Inside How OpenAI Uses Codex to Do Product Work | Rohan Varma

频道: Peter Yang
视频: https://www.youtube.com/watch?v=fAdFE7y6K2o
原文语言: en
统计: 共 114 轮 · Rohan 57 · Peter 50

[0:00] Rohan

Aim for something that seems 10x more ridiculous than you ever think it could and it probably can do like 90% of it. You cannot walk from point A to point B in the OpenAI office without hearing Codex get mentioned. Over the course of the day, I'm probably stewing three or four x more than I would have without Codex. Codex actually knows how to set up its own automation. It just comes back to you. You didn't have to think about what it was doing while it was doing it. It's literally not a question of what is possible. It's just everything is possible. It's just what should we do?

目标要定得比你想象中夸张 10 倍，它八成也能做到九成。在 OpenAI 办公室里，你从 A 点走到 B 点，路上一定会听到有人提 Codex。一天下来，我的产出大概是没有 Codex 时的三四倍。Codex 甚至知道怎么给自己搭自动化，它会自己跑完再回来找你，过程中你根本不用操心它在干什么。现在真的不是「什么能做到」的问题了——一切都能做到，问题只剩「我们该做什么」。

[0:26] Peter

All right. So, why don't we start by talking at a high level, you know, I've been a PM for over a decade too and and like a lot of the work do is just like, you know, editing Google Docs, making plans, doing alignment meetings, like all this kind of stuff, right? But like, how has Codex fundamentally transformed how you work as a PM?

好，那我们先从宏观聊起。你知道，我也当了十多年 PM 了，很多日常工作无非就是改 Google 文档、做计划、开对齐会这些，对吧？那 Codex 是怎么从根本上改变你作为 PM 的工作方式的？

[0:42] Rohan

I would say that there's two levers by which it's changing how we work as a product managers. One is the actual how we do the the specifics of the job. But then secondarily is that now with the entire team, you know, engineers, product managers, designers, all using Codex, it's just like the role that everyone's playing is slightly different and so it's not just doing the old stuff differently, but even just the way we operate is very different. Uh I think that I mean fundamentally, I would say a lot of if you look at what product managers were historically doing, what we've been up to, it's a lot of information synthesis and kind of like building ideas around lots of different interactions with different customers and I think that at least at OpenAI, we are pretty light on the product side, you know, it on Codex, we only have a few PMs um working across a lot of different areas and so uh fundamentally, this is only possible because of Codex. Um at a baseline level, you know, we get thousands of every day, like probably hundreds of different problems or data points or bugs that we're getting from like enterprise customers or you know, obviously customers on Twitter and things like this. Um and so, I'd say at the the front end of this, it's Codex is extremely valuable for just synthesizing tons of information um and kind of enabling someone like myself to operate with way more context than I would have been able to previously. I'll be added last minute onto an entirely new project that I in 20 minutes can immediately just have all of the context of stuff that happened before I joined, you know, cuz Codex can go into all of our tools, connected into things like Notion and Linear and email and Google Drive. Um and so, I think that everybody, especially the product managers, just they're able to be way more uh leveraged and like quickly jump into different things. Um and then I'd say the actual, you know, beyond just like being able to make decisions with that information, it's all of the ways that we generate artifacts that fundamentally change, right? Whether it's slides, uh you know, prototypes, docs, PRDs, all of these things I think are now either a lot of these things that previously required a lot of effort now can actually be done via automation.

我觉得它从两个杠杆上改变了我们做产品经理的方式。一个是具体怎么干活本身；另一个是，现在整个团队——工程师、产品经理、设计师——全都在用 Codex，每个人扮演的角色都有点不一样了，所以不只是用新方法做老事情，连整个协作方式都变了。从根本上说，如果你看产品经理过去在做什么，其实大量工作是信息综合——从和不同客户的各种交流里拼出想法。至少在 OpenAI，我们产品侧配置非常轻，Codex 团队只有几个 PM 要覆盖很多不同领域，说白了这只有靠 Codex 才可能。最基本的一层是，我们每天会收到成千上万条信息，大概几百个不同的问题、数据点或 bug，来自企业客户，当然还有 Twitter 上的用户。所以在最前端，Codex 在综合海量信息上极其有价值，让我这样的人能带着远超以往的上下文去工作。我可能临时被拉进一个全新的项目，20 分钟内就能掌握我加入之前发生的所有来龙去脉，因为 Codex 可以进到我们所有的工具里——它接入了 Notion、Linear、邮件、Google Drive 这些。所以每个人，尤其是产品经理，杠杆都大得多，能很快切进不同的事情。再往后，除了能基于这些信息做决策，我们产出各种交付物的方式也从根本上变了——不管是幻灯片、原型、文档还是 PRD，很多以前要花大力气的东西，现在真的可以靠自动化完成。

[2:38] Rohan

Um

嗯……

[2:38] Peter

Yeah.

对。

[2:39] Rohan

So then it allows us to actually just spend a lot more time focusing on decision-making and kind of thinking through things and kind of we're talking to customers more importantly, so. Um yeah, I mean, happy to we can dig up There's a lot of different ways that it's changed, but I think fundamentally it's I think we're just spending a lot more time doing the more important stuff, which is like talking to customers and working with our teams to build product and a lot less time doing a lot of the more manual stuff.

所以这让我们能把更多时间花在做决策、深入思考上，更重要的是花在和客户交流上。当然，具体改变的地方有很多，我们可以慢慢展开，但根本上我觉得就是：我们把更多时间用在更重要的事上——和客户聊、和团队一起打磨产品——而花在那些偏手工活儿上的时间少多了。

[3:01] Peter

Yeah. Dude, there's like a meme going around that like uh the plans are for agents, they're not even for humans anymore. So,

对。哥们儿，现在有个梗在流传，说计划是写给 agent 看的，根本不是给人看的了。

[3:07]

[laughter]

[笑声]

[3:08] Peter

and you know, dude, like there's like so much time wasted on just like writing these like in the old days writing 10-page PRDs or like long-term strategy or whatever. Like I'm I'm I'm really curious cuz I know the Codex team works really hard and moves really fast. So, like how how much planning do you guys do and like how how is Codex involved in the process?

而且你知道，以前写那种 10 页的 PRD 或者长期战略文档，真的浪费了太多时间。我特别好奇，因为我知道 Codex 团队干活很拼、节奏很快——你们到底做多少计划？Codex 又是怎么参与这个过程的？

[3:25] Rohan

I mean, we're definitely I'd say in general being at OpenAI, we are all we we operate very much a research lab even what how we develop product. And so, we're constantly, you know, trying things. And so, we're not planning things out 6 months in advance. I think I've I've I don't know what the right I've always I've been thinking we need a new new a new word for roadmap because roadmap kind of implies this 12-month out thing. I think it's more what's the 48-week plan. We know what those priorities are and let's execute against those. And then 4 weeks from now we'll see what the new priorities are. Um I think that the way that Codex really influences a lot of this stuff is basically one is we can just build things pretty pretty much instantly. And so often times I'd actually say that we spend a lot of time looking at solutions, deciding if we want to actually ship it. I think that a lot of the work, you know, there'll be an idea for example, the actually the in app browser is a really good example of this that we added a couple months ago, which one of the engineers on our team, Adam Fraser, he's awesome. He Basically one morning he I showed up to work and he's like, "Look, I built the browser in this." Cuz it was really annoying doing front end iteration. "What do you think?" And I was like, "That's awesome. We should definitely figure out how to ship this." And so then that was almost a starting point was having an actual in product branch that we could actually develop off of. And I think that's a good kind of mental model is there's just a lot It's like the product development life cycle is kind of inverted in a lot of ways because before it was very much spend all this upfront time planning and really making sure that engineers are only doing the most important things. And now it's actually a flip that it's like, "Do everything and then let's figure out what actually should go through and in what ways it should."

在 OpenAI，我们整体上非常像一个研究实验室在运作，连做产品的方式也是。我们一直在不停地尝试，不会提前规划 6 个月。我一直觉得我们需要给 roadmap 换个新词，因为 roadmap 总让人联想到 12 个月的长线规划，而我们更像是「未来 4 周的计划」：明确这几周的优先级，全力执行，4 周后再看新的优先级是什么。Codex 对这套打法的影响在于：我们几乎可以瞬间把东西做出来。所以很多时候，我们其实是花时间在评估已经做出来的方案、决定要不要真的发布。举个特别好的例子——我们几个月前加的应用内浏览器。团队里有个很棒的工程师叫 Adam Fraser，有天早上我到公司，他说：「你看，我把浏览器做进去了」——因为之前做前端迭代太烦了——「你觉得怎么样？」我说：「太棒了，我们一定要想办法把它发出去。」于是起点就变成了一个可以直接在上面继续开发的产品内分支。我觉得这是个很好的心智模型：产品开发生命周期在很多方面被倒转了。以前是花大量前期时间做规划，确保工程师只做最重要的事；现在反过来了——先把所有东西都做出来，再决定哪些真正该往前推、以什么形式推。

[4:55] Peter

Got it. That's actually really amazing for our profession I guess cuz cuz you're saying basically like the engineer basically built the MVP already and he can gain that iteration loop internally of like giving feedback and make it better and just like working on the actual product instead of like some document, right?

