Codex の Record and Replay：一度見せるだけでエージェントに面倒なワークフローを教える（リプレイ）

As smart as today's AI systems are, do you ever wish that they would just watch you do all your work and then magically do it all for you? Well, that's what we have now with a new feature from OpenAI and Codex called Record and Replay. because here's the reality. As we are continuing to shift from the responsive chatbot era to the proactive agent era, there's still a lot of manual work that we have to do, whether that's scheduling and automating your workflows, creating new skill files by hand or whatever it may be, Whether you're using ChatGPT, Codex, Claude, Copilot, or Gemini, there's a lot of manual work that needs to go in to create these repeatable and autonomous processes. but I think that may start to change and I think that with what this new feature is in Codex Record

and Replay we have maybe a preview of what's coming next as we have these extremely powerful super apps that live on our desktop and can do work for us. So that's what we're going to be going over on today's show but here is the big picture of Codex Record and Replay. So I think Prompting is kind of losing to demonstration, and maybe that's not a bad thing, even if you've spent a lot of time refining your prompt engineering skills or how you type to a chatbot that could be leaving by the wayside. So OpenAI shipped this new feature less than a week ago, and essentially you demonstrate something on your actual computer, on your Mac right now, and it becomes an editable skill.md file, then the agent can reuse. So the AI interface, I think, is moving away from just typing instructions to just showing your work and showing a series of tasks that go into a project that you may want to start automating.

And I think that this right now, it seems like a small feature, right? And I'm actually surprised this one didn't blow up more than it did. But I think this is a big step in transferring agency and workflows to autonomous systems. Because the other thing why I think this is actually a big deal, because this kind of OpenAI is showing their cards in terms of model capabilities, because now through whatever magic they did, they are now can understand and ingest video, which is huge because this has been a big advantage that Google Gemini has had for the last two years as the only kind of big player that you could upload a video to and it could actually understand what was going on inside that video. So that's a bigger picture. Maybe we'll talk about that a little bit more. But here, stick with me for the next 25 minutes. This is going to be a short one on today's show. Here's what you're going to learn. You're going to learn what OpenAI actually changed inside of Codex beyond just the launch demo.

You're going to know why the skill MD or the skill markdown file is the real business artifact here that leaders should understand. You're going to know how the shared format changes the Codex versus Claude comparison. Yeah, stick around for that. And you're going to know which workflows you should try to record first without creating unnecessary security risk. All right. I hope you're excited for this. This is Everyday AI in our weekly segment, Putting AI to Work on Wednesday. So once a week, I get a little brave. I demo some live technology. It's unedited, unscripted. We'll see how it goes. Sometimes these hands-on demos are amazing. Sometimes we get bugs, but that's the reality of generative AI and working on the cutting edge. sometimes you get cut. All right. So if you haven't already, please make sure to subscribe to the podcast if this is helpful. And then make sure you go to your everyday AI.com. Sign up for the free daily newsletter. And we're going to be recapping not just the highlights from today's

show, but all the other AI news and developments you need to know to be the smartest person in AI at your company. All right. So without further ado, yeah, let's get dirty, shall we? All right. So So this one's going to be a little tricky because I'm going to be jumping back and forth between some different screens, some different things I have going on. But we're going to try to make this work. And this is one of those that if you are listening on the podcast, this might be one of those ones that you should tune into for the video version. So you can always see that video version of this on our website at your everydayai.com. All right. I'm just making sure doing one more check here before I start sharing my screen that I don't have anything too private. All right. So, yeah, not going to be doxing myself or anyone else or I'll only be doxing myself slightly. All right. So let's go ahead and share this screen.

And one thing to know, this is a plug in. OK, so if you don't know too much about Codex, I've done a couple of shows recently on Codex. But if you have a paid ChatGPT subscription, then you have a Codex subscription. It is the same. You know, it uses the same subscription, which is really nice. So you can go download. Codex is a desktop program, whether you use Mac or Windows. Although these a lot of the features come to Mac first, like record and replay. All right. And then after you are logged into Codex, if you haven't downloaded it already, go ahead and do that. Follow along with me. And then you're going to make sure to go into your upper left hand corner. There's a plug in section. All you're going to do is search search for record and replay. Then you are going to add the plug in. After you do that, if you are brand new to Codex, you're probably going to have to go through your Mac security settings and grant it some access to some different privacy features that it needs. because as an example, this uses computer use.

