回一句就行。” 这大概是处理我收到的那一大堆 LinkedIn 消息最省事的办法了。爱你们，谢谢大家在 LinkedIn 上加我。

[19:24]

I'm happy to be connected, but I get hundreds of these. And so, I love to use Browser Use to just go through LinkedIn, click through stuff, read it. Again, LinkedIn doesn't have an MCP or an API that's official. So, we're using unofficial channels here to go through stuff. One of the other things that I notice is I am using high effort on 56 Soul for this. I probably can use Tara on medium effort, and it might be a lot faster. And so, again,

我很乐意跟大家连上，但我这一天能收到好几百条。所以我特别喜欢用 Browser Use 直接去 LinkedIn 上点来点去、把消息读一遍。再说一次，LinkedIn 没有官方的 MCP，也没有官方 API，所以我们这儿算是走了非官方渠道。我还注意到一件事：我这次用的是 GPT 5.6 的 high effort。其实换成 medium effort 大概就够用了，速度还会快很多。所以说，

[19:53]

you can calibrate your Browser Use effort and level of intelligence to whatever you need. I was definitely overusing it for LinkedIn replying. Um, but again, you know, maybe it requires a little bit of work. All right. So, as you can see, it's pulling up LinkedIn. It's going through my messages. Christie is my friend. We used to be um colleagues. She's asking me to meet with her CEO. We're going to see it reply back a message, go to

你完全可以按自己的需要，去调 Browser Use 的 effort 等级和智能水平。回 LinkedIn 消息我这明显是杀鸡用牛刀了。不过话说回来，这活儿可能确实得多花点心思。好，你看，它把 LinkedIn 打开了，正在过我的消息。Christie 是我朋友，我们以前是同事，她想让我跟她们 CEO 聊聊。我们看着它回一条消息，然后去看

[20:22]

unreads, um and then actually filter through this. So, if you're like me, if you have tons of LinkedIn connections, LinkedIn comments, need to figure out how to get through this stuff, I love using Browser Use to go through those use cases.

未读消息，再把这些筛一遍。所以如果你跟我一样，LinkedIn 上堆着一大堆好友请求和评论，不知道该怎么消化，我特别推荐用 Browser Use 来处理这类事情。

[20:38]

Going to wrap this up with a Browser Use case for the girls. Um, gents, others, if you want to use this one, please do. But, I'm going to use @chrome. So, I'm going to show you how you use the computer use, but specifically for Chrome. So, I've been using @browser,

最后再讲一个 Browser Use 的用法，献给姐妹们。当然，各位老哥、各位朋友想用也尽管用。这次我要用 @chrome。我给你演示一下 computer use 怎么玩，不过是专门针对 Chrome 的。我之前一直用的是 @browser，

[20:54]

which is the side um window of Codex. I'm going to use @chrome, and I'm going to say, "I'm going to Hawaii this weekend, and I need new clothes. Good news!" Exclamation mark. "Free People has 25% off their sale. I would love for you to go through using Chrome browser use and pick out 10 items in my size.

也就是 Codex 的侧边窗口。我用 @chrome，然后我说：「我这周末要去夏威夷，得添几件新衣服。好消息！」感叹号。「Free People 的打折区现在还有额外 25% off。麻烦你用 Chrome browser use 逛一逛，挑 10 件我尺码的。

[21:18]

You can pick size medium. That would be great for Hawaii. Remember, I'm a breastfeeding mom. I just had a baby 6 months ago. So, pick things that are comfortable, that are breastfeeding friendly, um and that would be great for the weather in Hawaii in late July. What I want you to do is end up with 10 things in my cart that I can pick from and buy. Thanks.

尺码选 M 就行。要适合夏威夷穿的。记住，我还在哺乳期，6 个月前刚生完孩子。所以挑那种穿着舒服、方便哺乳、而且适合 7 月底夏威夷天气的。我要的结果是：购物车里躺着 10 件，我可以从里面挑着买。谢啦。」

[21:42]

Always say thank you to your AI. Okay, so I'm going to put that in. It should spin up. It's going to use Chrome to curate 10 medium sale eligible pieces that are breathable, comfortable, and easy to nurse in and then add them to my cart without checking out. And so,

记得对你的 AI 说声谢谢。好，我把这段发出去，它应该马上就跑起来了。它会用 Chrome 帮我挑 10 件 M 码、参与打折、透气舒适、方便哺乳的单品，加进购物车但不下单。所以说，

[21:59]

again, this is something that can be happening over on the side. I do not need to think about it. I do not need to manage it or monitor it. It is, as you can see here, this little browser window will pop up and actually show you what it's doing. I think this is such a cool feature of Codex is that it will actually show you what it's browsing.