明白了。这对我们这个职业来说其实挺美妙的，因为你的意思是，工程师已经把 MVP 做出来了，你可以在内部直接进入迭代循环——给反馈、把它改得更好，是在真正的产品上打磨，而不是在某个文档上，对吧？

[5:09] Rohan

Totally. And in general I would say we try to have work You know, everyone within OpenAI is extremely AI pilled and very much at the frontier of thinking about curiously how can we leverage AI better, how can we improve our products in this way, but also our engineers are someone that is very product minded. And so which I think is actually really important with AI with tools like Codex because human to human collaboration becomes a big bottleneck. If you think about each of our projects we really are just have the one or two engineers working on it in its entire, you know, product lines. And so in a mode where there's so many engineers there's actually a ton of human, you know, collaboration overhead. But then also you're moving slower and so then the the product decisions might happen a couple times over a week. Whereas when an engineer can move extremely quickly, there's a ton of micro decisions that ideally they could just make themselves and not need to work with me. And so often when we start when we start a project we really what I try to do is just think through what are the most important questions that need kind of decisions that we need to make about this thing to feel it's cohesive to our overall strategy, is successful, all of this stuff almost like guardrails of what we're doing. But then beyond that I think our engineers and our designers move pretty autonomously and product is more kind of what the other the two ends of the process taking it to market but also kind of the more strategic planning up front.

完全是。总的来说，OpenAI 里每个人都极度拥抱 AI，一直站在前沿去琢磨怎么更好地利用 AI、怎么用它改进产品，而且我们的工程师也非常有产品思维。我觉得这一点在用 Codex 这类 AI 工具时特别重要，因为人与人之间的协作会变成大瓶颈。你想想，我们每个项目——甚至整条产品线——其实就一两个工程师在做。如果工程师很多，人际协作的开销就会非常大；而且节奏一慢，产品决策可能一周才发生几次。但当一个工程师能跑得飞快时，会冒出大量微决策，最理想的情况是他们自己就能拍板，不需要来找我。所以每次启动一个项目，我会做的就是想清楚：关于这个东西，哪些是最重要、必须做的决策，才能让它契合整体战略、能够成功——差不多是给这件事立好护栏。除此之外，工程师和设计师就相当自主地推进，产品更多是把住流程的两端：一端是推向市场，另一端是前期偏战略的规划。

[6:25] Peter

Got it. You just kind of trying to keep track of what like all the stuff that the engineers are working working on, right? And kind of

明白了。所以你主要是想跟得上工程师们手头在做的所有事情，对吧？然后……

[6:30] Rohan

Yeah, well I mean honestly there's a way to keep track of all of it.

对，不过说实话，还真有一个办法能把这些全跟住。

[6:32]

[laughter]

[笑声]

[6:32] Rohan

Codex is my only way. Every day there's probably something a term I hear and and this is one of the ways I use Codex which is whenever I'm in meetings or literally anytime I don't understand something or I hear something I didn't know about I just have Codex go and gather all the context. And I would say a couple times a day there's like something that someone's doing where I'm it's incredible that they're doing it. I had no idea it was happening and that's amazing.

Codex 是我唯一的办法。几乎每天都会碰到某个我没听过的词——这也是我用 Codex 的一种方式：开会的时候，或者任何时候只要有我不懂的东西、听到没接触过的事，我就让 Codex 去把所有上下文收集回来。而且我得说，一天里总有那么几次，我发现有人在做某件事，做得特别厉害，而我之前完全不知道它在发生——这种感觉很棒。

[6:50] Peter

That that's amazing, dude. So so yeah, it really has to flip, right? It's not like the PM is like trying to come up with all the ideas. It's like the opposite, you know? It's it's

太酷了，哥们儿。所以整个逻辑真的反过来了，对吧？不再是 PM 拼命想出所有点子，而是恰恰相反。

[6:57] Rohan

No, I mean and I mean we're like super lucky. I mean this is also true at Cursor. Obviously we're building a tool for ourselves, you know? Fundamentally it's funny because you cannot walk from point A to point B in the OpenAI office without hearing Codex get mentioned and someone talk to someone else be like oh I asked Codex to do this or like I I really think that if Codex went down or we didn't have access to it like at this point our company would be we'd be operating it would be it would be a pretty tough time and I think I think that that was originally why we built it in the first place is like we want to use it to you know, years ago we started Codex as this project to accelerate our own research. Um now I mean it turns out that that can just help us accelerate everything. Um and so then we can you know, we have the intuition over how it should work and what should be better about it.

没错。而且我们真的特别幸运——这一点在 Cursor 那边其实也一样——本质上我们是在给自己造工具。说来好玩，在 OpenAI 办公室里，你从 A 点走到 B 点，路上一定会听到有人提 Codex，有人跟别人说「哦我让 Codex 干了这个」。我真心觉得，如果 Codex 挂了，或者我们用不了它了，公司运转起来会相当艰难。当初我们做它也正是为了这个——几年前 Codex 立项就是为了加速我们自己的研究。现在发现，它能加速我们的一切。也正因如此，我们对它该怎么工作、哪里该做得更好，有非常直接的手感。

[7:39] Peter

This episode is brought to you by Oceans. I hired someone through Oceans for podcast post-production a few months back and can't imagine running the podcast without his help. He's proactive, picks up new tools fast, and uses AI to compound everything that he ships. Oceans doesn't just place [music] assistants, they place operators. Their talent is AI fluent and delivers the same output as a senior US hire at three to five x less cost. [music] They reject 99% of applicants, so the person who lands on your team is already operating from day one. If you're scaling and need marketing, ops, finance, or EA help, I highly recommend giving Oceans a try. Check it out at oceanstalent.com/peter. Now, back to our episode.

本期节目由 Oceans 赞助。几个月前我通过 Oceans 招了一位负责播客后期制作的伙伴，现在已经完全离不开他了。他主动性强、上手新工具快，还会用 AI 让每一项产出都事半功倍。Oceans 提供的不只是助理，而是真正的操盘手。他们的人才精通 AI，产出媲美美国资深员工，成本却只有三到五分之一。他们会筛掉 99% 的申请者，所以最终加入你团队的人从第一天起就能独当一面。如果你在扩张、需要市场、运营、财务或行政助理方面的帮手，我强烈推荐试试 Oceans。访问 oceanstalent.com/peter 了解详情。现在回到节目。

[8:19] Peter

I imagine it must be incredible working there because um it's it's kind of an interesting spot cuz like yeah, I switched completely to Codex and like pretty much anything you want to automate or streamline, you can just ask Codex to figure it out and like build a scale or something around it, right? At the same time, I feel like OpenAI has like what, like like a billion ChatGPT users. So, how do you get the nor- normies to, you know, get Codex built without overwhelming them? Like that that's that's kind of like the key the key problem, right? So, it's like yeah.

我想在那儿工作一定很爽，不过这里有个挺有意思的点：我自己已经完全切到 Codex 了，基本上任何你想自动化、想提效的事，都可以直接让 Codex 想办法搞定，围绕它搭点东西，对吧？但与此同时，OpenAI 有差不多 10 亿 ChatGPT 用户。你们怎么让普通人也用上 Codex，又不把他们吓跑？这才是关键问题吧？

[8:46] Rohan

Yeah. I mean, it's it's really interesting because I think I mean, I've worked in AI coding for a while and now it's like turns out these agents are super useful for everybody. Um but I think developers even haven't fully warmed up to the level of ambition you can have with Codex. Um you know, if you go within to any large enterprise, maybe the top 5 to 10% most kind of AI pilled frontier engineers are maximally leveraging Codex, but then there's a long tail of people that are still, you know, maybe pair programming, doing less delegation, things like that. But it's interesting because engineers have kind of seen AI incrementally expand its capabilities over the last 3 years. But then if you look to uh a non-developer, it's kind of been just ChatGPT this whole time. And so then the type of things that they're asking chat are very different than what you can actually delegate to Codex.

对，这确实很有意思。我在 AI 编程这个领域做了挺久，现在发现这些 agent 其实对所有人都超有用。但我觉得，连开发者都还没完全适应 Codex 能承载的那种野心。你到任何一家大企业里看，可能只有最前沿、最拥抱 AI 的那 5% 到 10% 的工程师在把 Codex 用到极致，剩下一大批人可能还停留在结对编程的用法上，很少做真正的任务委派。有趣的是，工程师们过去三年是眼看着 AI 的能力一点点扩展上来的；但对非开发者来说，这一路以来 AI 就等于 ChatGPT。所以他们问 ChatGPT 的那类问题，和你真正能委派给 Codex 的事，是完全不同量级的。

[9:32] Peter

Yeah.