So you do give it access to essentially read and write to your entire computer. So, you know, if you are using this in an enterprise setting, make sure to get the permissions you need before doing this. Understand the pros and the cons and the risks. And that just goes for not just the record and replay in codex, but for anything desktop, right? So whether you talking about codex cloud you know desktop using cloud code cloud cowork anti whatever you using on your desktop always keep that in mind So after you have Codex you logged in with your ChatGPT plan you installed the record and replay plugin Here's what we're going to do. So I'm going to do this. It might seem a little backwards. All right. But stick with me here. I'm doing it like this for a reason, mainly so hopefully the demo can work. All right. So live stream audience, let me know if you can see my screen here. All right. I think we're pretty good. Not giving away too much. Aside from the fact, yes, I do have, you know, Codex and Cloud Code running here at the bottom, like I always do. And a couple other secret projects I'm working on down there.

Okay, so I'm going to put in a prompt first describing what I'm going to do. And you might be saying, Jordan, that defeats the entire purpose. Yes and no. Okay, because my example here requires a little nuance. So also think of what this is and what it isn't. Because right now, this feature, all it does is it sees your screen. So I do think that they're going to add in the feature where you can talk to it at the same time. So you can talk it through and walk it through visually at the same time. But essentially, this is going to record my screen, everything that I do. So because I'm doing a live demo and I want this to hopefully work and, you know, not have to accidentally log into all these different places, I'm first going to give it a precursor on what I'm doing. So I'm just putting in a simple prompt. And for this, I'm going to go ahead. I'm going to go ahead, put this prompt in. OK, nothing crazy here. I'm just telling you what I'm going to do. So I want you to think of that manual thing that you do that that usually requires what I call kind of this context carry.

So, you know, carrying contacts from different programs. I'm doing something extremely simple, very straightforward. All right. I'm just having my thing that I'm trying to automate. This is something I don't have time for, even though it's not super time consuming. It's just all of these small little tasks. When I think about them in my head, I'm like, oh, this requires three, four, five different websites, all these different steps. Even if it only takes 10 minutes, a lot of times I just don't do that because my day to day schedule is absolutely bugger bananas. Right. In my mind, even if a task is 10 minutes, if it requires me going into five, six, seven, eight different systems, doing a lot of copying and pasting, carrying contacts over, even a little bit of brain power. Sometimes I just don't do them, which is why I really like this new feature. So all I'm doing in this prompt, I'm saying for this chat, I'm going to be using the record and replay eventually, but do not use it yet. And I'm just saying, here's a primer on what I'm going to do before I hit record. Yes, I know this is kind of counterintuitive, but because it requires a little bit of nuance, this is why I'm typing it out. And the reason being what I'm trying to do

is I'm going to go into my Beehive newsletter. That is our email newsletter provider. And I'm going to find the five most clicked on stories from our last newsletter. But here's the thing. Most days, that's our actual website, right? It's everyone clicks on the episode page. They go to our website. I don't care about that, right? So what I'm ultimately trying to do is I want to see the five most clicked stories that are not ours, right? that aren't also not something, you know, sometimes we link out to Twitter or, you know, a lot of times it's Spotify. People click on our Spotify link. So, you know, I can't just go click on the five, you know, top stories. I do have to explain that nuance. And Codex would get this right because this is an editable skill that ultimately this is going to do that then I can run, you know, whenever I want by calling on the skill or I can make this a daily automation, which is what I'm going to do. But all I'm going to do, the five most popular stories, take all the text in there, copy and paste it, put it into Notebook LM, and then I'm going to create an audio overview, the debate style.

So essentially every day I can wake up, I can know the five most important stories that our audience thinks about. Not only that, the content of everything, have it be extremely relevant, but then have two hosts, right? I love Notebook LM. To have two hosts kind of debate different sides of those stories. so I can hopefully, right, better understand how our audience is thinking about these things. All right. So that's all I'm doing. So I went through, I told it, this is what I'm doing. So now I'm going to click the app button. All right. And then I'm going to click record and replay. All right. And so all I'm doing is saying, you know, let's start. All right. So I just clicked the app button. I typed in record and replay. I click let's start. So now what should happen is it's going to say, okay, we're recording now. And then it's going to create a little recording thing at the bottom. And I want to make sure I have everything ready. Yes, I do. I'm going to drag this over here, make sure I'm not showing anything. All right, we're good. Yeah, I can't even show all my notebook LFs because, yeah, there's some private things in there.

All right, so here we go. I'm just doing inside codex. I at tag record and replay. I said, let's start. It should. Let's see if it goes here. All right. It says it's reading the skill. It's thinking. There we go. And so now I have a little recording at the bottom of my screen. So I'm going to try to go as fast as I can, but I'll talk you through it for a podcast audience. Nothing crazy is happening here, but Codex is now recording my screen. I do have dual monitor, so hopefully this will work OK. All right. So now I'm going into my Beehive account. I'm logged in. I'm going to the newsletter. I'm going to yesterday's. yesterday's newsletter. I'm going to, where am I going? Performance. Yes. Okay. So now here's all the different ones, the different clicks. These are all the different stories. So I'm going to click what I think are five of the most read stories. I'm just going in order here. All right.