同样，这件事完全可以在旁边自己跑着，我不用去想它，不用管它，也不用盯着它。你看，这个小浏览器窗口会自己弹出来，实时展示它在干什么。我觉得这是 Codex 特别酷的一点——它会把自己浏览的过程直接放给你看。

[22:22]

It's going to go through the sale and it's going to pick out outfits for me. Now, we're off to a bad start because I did say that I was a relatively new postpartum mom and it did pick a mini skirt. So, I'm not quite sure. We'll see what good fashion sense that 5.6 ter has. TBD. But, again, just like think about these kind of like annoying shopping use cases that you have to use a browser for that require a lot of like filtering and

它现在开始逛打折区，帮我挑搭配了。不过开局有点翻车，我明明说了自己是刚生完孩子没多久的妈妈，它却挑了一条超短裙。所以我也说不好，咱们就看看 5.6 的时尚品味到底怎么样吧，有待观察。但还是那句话，想想生活里那些烦人的购物场景——非得开浏览器、要一层层筛选、要点来点去的那种——

[22:54]

interaction and you can just go through and have it shop for you. And now, I'm going to pull up this browser and show you Oh, uh-oh. Bad news bears. Look, it thinks that I'm an agent, which I am. Sorry. This is much cuter. So, you might have to be the hands for your AI. I tell people this all the time. Sometimes, we put AI to use for us, and then sometimes, we need AI to put us to use. And so, sometimes,

现在你完全可以让它替你去逛去买。接下来我把这个浏览器窗口调出来给大家看看。哎哟，不妙。你看，它认为我是个 agent——我确实是，抱歉了。这个验证图案还挺可爱的。所以有时候你得给你的 AI 当那双手。我经常跟别人讲：有时候是我们让 AI 干活，有时候反过来，是 AI 让我们干活。所以有些时候，

[23:28]

I'm just the human verification layer on top of Chat GPT browser use. So, again, this is Sneak a Shop. We will see how successful it is. If it adds stuff to the cart that I would really like. But again, I hope this is getting your mind open on what browser use and computer use can do for you. Okay, I'm going to list just a couple other things that I use browser and computer use for that you might find interesting. One,

我就是套在 ChatGPT browser use 外面的那层「人工验证」。好，这就是所谓的偷偷摸鱼式购物。咱们看看效果如何，看它往购物车里加的东西是不是真合我心意。总之我希望这些例子能打开你的思路，让你看到 browser use 和 computer use 能帮你干什么。好，我再列几个我自己常用 browser use 和 computer use 的场景，你可能会感兴趣。第一，

[23:53]

filling out really, really annoying forms. So, if you ever have to do procurement forms, if you have to ever ever have to register your kid for camp, if you ever have to fill out an old, old, old website, um browser and computer use is really great for you.

填那种烦到极点的表单。比如你要走采购流程的表单，比如你要给孩子报夏令营，比如你要在一个古董级别的老网站上填一堆东西——browser use 和 computer use 简直是救命。

[24:08]

Second one, browser and computer use can actually run your phone. I did this the other day. So, if you're on a Mac, you have something called iPhone mirroring where you can actually pull up a version of your phone on your desktop. Computer use can use iPhone mirroring. So, your computer can technically use your phone.

第二个，browser use 和 computer use 其实可以操作你的手机。我前几天就这么干了。如果你用 Mac，有个功能叫 iPhone 镜像，可以把手机界面投到电脑上。computer use 是能操作 iPhone 镜像的。所以严格来说，你的电脑可以替你用手机。

[24:30]

How did I use this recently? Well, I was out of state, and my Mac minis were sitting here on my desk here in California, and I needed to hack into my Wi-Fi. So, I needed to open up some firewall gateways so I could remotely SSH in. Sorry, I'm just going to really need to do so I can like remotely SSH into my MacBooks and like revise the open clause cuz it wasn't on the same network. My Wi-Fi I software is on my phone, so I opened up Codex, I opened up

我最近怎么用的？当时我人在外地，我的几台 Mac mini 摆在加州家里的桌上，我需要「黑」进自己家的 Wi-Fi。我得把防火墙的几个端口打开，才能远程 SSH 进去。抱歉，说清楚点——我要远程 SSH 进我那几台 MacBook，去改一下上面跑着的 OpenClaw，因为它们不在同一个网络里。而我的 Wi-Fi 管理 App 在手机上，于是我打开 Codex，打开