对。

[9:33] Rohan

And so honestly, I think that is the big opportunity for us is to figure out um you know, what are the ways to just make people feel more ambitious with what they can ask. Uh because ultimately it's there's actually nothing almost everything you should you need to do professionally should almost always just immediately have Codex start doing it. So I I I think that it that is an interesting kind of both product and kind of enablement overall just mentality challenge.

所以说实话，我觉得我们最大的机会就是想清楚：用什么方式能让人们敢提出更有野心的要求。因为归根结底，几乎所有你工作上需要做的事，都应该第一时间直接丢给 Codex 去做。所以这既是产品挑战，也是一个整体上的赋能和心态挑战。

[9:58] Peter

Yeah, I've gotten more ambitious with my requests for it over time. Like I have a skill that, you know, it posts like seven social media platforms at once. And most of them don't even have APIs. So it just use like the browser use to to to to Totally. Yeah.

对，我自己对它的要求也是越来越大胆。比如我有一个 skill，可以一次性往七个社交平台发帖，而且大多数平台连 API 都没有，它就直接用浏览器操作去发。——完全是。对。

[10:09] Rohan

No, the I mean that the browser and that's what I think is so cool is like the approach we've taken very much has been build these really useful primitives for Codex to interact with your computer. And then the magic is when they all kind of combined into this thing where you don't even know you don't even need to know how anything works. All you do is just ask Codex for what you need. And under the hood Codex actually knows how to use itself to then actually do the things you need, right? So the most magical experiences I've had with browser use are not when I ask it to use browser or ask it to use my computer. It's just when it was trying to do something once I was I I built I wrote a Notion doc um and I basically wanted to create a site off of the Notion doc to kind of, you know, prototype something. And so I asked Codex, "Can you turn this Notion doc into a site?" Um which Sites is kind of the the hosted websites we we launched recently. And it was funny because then I went it it did it and then I I went back and looked at the trace and it basically Notion MCV didn't expose getting the actual images in the doc. And so Codex opened up a browser and then just went into the DOM and got the the actual file images. And then used that in the website. And I was like, "That's amazing. I'm glad it did that." Uh so that's kind of that type of stuff feels really magical to me.

对，浏览器这块——我觉得特别酷的一点是，我们采取的思路一直是：为 Codex 打造一批真正好用的基础能力（primitives），让它能和你的电脑交互。而魔法在于这些能力组合到一起之后，你根本不需要知道任何东西是怎么运作的，你只管向 Codex 提需求，底层 Codex 自己知道怎么调用自己的能力去把事情办成。我用浏览器功能最神奇的几次体验，都不是我主动让它去用浏览器或操作电脑，而是它自己在完成任务时想到的。有一次我写了一个 Notion 文档，想基于这个文档生成一个网站来做原型。于是我问 Codex：「你能把这个 Notion 文档变成一个网站吗？」——Sites 就是我们最近发布的托管网站功能。有意思的是，它做完之后我回去看执行轨迹，发现 Notion 的 MCP 没有暴露获取文档内图片的接口，于是 Codex 自己打开了浏览器，钻进 DOM 里把真正的图片文件抓了出来，然后用在了网站上。我当时就想：「太绝了，幸好它这么干了。」这类事对我来说才是真正有魔力的。

[11:20] Peter

Yeah, dude. It's like it's a lot more persistent than humans in some ways. It just keeps trying to figure it out.

是啊哥们儿。它在某些方面比人还有韧劲，会一直想办法把事情搞定。

[11:25] Rohan

Totally. And I think that's I mean it's funny like I mean I think that's why our goal feature is, you know, feels really useful because there is generally the our goal with OpenAI and with Codex and generally with our agents is we want it to just keep getting better at doing harder and harder, more ambitious, longer hard time horizon tasks. Um and often some of that's at the harness level, but some of that's also just at the scope of, you know, continuing to prompt it. Before we had goal, I think a lot of what people would do is just add a bunch of steers where it's okay, keep going, and then keep going, and then keep going.

完全是。这也是为什么我们的 goal 功能让人觉得特别有用。OpenAI 也好、Codex 也好、我们的 agent 整体也好，目标都是让它在更难、更有野心、时间跨度更长的任务上不断变强。这一部分靠 harness 层面的改进，一部分也靠持续地 prompt 它。在有 goal 功能之前，很多人做的其实就是不停地追加指令：「好，继续」，然后「继续」，再「继续」。

[11:54] Rohan

And I think now it's, you know, it'll just it'll just keep going now until it actually meets the criteria, which I think is awesome.

而现在，它会一直自己干下去，直到真正达到设定的标准为止，我觉得这很棒。

[11:59] Peter

Got it.

明白了。

[12:00] Peter

I think you prepared like a few demos for us. Do you want to kind of share some PM workflows that you use?

我记得你为我们准备了几个 demo。要不要分享一些你在用的 PM 工作流？

[12:04] Rohan

Yeah. I'm going to share kind of in a dummy environment cuz we're not able to share the things I'm doing in OpenAI directly, but it'll hopefully be able to illustrate some of this stuff. So, there's a few different ways that I I mean, there's a lot of different things that I do with Codex. I think one thing actually, and so just to orient everyone, this is the Codex app. Um this is kind of, you know, an an agent development environment that we've built. It's kind of an agent control plane. Um and so there's a lot of different things. The most fundamental thing is, you know, you can start a chat um with a model, all this stuff. It can work within a file a project or just on its own, but um so there's a few different ways that I use Codex that are really valuable to me. So, one is that at OpenAI, you know, we have a ton of chat We use Slack very heavily. Um and so we have a ton of Slack channels um that we get feedback from customers, our internal team send feedback, that we discuss ideas. And so one of the things I use Codex for is actually uh to just synthesize feedback, kind of highlight important things, um and then actually upload it into a linear board that I can review later. And so this is kind of illustrative of the what I did, which is essentially there's a thread that I wrote in Codex where I asked it for feedback. Um I actually, you know, also spent a lot of time asking it to figure out how to correctly input it into linear and what are the ways that make sense. Um and then often, and this is very true for a lot of other workflows, I'll then just ask Codex to set up an automation to do this continuously. And so so, great is that Codex actually knows how to set up its own automations. And so, Codex was actually able to go in and say, "Great, now do this every day." Um, or do this every week. And then, you know, things that I would often also do is say, "You know, also send a Slack message uh to me every time it finishes." Um, and then Codex can go in and update the automation. Um, and so, there's probably a five or six different automations like this that I've set up where it's like a manual process that I then ask Codex to do once, and then I just ask it to, you know, go and set up the automation. Um, and so, this is probably a big I would say this is a huge unlock um for a lot of things. And there's also really interesting creative ways that you can use automations. Something that um I'll do often is, you know, basically tell Codex, uh you know, there's some there was some sort of message that got sent or well, I'll have sent a message, and then based on the result of that message, I'll ask a question to one of my coworkers. So, I'll be like, "Hey, when Alex responds to this message, please draft two the last DM I sent him. Um, drop a response email to the client." Um, and you know, so, like basically what Codex will do is actually use automations under the hood to kind of keep checking uh the Slack message that I had with Alex, and then essentially once it triggers, it'll draft an email, and then actually delete the automation. And so, I think the the thing that's really cool is using Codex knows how to use itself. Um, and so, you can just ask it to do pretty much random things that aren't very specified, and it'll just use its own primitives to do that really well.

好的。我会在一个演示环境里展示，因为我在 OpenAI 内部做的东西没法直接给大家看，但希望能把这些用法讲清楚。我用 Codex 的方式有好几种——其实是非常多。先给大家定个位：这是 Codex 应用，可以理解为我们打造的一个 agent 开发环境，或者说 agent 的控制平面（control plane）。里面能做的事很多，最基础的就是你可以和模型开一个对话，它可以在某个文件、某个项目里工作，也可以独立运行。我用 Codex 有几种对我特别有价值的方式。第一，在 OpenAI 我们重度使用 Slack，有海量的频道——客户反馈、内部团队反馈、各种想法讨论都在里面。所以我用 Codex 做的一件事就是综合整理这些反馈，把重要的东西提炼出来，然后自动录入到一个 Linear 看板里，方便我之后回顾。这里展示的大致就是我做的事：我在 Codex 里开了一个 thread，让它去收集反馈。我还花了不少时间让它琢磨怎么正确地把内容录入 Linear、什么样的组织方式才合理。然后——这一点在很多其他工作流里也一样——我会直接让 Codex 把这件事设置成一个持续运行的自动化任务。妙就妙在 Codex 知道怎么给自己设置自动化，它可以自己进去配置成'每天跑一次'或者'每周跑一次'。我还经常会加一句：'每次跑完之后给我发条 Slack 消息'，Codex 就会自己去更新那个自动化配置。像这样的自动化我大概设了五六个，套路都是：先手动让 Codex 做一遍，然后直接让它把这个流程自动化。我觉得这是一个巨大的解锁点。自动化还有很多有创意的玩法。我经常这么干：比如我发了一条消息出去，然后基于对方的回复我要问同事一个问题。我就会说：'嘿，等 Alex 回复这条消息之后，参考我上次给他发的 DM，帮我起草一封给客户的回复邮件。'Codex 实际上会在底层用自动化机制不停地去检查我和 Alex 的那条 Slack 消息，一旦触发条件满足，它就起草邮件，然后把这个自动化删掉。所以真正酷的地方在于：Codex 知道怎么使用它自己。你可以丢给它各种没怎么明确定义的任务，它会用自己的原语（primitives）把事情办得很漂亮。

[14:56] Peter

Oh, oh, that's interesting. Okay, yeah, I've set up like daily jobs and weekly jobs before, but not this kind of trigger stuff. Like, just like Alex responding and trigger something. So, so, it'll actually like it'll actually do it, and then it'll actually remove it afterwards?

哦，这个有意思。我之前设置过每日任务、每周任务，但没玩过这种触发式的——比如 Alex 一回复就触发某个动作。所以它真的会执行，执行完之后还会把自动化删掉？

[15:08] Rohan

Exactly. Cuz it can update The automation can have it'll it'll set up the automation to actually tell itself to delete the automation when it it it meets the acceptance criteria. Um, and so, you can do really it basically anything that's time based or requires kind of a trigger, you can actually Codex will set up automations to do those things and you don't even need to kind of like be very specific, um which is cool. Um

没错。因为它能更新自动化配置——它在设置自动化的时候，会告诉自己：一旦满足验收标准，就把这个自动化删掉。所以基本上任何基于时间、或者需要某种触发条件的事情，Codex 都能通过自动化来搞定，而且你甚至不需要说得很具体，这一点很酷。

[15:29] Peter

Dude, so you can you can you can do like an automation where like if someone tags Rohan in Slack on long threads, just like pretend to be Rohan and respond. That's cool.

老兄，那岂不是可以设一个自动化：只要有人在 Slack 的长 thread 里 @Rohan，就假装是 Rohan 去回复。太酷了。

[15:36] Rohan

Exactly, yeah. So actually funny Alex also he was just sharing this with a couple of us where he'll like do @Codex but it's not actually a Slack app, it'll just be an automation running every few minutes to see if he tagged Codex and then it'll actually just respond. Um but there's a lot of Yeah, I think automations are honestly kind of an not under uh appreciated but kind of under leveraged, I think, feature that we have because it's so flexible because of the fact that Codex can intelligently kind of update it. Um but so yeah, this is one thing. Or sorry, go

正是这样，哈哈。有个好玩的事：Alex 前几天跟我们几个人分享，他会在 Slack 里 @Codex——但那其实不是一个真的 Slack 应用，就是一个每隔几分钟跑一次的自动化，检查他有没有 tag Codex，有的话就直接回复。所以，我觉得自动化其实是我们一个被低估——或者说没被充分利用的功能，因为 Codex 能智能地更新它，所以它极其灵活。这是第一件事。抱歉，你先说。

[16:05] Peter

Oh, real real quick. My my one problem with automations is like um like a lot of them are local, right? So they'll only run when the computer's on. So some sometimes if I do like a Friday morning thing and my computer's not on, it it doesn't work, you know?

哦，插一句。我用自动化的一个痛点是，很多自动化是本地运行的，对吧？只有电脑开着才能跑。有时候我设了个周五早上的任务，结果电脑没开，它就没跑，你懂吧？

[16:17] Rohan

Totally.

完全理解。

[16:17] Peter

So so I guess I have to put it on a Mac Mini or something.

所以我估计得把它放到一台 Mac Mini 之类的机器上跑。

[16:20] Rohan

Yeah, that is one option. I mean, we did launch uh mobile support for mobile. Um and kind of in you can also control remote uh instances of Codex uh via

对，那是一种办法。不过我们已经推出了移动端支持，你也可以远程控制云端运行的 Codex 实例。

[16:30] Peter

In in in the cloud?

在云端吗？

[16:31] Rohan

Yeah.

对。

[16:31] Peter

Got it.

明白了。

[16:32] Rohan

Um but we are going to So right now we don't have native cloud automations, but that is something that, you know, is in the pipeline, so.

目前我们还没有原生的云端自动化，但这个已经在规划里了。

[16:39] Peter

Okay.

好的。

[16:39] Rohan

Uh but I think it's it's kind of interesting because um I know there's like a meme of a laptop kind of, you know, slightly ajar. Um but I mean, to be honest I originally when I joined I was kind of like, "Yo, you know, the automations being local feels limiting. But I think that in the if you kind of are just using them I mean, for the most part in my whole day my laptop's open, so it's it actually hasn't been a huge problem. Um but I do think it's, you know, we are going to have some solutions there. Um

不过挺有意思的，我知道有个梗——笔记本电脑半开着盖子（让自动化一直跑）。说实话，我刚加入的时候也觉得'自动化只能本地跑，太受限了'。但实际用下来，我一天里笔记本基本都是开着的，所以并没有造成太大问题。当然，我们后面会有解决方案的。

[17:04] Peter

Okay.

好。

[17:04] Rohan

for sure, so. The other the other thing that I think is really interesting that I think is um has really fundamentally changed the way that we work or at least the way that I work is ImageGen. Um so you can actually use ImageGen directly in Codex um if you just tag imagegen the skill it'll actually trigger using the model. Um and so it's it's really interesting because I think this is if you think about the way that we were thinking about using AI to kind of do quick design iterations, it often was writing code to prototype things. But I actually think imagegen is a much faster way to prototype ideas. Um and so this is an just illustrate the example. So I took a screenshot of um you know the the composer where you can type in a new message into Codex and I basically said, "Hey, you know, when you when you launch Codex, there you can select the project." And so I asked uh imagegen or asked Codex using imagegen to basically prototype four or five ideas to improve the way to select the project. Um and so what it did is actually, you know, generated a bunch of different ideas. Um you know,

第二件事我觉得非常有意思，它从根本上改变了我们——至少是我——的工作方式，就是 ImageGen（图像生成）。你可以直接在 Codex 里用 ImageGen，只要 tag 一下 imagegen 这个 skill，它就会调用模型。有意思的地方在于：过去我们用 AI 做快速设计迭代，通常是写代码做原型；但我发现 imagegen 是一种快得多的原型方式。举个例子：我截了一张 Codex 输入框（composer）的截图，然后说，'启动 Codex 的时候可以选择项目，请用 imagegen 出四五个改进项目选择方式的方案。'它就真的生成了一堆不同的设计想法。

[18:05] Peter

Nice.

漂亮。

[18:07] Rohan

immediately. [snorts] And I think what I mean to me it actually this was kind of an AGI moment for me where I was like cuz I think we always think of imagegen as this thing which is useful for making I don't know, turn my hair blue or something. Um but the fact that it can create such good digital mocks essentially um really allows for like super quick iteration, right? This feels like a much quicker iteration to me iteration than kind of generating five different React dummy websites. Um And what I'll often do is say, you know, cool, that was awesome. Now take the first one and make a prototype of it on a um on a site. And so then then something that it creates something that can be shared. Um and then, you know, we can kind of reason over it. So I'd say imagegen was definitely been a huge unlock for prototyping and just generally expressing ideas in product.

而且是立刻就出来。对我来说这算是一个 AGI 时刻——我们总以为 imagegen 就是用来'把我头发变成蓝色'之类的玩具，但它居然能生成质量这么高的产品界面 mock，这让迭代速度快到飞起。比起生成五个不同的 React 假网站，这种迭代快多了。我经常接着说：'很好，把第一个方案做成一个网页原型。'这样它就产出了一个可以分享的东西，我们就能围绕它讨论推敲。所以 imagegen 对做原型、对表达产品想法来说，绝对是一个巨大的解锁。

[18:56] Peter

Like do you just give it a reference screenshot or do you have to give it some sort of like the design system thing or no? Just just a reference screenshot?

你是只给它一张参考截图，还是得给它设计系统之类的东西？就一张参考截图就行？

[19:02] Rohan

Yeah, for this one I mean so just gave the screenshot and it was able it's able to, you know, do a pretty good job of staying within the design language. Um we've also built skills internally at open AI that basically kind of enforce our our design language. When then or it enables Codex to kind of access it when it is generating design assets. The other thing that's interesting is you can integrate with into with plugins like the Figma plugin. And so then that will actually be able to pull things or design tokens and things like that.

对，这个例子里我就只给了截图，它就能相当好地保持在原有的设计语言之内。另外我们在 OpenAI 内部也构建了一些 skill，用来强制执行我们的设计语言——让 Codex 在生成设计资产的时候能够调用它。还有一点很有意思：你可以接入插件，比如 Figma 插件，这样它就能直接拉取设计稿或者 design token 之类的东西。

[19:31] Peter

But like for for the audience just just like a really basic skill is just like here's the colors, here's the fonts, and here's like you know. That's it right? Like that that kind of stuff.