[24:58]

screen mirroring, and I had Codex use computer use to set up the right settings on my home Wi-Fi, update my router, SSH into my MacBooks, and then close all the ports when it was done. It was pretty amazing. It was really good. So, you can definitely use computer use to use your phone. And the last thing that I do is I use computer use a lot to make like Google documents and Google Sheets. So, sometimes I'm just not happy

iPhone 镜像，让 Codex 用 computer use 去把家里 Wi-Fi 的设置改好、更新路由器、SSH 进我的 MacBook，弄完之后再把那些端口全部关上。真的挺惊艳的，效果非常好。所以 computer use 确实可以拿来操作你的手机。最后一个我常干的事，是用 computer use 做 Google 文档和 Google Sheets。有时候我就是不太满意

[25:26]

with what is being used with like the MCPs and the plugins in Codex. So, if it hits kind of like access use or it can't do something via an MCP, if you are permissive enough, you can just say like, "Please just open up my browser and do it in the browser." and it will do a very very good job. So, look, now my Free People cart has 10 sale items,

Codex 里那些 MCP 和插件跑出来的结果。所以如果它卡在权限问题上，或者某件事通过 MCP 做不了，只要你的权限设置足够宽松，你就可以直接说：「你就打开我的浏览器，在浏览器里做吧。」它会做得非常非常漂亮。好，你看，我的 Free People 购物车里已经有 10 件打折商品了，

[25:47]

all size medium, totaling $352 before tax. Tax, the extra 25%. What did it give me? A maxi dress, great. A skirt set, great. Loosey-goosey pants, amazing. Shorts, great. T-shirt, oh, look at this tank. Perfect. Perfect, perfect. A blouse, a button a couple button-downs, a little V-neck.

全是 M 码，算上额外那 25%、税前一共 352 美元。它给我挑了些什么？一条长裙，不错。一套裙装，不错。宽松阔腿裤，绝了。短裤，可以。T 恤——哎哟看这件背心，太完美了。完美，太完美了。一件衬衫，还有两件开襟衬衫，一件小 V 领。

[26:11]

It did a really really good job. So, this turned this started with browser use as a debugging tool and ended with browser use as your personal shopper. Again, this is my favorite feature, y'all. If you have not tried it, especially with Codex, it is a 100 out of 10.

它做得真的非常非常好。所以你看，这期从把 browser use 当调试工具开始，最后落到把 browser use 当私人采购。再说一遍，这是我最喜欢的功能，各位。如果你还没试过，尤其是没在 Codex 上试过，满分 10 分我给它 100 分。

[26:34]

Everything is computer, and I really think that this is going to change how we interface with our digital lives. And again, my goal for most of AI is I can be not on my computer and stuff can still get done. So, I can enjoy my Hawaii vacation. This has been a very fun mini episode on browser use and computer use. If this is something you find interesting, let me know in the comments. I would love to hear how you use computer browser use,

Everything is computer（万物皆电脑），我真心觉得这会改变我们跟数字生活打交道的方式。还是那句话，我用 AI 的大目标就是：我人可以不在电脑前，但事情照样能办完。这样我才能安心享受我的夏威夷假期。这期关于 browser use 和 computer use 的迷你节目就到这里，非常好玩。如果你觉得这个话题有意思，欢迎在评论区告诉我。我很想听听你是怎么用 computer use 和 browser use 的，

[27:00]

what questions you have about privacy, what questions you have about security, and other things I can show you about how I AI. Thanks for joining this week and I'll see you all soon. Thanks so much for watching. If you enjoyed the show, please like and subscribe here on YouTube or even better leave us a comment with your thoughts.

你对隐私有什么疑问，对安全有什么疑问，以及还有哪些「我是怎么用 AI 的」内容想让我讲。感谢这周的收看，我们很快再见。非常感谢观看。如果你喜欢这期节目，请在 YouTube 上点赞订阅，或者更好——留言告诉我们你的想法。

[27:20]

You can also find this podcast on Apple Podcasts, Spotify, or your favorite podcast app. [music] Please consider leaving us a rating and review, which will help others find the show. You can see all our episodes and learn more about the show at how I AI pod.com.

你也可以在 Apple Podcasts、Spotify 或者你常用的播客 App 上收听我们的播客。[音乐] 也欢迎给我们打分写评价，这能帮助更多人发现这个节目。所有往期节目和更多信息都可以在 howiaipod.com 上找到。

[27:37]

See you next [music] time.

下期[音乐]再见。