给观众解释一下——一个最基础的 skill，其实就是'这是我们的颜色、这是字体'，就这些对吧？就是这类东西。

[19:39] Rohan

I mean what's great with things like you know an often thing might happen is you do this iteration, you look at this and you say hey this wasn't right quite what I wanted. And so you can keep doing it until it is. And then what you can do is actually use we have something called a skill creator which is actually a skill to create skills. And what you can say is hey Codex look at this thread and create a skill to ensure you are consistent moving forward. And this is actually a pretty common thing that we do which is you you have some sort of interaction with Codex and you just tell Codex to kind of at the meta level do something, make a skill, make a plugin to ensure that this thing gets essentially templateized moving forward.

对，而且妙的地方在于：经常你迭代一轮之后，看着结果说'这不太是我想要的'，然后你就继续改，直到满意为止。接下来你可以用我们的 skill creator——一个专门用来创建 skill 的 skill——跟 Codex 说：'看一下这个 thread，创建一个 skill，确保以后都保持这个风格。'这其实是我们很常见的做法：你和 Codex 有过一段交互之后，就让它在元层面（meta level）做点事——做个 skill、做个 plugin，把这套东西模板化，以后直接复用。

[20:19] Peter

Yeah that that's what I do too. So like so like you know I have skills for doing stuff and then like it doesn't quite get it right one shot sometimes so I have to give it feedback. And then I'm like hey can you update the skill to account for this feedback? But but like dude one thing I worry about is is like cuz I don't actually read the updates that it makes.

对，我也是这么干的。我有各种干活用的 skill，有时候它一次做不对，我就给反馈，然后说：'嘿，能不能把这条反馈更新进 skill 里？'但老兄，我有个担心——因为我其实从来不看它改了什么。

[20:35]

[laughter]

[笑声]

[20:36] Peter

So that I'm I'm worried that it'll just like slopify the skill like you know if I keep making it do new stuff. Do you worry about that too or do you have a yeah?

所以我担心如果一直让它加新东西，它会把这个 skill 越改越水、变成一堆 slop。你也担心这个吗？还是有什么办法？

[20:43]

[snorts]

[轻笑]

[20:43] Rohan

No I I do think about that because I feel a lot of times um I mean and this is kind of a general uh thing that we're kind of thinking about which is a lot of things skills and plugins do introduce a lot of things to the prompt. And so we're starting to think about like you know especially as you have more autonomous agents running is are there ways to kind of eval the output when you make changes to it or you know build confidence. Right now what I'll do is I'll just test it. And basically cool, looks good, um and kind of keep going.

我确实会想这个问题。这其实也是我们在思考的一个普遍性问题：skill 和 plugin 会往 prompt 里塞进很多东西。所以我们开始考虑——尤其是当越来越多自主运行的 agent 出现之后——有没有办法在你改动这些东西之后对输出做 eval（评测），建立信心。目前我的做法就是直接测一下，觉得'嗯，看着不错'，然后继续往前走。

[21:09] Peter

Okay. Cool.

好，明白。

[21:12] Rohan

Um the last thing that I'll share um on the just things that I do that I think are super leveraged is um essentially, often times, and this kind of goes to this idea of like kind of disposable software. It's so easy to create software and tools, and so using Codex, and so often I'll just create one-off things that aren't necessarily reproducible, but for example, you know, we use a lot, and so often there'll be just a ton of messages that I need to respond to. And so I'll actually just on the fly ask Codex to generate little apps for me to work through things that are happening in other tools. Um So for example, a common thing I'll be is, "Hey, look at all my Slack messages, uh figure out the most important ones to respond to, and then actually just show it to me visually within a local website." Um and so then it'll kind of generate, you know, different versions of things like this. Um something I actually have set up also is like when it generates something that feels useful, "Oh, I could actually use this repeatedly." Then I'll actually just like set up an automation to tell it to update this um every, let's say, hour. And so then I can actually just work out of this a little bit. Um so I think that there's a lot of And I'd say like this is where there's a lot of overhang of uh capabilities where I think we're used to thinking about software, these little apps taking a lot of effort. Um And but then now it's so easy, so then you can actually just create extremely custom versions of everything that you need for like the workflows that makes sense for you. Um

我要分享的最后一件杠杆特别高的事，跟'一次性软件'（disposable software）这个概念有关。现在创建软件和工具太容易了，所以我经常用 Codex 做一些一次性的、不一定要复用的小东西。举个例子：我们消息量很大，经常有一大堆消息等着我回。我就会临时让 Codex 给我生成一些小应用，帮我处理其他工具里的事务。比如一个常见的用法是：'看一下我所有的 Slack 消息，找出最需要回复的那些，然后在一个本地网页里可视化地展示给我。'它就会生成各种版本的这类小工具。我还有个玩法：当它生成的某个东西感觉很好用、'诶，这个我可以反复用'，我就顺手设一个自动化，让它比如每小时更新一次，然后我就可以直接在这个小工具里干活了。我觉得这里存在大量的能力过剩（overhang）——我们习惯性地认为做软件、做这些小应用要花很大力气，但现在它变得如此容易，你完全可以为自己的工作流量身定制一切极度个性化的工具。

[22:33] Peter

Got it. So this one is about like uh kind of like a prioritized list of things you need to respond to or actions you have to take.

明白了。所以这个功能其实就是给你一份按优先级排好的清单，列出你需要回复的消息或者要采取的行动。

[22:39] Rohan

Yeah.

对。

[22:40] Rohan

Yeah.

对。

[22:40] Rohan

Another [clears throat] version of this that I'll often do is look at the last day of messages and find all the commitments that I made. Um and then kind of give me a prioritized list of that. And so uh you know, I think that you know, I'm I would say I'm not the most uh organized person in general, but I think with Codex it enables me to just basically delegate organization and persistence to the tool because it can and this is I actually I think I just tweeted about this where it's there's so many things where you have so much to do and often it just the to-do list adds up and you kind of feel oh you know there's some anxiety there and I think what's really relieving is actually you can just ask Codex to do these things and ultimately half probably 80% of the time it'll actually just knock the thing off the list that you needed to do. Um and so for me that's you know basically frees up a ton of mental space to be able to focus on more important things.

我还经常用另一个版本：让它翻一遍过去一天的消息，找出我做过的所有承诺，然后按优先级列给我。说实话，我本身不算特别有条理的人，但有了 Codex，我基本上可以把「保持有序」和「持续跟进」这件事直接委托给工具，因为它做得到。我前几天还发推说过这个——很多时候事情太多，待办清单越堆越长，你会有种焦虑感。而真正让人松一口气的是，你可以直接让 Codex 去做这些事，而且大概 80% 的情况下，它真的能把清单上那件事直接干掉。对我来说，这就腾出了大量的心智空间，让我能专注在更重要的事情上。

[23:28] Peter

Yeah I feel like almost all these work streams including like strategy and everything is like the the machine can do like 80% of it and then the last 20% you kind of apply your touch to it you know and maybe it'll get fewer and fewer over time.

对，我感觉几乎所有这些工作流，包括战略之类的，机器都能干掉 80%，剩下 20% 由你加上自己的判断和风格，而且这 20% 可能还会越来越少。

[23:40] Rohan

Oh yeah.

没错。

[23:41] Rohan

Yeah.

对。

[23:41] Rohan

Uh and I think also there's Yeah it's 80% of the time and maybe some of the time the machine can do it at 10 but it's like I think also there's so much time where as especially as product managers where we just spend gathering context. Uh you get added to a Slack thread and it has like 50 messages you would normally have to sit there and spend like 50 minutes just reading them. Um often what I'll do is just go into Slack copy a thread and then ask Codex to just give me a little TLDR of this. Um and then I can immediately kind of respond and so um I think you just you know over the course of the day I'm probably doing like three or four X more than I would have without Codex um which is a pretty crazy.

而且我觉得，80% 的情况是这样，有些时候机器可能只能做到一小部分。但还有一点：尤其是做产品经理，我们花了太多时间在收集上下文上。你被拉进一个 Slack 讨论串，里面有 50 条消息，按以前你就得坐那儿花 50 分钟把它们读完。我现在经常就是进 Slack 把整个讨论串复制下来，让 Codex 给我一个简短的 TLDR，然后我马上就能回复。所以一天下来，我的产出大概是没有 Codex 时的三四倍，这挺疯狂的。

[24:18] Peter

And there's probably a ton of just like also the product development process I found AI super helpful for like you know like interviewing candidates like summarize interview notes or like writing my own performance review you can probably cuz cuz it has mem- memory right so you can probably just get it to like write a performance review.

而且产品开发流程里应该还有一大堆场景，我发现 AI 都特别有用——比如面试候选人之后总结面试笔记，或者写自己的绩效自评。因为它有记忆嘛，你大概可以直接让它帮你把绩效自评写出来。

[24:33] Rohan

I totally agree. Yeah I would cuz I'm also worse at actually documenting anything that I've actually done. Um I mean I've been a founder most of my career so I usually just do stuff and it's usually Um but it's funny because and that's yeah you can just tell Codex go look at what's all the things I've contributed in the last three months and it'll you know spend a couple hours and really just put together all the definitive list.

完全同意。因为我也特别不擅长记录自己到底做过什么。我职业生涯大部分时间都在创业，一般就是埋头把事干了。但有意思的是，现在你可以直接跟 Codex 说：去看看我过去三个月都贡献了什么。它会花上几个小时，真的给你整理出一份完整权威的清单。

[24:53] Peter

Yeah I imagine meetings are more fun too right cuz there's probably a lot of it is just like playing with prototypes and like playing with the actual product. Right? That's my guess. Yeah.

我猜你们开会也更有意思了吧？因为很多时候就是在把玩原型、把玩真正的产品，对吧？我是这么猜的。

[25:01] Rohan

Yeah, so I mean I think Yeah, especially I mean honestly a lot of because we work in a lot of we work very in person as well. So there's a lot of just serendipitous conversations that we have and that's someone showing the thing that they built and then it's like oh my gosh. You know, let's actually grab a little bit and kind of go from there.

对，确实是。尤其是我们很多时候都是线下当面办公，所以经常会有那种不期而遇的对话——有人给你展示他刚做出来的东西，然后大家「哇」一下，干脆坐下来聊一会儿，顺着就往下推进了。

[25:19] Peter

Dude, so like um So like you know Do you think your schedule is filled with meetings still or you have your actual build time? Like Yeah.

那你现在的日程还是被会议塞满吗？还是说你有真正动手做东西的时间？

[25:28] Rohan

I mean I'm so I mean at Open AI one of the things I I focus a lot on our enterprise motion and so I do spend I mean I do actually probably have a pretty large number of meetings relative to what would be optimal, but a lot of that's actually just with customers. Um So but yeah, I would say my build time is from 5:00 p.m. onwards. Got a lot of evening time, but but no, I think that It's funny because actually I'll even use Codex as my personal chief of staff. So in the beginning of the week I'll have it look at all my meetings, figure out you know, consolidate the ones where it can be more give me focus blocks and figure out which ones could be canceled as a DM or something. So it's definitely helping me find the focus time, but I think in general what we do with the free time is actually just spend way more time with our customers.

在 OpenAI，我很大一块精力放在企业客户业务上，所以相对理想状态来说，我的会确实不少，但其中很多是跟客户开的。我的动手时间基本是下午五点以后，晚上时间挺多的。不过有意思的是，我甚至会把 Codex 当成我的私人幕僚长（chief of staff）来用：每周一开始，我让它看一遍我所有的会议，能合并的合并，帮我腾出专注时间块，再找出哪些会其实发个 DM 就能解决、可以直接取消。所以它确实在帮我找回专注时间。但总的来说，省下来的时间我们其实都拿去花更多时间陪客户了。

[26:14] Peter

I mean yeah, I I think any meeting with customers is definitely worth taking. So yeah, no doubt about that. And let me ask some questions cuz like I I kind of become a Codex power user, too. Like Uh one one thing I struggle with is like I I don't know how people like Peter Peter Steinberger can do it, but like I I can pretty much only work on maybe like three or four Codex threads at the same time. You know? Like how many active conversations do you have going going on at the parallel?

对，我觉得跟客户的会议肯定都值得开，这点毫无疑问。我再问几个问题，因为我自己也算是 Codex 重度用户了。有一件事我一直搞不定：我不知道像 Peter Steinberger 那样的人是怎么做到的，我基本上最多只能同时跑三四个 Codex 线程。你平时并行开多少个活跃对话？

[26:39] Rohan

Yeah, I think I mean Peter's definitely on another level, so I would say maybe don't necessarily immediately start that try that at home, but I think for me um I would say I'm like probably and in general five or six kind of threads. But what's interesting is it's not it's very delegated. So I'm often not waiting for it. It's just I'll fire it off. And then at the end of a meeting or I have a free second, I go back to Codex, I see all the unreads, and I actually then have Oh, actually that was great. It did make the slide deck that I was preparing for for the next some meeting tomorrow, and then I can kind of iterate on it. So I kind of it's kind of a the way that I think about the the paradigm is much more you know, let's say that you have someone that you work with that you manage. If everything they were doing you had to maintain constant state on it, it wouldn't be a very effective delegation. Um and I think with Codex that's the goal is how do we move much more to that fully delegated state where it's it just comes back to you. You didn't have to think about what it was doing while it was doing it. And that I think kind of starts to address that uh just the mental overhead required. Um and like a good example of this is with let's say you're pushing a feature. Uh you want to get a prod get a PR landed. Um some of the stuff that we've built is you can actually ask Codex to nurse that PR, wait for the CI to finish, then fix the CI, and keep going, and respond to any human feedback on the PR, and then like once it's fully ready to go, then ping me on Slack and let me know that I can merge it. And so it's like now you would have previously had to spend 10 different interaction points thinking about that PR, but now it's just at the very end. It's like, "Okay, cool. Like it's good to go." Um and then you can kind of go.

哈哈，Peter 确实是另一个次元的，所以建议大家不要贸然在家模仿。对我来说，一般是五六个线程。但有意思的是，我的用法是高度委托式的——我通常不会守在那儿等结果，而是把任务发出去就不管了。等开完一个会或者有空闲的几秒钟，我回到 Codex 看所有未读，然后发现：哦不错，它把我明天开会要用的幻灯片做出来了，我再在上面迭代。我理解这个范式更像是：假设你手下管着一个人，如果他做的每件事你都得时刻盯着进度，那这个委托就是失败的。Codex 的目标就是往那种「完全委托」的状态走——它做完了自己回来找你，过程中你完全不用惦记它在干什么。这就解决了心智负担的问题。举个很好的例子：比如你要推一个功能上线，想让一个 PR 合进去。我们做了一些能力，你可以让 Codex 去「看护」这个 PR——等 CI 跑完，CI 挂了就去修，继续推进，回应 PR 上所有人类的评审意见，等一切完全就绪之后，再在 Slack 上 ping 我一声，告诉我可以合并了。以前你可能要为这个 PR 操心 10 个不同的交互节点，现在只剩最后一下：「OK，可以合了」，然后你就继续干别的去了。

[28:12] Peter

So have you managed to build a system to manage a system? Like people talk about loops and stuff on Twitter. Have you Are you kind of trending that direction, too? Or or like

那你有没有搭出一套「用系统管理系统」的东西？Twitter 上大家老聊什么 loop 之类的，你也在往那个方向走吗？

[28:20] Rohan

Yeah. I think honestly I haven't I'm I feel like I'm not at the the the next level that Peter is talking about, but it's So I'm still out here promising. I think the But the thing that for me I think I mean honestly I would say uh slash goal is a good version of this where it'll just keep iterating on itself. Um the thing that I'm trying to optimize more for is basically basically the the the general optimization function for me right now is every time I do anything I basically think could Codex have done this? And then there's like but then there's a whole class of so that's a whole class of things. Then there's a whole separate class of things that are like things that we don't even think to do today because they would just be way too painful. Um and so I think that that's actually a really interesting bucket of opportunity for automation. A real example that I've kind of started doing these days is we have project channels um you know, for different products that we're building and or you know, partnership things or whatever. And so I'll actually I've what I've started doing is actually set up for each of these channels a site that essentially has the full state. It's almost like a a project overview. And then I have Codex basically every couple hours for each of for each of these channels like go and essentially get all possible, you know, things that have changed or conversation that's happened and essentially integrate it into the overall site. And so then it becomes this context source for anybody that needs to onboard or on ramp or even for me if I want to see updates.

说实话，我觉得我还没到 Peter 说的那个境界，还在努力中。不过我觉得 /goal 这种功能算是一个不错的版本，它会自己不断迭代。我现在更想优化的是一个总的判断准则：每次我亲手做任何事，我都会问自己——这事 Codex 能不能做？这是一类事情。然后还有完全另一类：那些我们今天压根不会想去做的事，因为以前做起来太痛苦了。我觉得这才是自动化真正有意思的机会点。举个我最近开始做的真实例子：我们有各种项目频道，对应在做的不同产品、合作事项之类。我开始给每个频道搭一个站点，上面承载这个项目的完整状态，差不多就是一个项目总览页。然后我让 Codex 每隔几个小时，对每个频道去抓取所有的变化和新发生的讨论，把它们整合进这个总览站点。这样它就成了一个上下文源——任何需要 onboard 或快速上手的人都能用，我自己想看进展也能用。

[29:38] Peter

Okay.

好。

[29:39] Rohan

So there's that where it's we were never doing anything on this level before, but it's it's actually very useful.

所以这就是那种我们以前根本不会做到这个程度的事，但它真的非常有用。

[29:45] Peter

Yeah, that's actually very smart. So it's like that that becomes the source of truth, right? Cuz like somebody has like Cuz cuz cuz cuz like you know, yeah, like going back and forth in Slack and then manually updating a PRD is such a pain pain in the ass. It's it's just like just get AI to do it.

这招真的很聪明。它就变成了唯一可信的信息源（source of truth），对吧？因为在 Slack 里来回扒消息、再手动去更新 PRD，实在太折磨人了。这种活儿就该让 AI 来干。

[29:56] Rohan

I feel it was just a known thing among everybody that any document is effectively immediately out of date the moment you ship it. And it's but now we actually can create fairly up-to-date documents that are updating themselves, you know.

以前大家都默认一个共识：任何文档在你发布出去的那一刻就已经过时了。但现在我们真的可以做出相当实时、还能自我更新的文档。

[30:07]

[laughter]

（笑声）

[30:08] Rohan

Only they actually become a useful artifact.

这样文档才真正变成有用的产物。

[30:10] Peter

Oh, that that that's super cool. Yeah. What what what what kind of things do you use Go for? Do you do like Go like hit hit my OKR now? What kind of things do you Go Go for?

这太酷了。那你平时都用 /goal 干些什么？会不会来一句「现在去把我的 OKR 完成」这种？你一般拿它做什么？

[30:20] Rohan

Uh what's the most creative out there thing? Um I think for me it's honestly I just always use it. Cuz I think that I'm a fairly latency sensitive most tasks. For the most part the tasks that I'm asking Codex to do it's I just care about it completing it less than I do doing it in 10 minutes or something. Um and so I'd say it's a pretty standard thing that I'm using, but um I'd say the big ones are often when I'm generating from scratch an artifact where I actually don't have as much opinion. I mean often you know I'll go in and say this is what I want let's say I'm generating a side deck or some sort of doc. It's generally know what I want and so then it's just using voice to communicate you know transcription I'll ask it to do it. and brain dump but then sometimes I'll ask it to do things that I don't even know and it's very open-ended. Then I'll go especially for that cuz I'll say go really figure out what's going on everywhere then come up with an idea maybe ask me before you build it then go build it.

要说最有创意、最出格的用法？其实我就是随时都在用它。因为大多数任务上我对延迟不太敏感——我让 Codex 做的事，我更在乎它最终做完，而不是它十分钟内做完。所以我的用法算比较标准。要说大头，通常是让它从零生成一个我自己都还没有太多想法的产物。很多时候我心里清楚要什么——比如生成一份幻灯片或者某种文档——那我就直接用语音转录，把脑子里的东西一股脑倒给它，让它去做。但有时候我会让它做一些我自己都不知道该怎么下手的、非常开放式的事情，尤其是这种情况，我会说：你先去把各处的情况都彻底摸清楚，然后想出一个方案，动手之前可以先来问我，然后再去做。

[31:13] Peter

Got it. Okay. All right man just a few more questions. Like do you have a personal set of skills or like do you most use open AI skills? But like do you have like personal of raw hand skills to apply your own taste and stuff like that? Or Yeah.

明白了。好，再问几个问题。你有没有自己的一套个人 skills？还是主要用 OpenAI 官方的？就是说，你有没有像自己手搓的 skills，用来注入你个人的品味之类的？

[31:24] Rohan

I yeah I actually don't. So I typically am using so I mean we've actually built out a ton of interesting plugins internally. And actually some of them that we shipped externally for specific roles for example design or finance and things like this but for me I actually am pretty light on the skills. I think like part of that for me is I often find that for me it's really useful for me to see where does Codex fall over when it doesn't have any specific stuff. Um I'm I'm I mean honestly maybe I'm just like too you know small brain on the things I'm trying to do with it but it's generally pretty good at doing most things. Um there are a few times though where like I I think what I use skills for is uh defining workflows is maybe like one way to think about it. Um and often I would say it's either in a skill or an automation but it's like okay cool we just did this thing in a chat for like 20 minutes. I want to do this again probably a couple times a week. Now turn that skill or turn that into an automation or something like that.

其实我还真没有。我们内部确实做了一大堆很有意思的插件，其中一些还针对特定角色对外发布了，比如设计、财务这些。但我个人在 skills 上用得很轻。部分原因是，我发现观察 Codex 在没有任何专门配置的情况下会在哪里翻车，这对我特别有价值。当然也可能是我拿它做的事情太简单了，但它对大多数事情的完成度已经相当好了。不过确实有一些时候我会用 skills——我理解 skills 的一个角度就是「定义工作流」。通常要么做成 skill 要么做成自动化：比如我们刚在一个对话里花 20 分钟搞定了某件事，而我以后每周大概要重复做几次，那就把它固化成一个 skill 或者一个自动化。

[32:20] Peter

Got it got it. Okay got it. Um and then and then uh do you have any just like um high high level tips for using Codex for some of the more Codex-feel people? Like I I I remember Jason published an article about how he uses Codex and um something the first one was like just like keep talking to do a long running thread which I didn't really do before this. Like do you have any personal tips or anything? Yeah.

明白明白。那对那些还不太熟 Codex 的人，你有没有什么高层面的使用建议？我记得 Jason 发过一篇文章讲他怎么用 Codex，第一条好像就是「持续对话，维持一个长线程」，这个我以前还真没这么干过。你有什么个人心得吗？

[32:40] Rohan

Well, one thing that was a huge unlock is that we recently shipped Codex can now control threads. So, you can kind of have a high level orchestrator thread that's actually spawning other threads. Um and things like that. So, that's definitely I think a great way to use the tool. But for me, I would say my biggest advice would be or the biggest unlock is one, make sure that Codex is integrated into all your tools. So, sometimes that might require you know, building custom CLIs and things like that. But often it's just you know, as simple as we have a plugin for the tool and just go integrate it. Um and then I would say and then what's great though is even if it doesn't have the tool connected, you can just use computer use often. The other thing I mean is honestly just thinking through what are all the stuff that you're doing and why can't Codex do it? Or like every time you run into an issue or either it's like did you ask it to do it? If not, then great, just try. And then if you tried and it didn't work, then I think some folks it's easy to see it not do something correctly and then kind of just be like, "Great, that's I'll just do it myself." But it's I think if you always ask the question of like what was missing? What almost as a human, how do we know that it's a didn't do the right thing? Probably we have some context or some information that Codex didn't have. Yeah. And then it's just answering the question of what what what did Codex not have and then figure out how can I give it to it so that in moving forward it has it. Then it's kind of creates this acceleration towards just automating more and more of the things that you're doing.

有一个巨大的解锁点：我们最近上线了 Codex 控制线程的能力——你可以有一个高层的 orchestrator 线程，由它去派生其他线程，诸如此类。这绝对是用好这个工具的一个好方式。但对我来说，最大的建议或者说最大的解锁点有两条。第一，确保 Codex 接入了你所有的工具。有时候这需要你自己写一些定制的 CLI，但很多时候就是简单地装个现成的插件、把它接上就行。而且很棒的是，就算某个工具没接上，你通常还可以用 computer use 兜底。第二，认真想想你在做的所有事情，问一句：为什么 Codex 不能做？每次遇到问题，先问自己：你让它试过吗？没试过？那太好了，直接试。如果试了没成功——我觉得有些人看到它没做对，就很容易「算了，我还是自己来吧」。但你应该始终问一个问题：它缺了什么？作为人类，我们凭什么知道它做错了？多半是因为我们掌握了某些 Codex 没有的上下文或信息。那接下来就是回答：Codex 缺的到底是什么？我怎么把它喂给它，让它以后一直都有？这样就会形成一种加速循环，让你把越来越多的事情自动化掉。

[34:03] Peter

Yeah, you can also like string a bunch of skills together into like a whole workflow. Just have it like keep keep running. Like I I found that super help- helpful. Yeah.

对，你还可以把一堆 skill 串成一整条 workflow，让它一直跑下去。我发现这招特别好用。

[34:12] Rohan

I'm curious for you what changed the shift when you were like, what made the shift of for you or the aha moment of this feels really useful or Codex specifically can really help me with things?

我很好奇，对你来说转折点是什么？就是那个「aha 时刻」——什么时候你觉得这东西真的有用，或者说 Codex 真的能帮上你的忙？

[34:20] Peter

Um I I think what blew me away was like browser use. I really think I don't know who worked on it, but I really think it's like the number one feature Codex has. Uh so, it totally blew me away because for example, even for this podcast, like to turn this transcript that we're talking about into a turn it a bunch of assets, right? Like newsletter post, like social post, like all this kind of stuff. I got to post it on all these different platforms. And then I I built this like whole automation where it it it strings together a bunch of different skills. Like it turns this transcript into a newsletter post based on my examples. There's like a remove AI slop skill that that I run to remove all the slop stuff. And then and then there's another skill to post on the social media platforms. Like a lot of it like I said it doesn't have APIs, so it uses browser uses picker out. Recently I made another skill, this thing called hyperframes. So you can actually make video assets based on like text, right? I have another skill to make slides, like HTML slides. It is crazy man. You can string all this together. Oh yeah. I mean like you know, I don't I don't just like auto automate the stuff. Like there's a bunch of creators out there who just automate everything like just pump out tons of slop all the time. So I actually check the stuff, but like it it seems a lot of busy work. You know?

嗯，真正让我震撼的是 browser use（浏览器操作）。我不知道是谁做的这个功能，但我真心觉得它是 Codex 的第一功能。它彻底把我震住了。举个例子，就拿这期播客来说，要把我们现在聊的这份 transcript 变成一堆素材，对吧——newsletter 文章、社交媒体帖子之类的，还得发到各个平台上。于是我搭了一整套自动化流程，把好几个不同的 skill 串在一起：先有一个 skill 参考我以前的范例，把 transcript 改写成 newsletter 文章；然后有一个「去 AI 味」的 skill，把那些 AI 腔的东西清理掉；再有一个 skill 负责发到各个社交平台。像我说的，很多平台没有 API，它就用 browser use 自己去点。最近我又做了一个叫 hyperframes 的 skill，可以直接根据文字生成视频素材；还有一个做 HTML 幻灯片的 skill。太疯狂了，兄弟，这些全都能串起来。当然，我不是无脑全自动——外面有一堆创作者什么都自动化，天天批量生产垃圾内容。我是会亲自把关的，但这些确实原本都是大量的琐碎杂活，你懂吧？

[35:23] Rohan

Really? I mean it's cool because I don't even for product It's funny because we talk about the bottlenecks. You write code more fast with Codex. There's all these like SDLC bottlenecks. But then it's funny because I think there's then becomes even bottlenecks in other parts of the org when it comes to shipping so much software, right? Um for example, we ship so much product at a pace that it's hard to create marketing assets for them. Really high quality videos and demos and things like this. And so our DevEx team, you know, from our own team, they developed some skills to actually use Codex to generate nice kind of visual graphic videos.

是吗？这挺有意思的，因为我们经常讨论瓶颈问题。有了 Codex 你写代码更快了，软件研发流程（SDLC）里那些瓶颈被打通了。但好玩的是，当你能发布这么多软件的时候，组织里其他环节反而成了新瓶颈。比如我们发产品的节奏快到市场素材都来不及做——那种高质量的视频、demo 之类的。所以我们的 DevEx 团队，就是我们自己团队的人，开发了一些 skill，直接用 Codex 来生成漂亮的可视化视频。

[35:56] Peter

Nice.

赞。

[35:57] Rohan

Which also now they can actually ship demos much faster and things like this. So I mean I think it's like every time there's a bottleneck it's you know, the solution is just find a way for Codex to help unlock it a little bit.

这样他们现在也能更快地做出 demo 之类的东西。所以我觉得，每次出现瓶颈，解法都是想办法让 Codex 来帮你把它解开一点。

[36:06] Peter

Yeah. Yeah, I I feel like the models are already beyond kind of like most people's ambition of what they should be doing. It's just like I think that's that's my point and number one tip like it's just like ask Codex to try to do it. That's just it's like be persistent. Yeah.

对对。我觉得模型的能力其实已经超出了大多数人的想象力边界。这就是我的观点，也是我的第一条建议：直接让 Codex 去试就完了，要有耐心、够坚持。

[36:21] Rohan

Yeah, you almost you should aim for something that seems like 10x more ridiculous than you ever think it could and it probably can do like 90% of it. And it's like, okay, great. Now I need to reset to be 10x more than that.

对，你几乎应该瞄准一个比你想象中夸张 10 倍的目标，它大概率能完成其中 90%。然后你会想：好，太棒了，那我得把目标再调高 10 倍。

[36:30] Peter

So what do you think, dude? So So like do you think cuz I I feel like the PM profession like has gone through like a troll where like you know, people just like grow up like cross-functional alignment secretaries like for for for a lot of roles. But I actually feel pretty optimistic about PM now because of all these new tools that you guys are shipping. Like what what do you what do you think? Do you think it's going to become a lot more fun as a job or it probably already is for you, right? It's it's just Yeah.

那你怎么看，兄弟？我感觉 PM 这个职业经历过一段低谷，很多岗位上的人慢慢活成了「跨部门对齐秘书」。但现在因为你们在发布的这些新工具，我对 PM 这个职业反而挺乐观的。你怎么看？你觉得这份工作会变得有趣得多吗？或者对你来说可能已经是了，对吧？

[36:54] Rohan

I mean I think I mean yeah, I'm having fun.

嗯，我确实干得挺开心的。

[36:57]

[laughter]

（笑声）

[36:58]

Yeah.

是啊。

[36:59] Rohan

A good time, but I mean I think yeah, I feel everything that's happening only makes pretty much every role more leveraged. Not even just PM, but every role that's kind of like I mean really what it unlocks is being able to just do way more and own the end-to-end way more and work on way more things. Like I think what it and it's I kind of interesting. I would say AI you can feel AI leverage meaning you can do the same thing faster. But I actually think where we see what we've been able to do and like the most AI companies I've worked at Open AI and Cursor, it's the way that we operate is not doing the same thing small thing faster, it's actually just doing way more things and doing all of them faster relative to the number of people you have. And so I think as a someone who just loves building useful tools and working with customers and just I don't know, just like delivering useful experiences to people that delight them, it's cool to be able to just do way more of that. Um and you're feeling unconstrained, you know, I think uh there's literally I can't remember the last time where it was like it's literally not a question of like what is possible. It's just everything is possible. It's just like what should we do? Um pretty exciting.

确实很享受。不过我觉得，现在发生的一切几乎让每个角色的杠杆都变大了——不只是 PM，是所有角色。它真正解锁的是：你能做多得多的事，能更完整地端到端负责，能同时推进多得多的项目。有一点挺有意思：大家一般理解的 AI 杠杆是「同样的事做得更快」。但从我们的实际经验看——我待过的这几家最典型的 AI 公司，OpenAI 和 Cursor——我们的运作方式并不是把同一件小事做得更快，而是直接做多得多的事，而且相对于团队人数，每件事都做得更快。所以作为一个热爱打造有用工具、喜欢跟客户打交道、喜欢给人交付让他们眼前一亮的体验的人，能做更多这样的事真的很酷。而且你会感觉毫无束缚——我已经想不起来上一次纠结「这事到底可不可能做到」是什么时候了。现在根本不是可能性的问题，一切皆有可能，问题只剩下：我们应该做什么。挺让人兴奋的。

[38:00] Peter

Yeah, you just got to set up a system to do it. Yeah. But but like But like I I I I think I think maybe a downside or maybe a plus is like you don't have you don't actually end up working less. You actually end up work working more cuz there's so much stuff that you can do, right?

对，你只需要搭好一套系统去做就行。不过——这也许是个缺点，也许是个优点——你并不会因此工作变少，反而会干得更多，因为能做的事情实在太多了，对吧？

[38:12] Rohan

Yeah, but I mean I think yeah, I mean I would say for me it wasn't a acceptance criteria was to work less. So I I mean I'm not doing this, you know. That I've always been someone who works a lot, but I think um yeah, I mean I don't know that's like I think that we're going to see some interesting shifts as these tools get better and better at automating more and more, for sure.

是啊，不过对我来说，「少干点活」从来不是验收标准，我做这些不是为了这个。我一直是个工作量很大的人。但我觉得，随着这些工具在自动化上越来越强，我们肯定会看到一些有意思的变化。

[38:28] Peter

I will have all I mean Chris you and team man is is really you guys really turn cortex around it's like probably my number one tool. Uh and and um where can people find you online and um keep following your journey?

Rohan，你和团队真的把 Codex 做起来了，它现在大概是我的第一号工具。那大家可以在哪里找到你、持续关注你们的动态呢？

[38:40] Rohan

Yeah, I mean I'm on Twitter and LinkedIn just [laughter] Rohan Verma. Um yeah, I'm posting every now and then trying to just let people know what we're shipping, but um yeah, I really appreciate you having me on and it's been fun to see all the stuff you're progressing kind of it's been a it's been a fun journey.

我在 Twitter 和 LinkedIn 上，搜 Rohan Varma 就行（笑）。我时不时会发点东西，让大家知道我们在发布什么。真的很感谢你邀请我来，看着你一路做的这些东西也很有意思，这一路挺好玩的。

[38:55] Peter

Yeah, dude I I love the memes like people resetting limits all time and also some of the snark too. I I love the I love it. Yeah.

哈哈，我特别喜欢那些梗图——大家整天刷新用量额度的段子，还有那些小吐槽，我太爱了。

[39:02] Rohan

It's funny because I dropped in we're just actually communicating on Slack or sorry on Twitter. Currently I'll someone post feedback to me or someone will tag me like yo, can you look at this on Twitter? Yeah, I got you.

挺好玩的，我们其实就直接在 Twitter 上交流——现在经常有人给我发反馈，或者 @ 我说「哥们，能看看这个吗」，我就回「好，我来处理」。

[39:12] Peter

Yeah, that's what the best product teams do they're all over Twitter, dude. So it makes sense. Yeah.

对，最好的产品团队就是这样的，天天泡在 Twitter 上，兄弟。所以这很合理。

[39:16] Rohan

Oh, memes are great. So.

哈哈，梗图确实很棒。

[39:17] Rohan

I did good chatting man.

聊得很开心，兄弟。

[39:19] Peter

Yeah, appreciate it. Cheers, dude.

谢谢你来，干杯，兄弟。

[39:21] Rohan

Cheers.

干杯。