I want to make sure all of these are actual news stories. There we go. It looks like we have them all. Okay. One is just for GLM, which just takes you to their website. So that's not going to tell me anything. So I'm just actually going to Google it because it looks like we just linked to their actual website. But let me just go ahead and grab the actual announcement post there. There we go. Okay, so I think I'm good. All right, this one couldn't be reached for whatever reason. All right, we're good, we're good. What is this This was a random tool All right so I not gonna do that one We gotta verify myself via Cloudflare Yeah this is fun All right Let do one more I do want to make sure that we have five that we can actually copy and paste and grab the text All right. It looks like this Fortune article is not working. Cool. So maybe Codex will see that I can't copy and paste this because it's, you know, apparently behind a paywall.

All right. We'll see if Codex picks up on that or not. but again, this is editable. So I can always go and say, if you can't access, um, the text, then, you know, go down to the next story. Okay. So it looks like we're good here. So I'm going to go ahead and grab all of the texts from these. So I'm going to copy and paste. Um, I'm going to go into copy text. I'm going to paste insert next. I'm clicking, uh, let me just X out of that one. That was something about NVIDIA, about the water sustainability. I'm going to click on add sources. Here's OpenAI announcing how they're using AI to diagnose rare genetic diseases. I'm pasting that into Notebook LM. Again, I'm just copying and pasting the context of all of these pages. All right, next, there was a micron and anthropic announcement. I'm copying and pasting that. I'm adding this inside of notebook.lm. We have three stories so far.

Let's get two more. Copy and pasting the GLM 5.2. All right. Putting that in there. There we go. And then our last one here, something about someone leaving DeepMind for Anthropic. Let's go ahead. Actually, there's a ton of extra stuff on that page. So we'll see if Codex picks this up that there's a lot of extra stuff here. It's a shorter article. So I'm going to copy and paste that in. And done. All right. The next thing, and this is all laid out in the instructions. I want Codex to go on here, click audio, and then, but I want to customize this. So there's a little carrot. I'm going to click that. I want it to go to the debate. All right. And then I actually have a prompt that I'm going to want it to copy and paste. And I told it this. So the prompt, the custom prompt that I'm using for this audio overview is this. The included sources are web page contacts of the most clicked articles from the Everyday AI newsletter. Please simulate how business, I'm going to say how busy business leaders who are leading AI implementation at their companies.

That's you guys, right? Maybe interpreting these stories and debate how each side thoughtfully thinks about these. With pinpoint detail, avoid generalities or soft takes. All right. And then I told Codex once I hit generate, it is done. So let me go back. All I have to do is I just have to click the record button one more time. And then what that does is it sends a message to Codex actually that says I'm done recording. OK, so now what it's what it's saying is it says I'm stopping the recorder now. Then I'll inspect the capture event stream directly. After that, I'll build the reusable skill rather than just summarizing the demo. So pretty good. So I'm going to give that a minute to cook as we go over some of the other details. And then hopefully we'll see it at the end. And I'll see if I can actually, you know, go and run this skill.

We'll see how long it takes. All right. Also, I just realized I forgot to do a couple of things that I told Codex I was going to do, like renaming the notebook, which I didn't do. So we'll see what it picks up on some of the nuance there. But let's just quickly go over what's new. I talked about a lot of this as I was going through. But right now, this is Mac only. And during the recording, Codex observes everything you do. So actions, what you click on, you can do different programs. So I just told it to go to a specific Chrome profile. So when it builds this skill, it is going to tell itself to use that Chrome profile. So Codex has all of these different skills, right? So the big, one of the big differences between Codex and ChatGPT is, well, Codex is an autonomous coding agent. Don't worry if you're not technical. All that means is it can use your computer. It can use any programs, any other desktop programs. One of the most used things I have Codex do is to use Claude Code on the desktop, to use Google Gemini on the desktop.

So I just haven't used all my other desktop programs. It can read and write files, all that good stuff. But Codex observes everything you do. So different actions you take, what you click on, what you highlight, what you type in, et cetera, and also the window content. And then those observations in order sequentially essentially become this skill file that changes what you're doing, right? So when the recording stops, and this is what's happening now live on my other screen, Codex starts to create this skill. And this is what I think is really important because a skill is an editable file, which is really nice. At any time I can go in and I can update this, but here's the big unlock that I don't think most people are talking about. Aside from the fact like, oh my gosh, apparently, uh, JPD55 via Codex can ingest and understand video files, which is a first for OpenAI at scale. So before it could use Python, it could go in, you could upload a video, It could look at things frame by frame. But as far as I know, this is the first time that OpenAI slash Codex slash ChatGPT can actually ingest and understand video.

And like I said, this is a huge advantage that Google has had for like a year and a half. It's something, this process here, I've actually been doing this well before Codex recorded replay. I was using Google Gemini because I would dictate something. I would upload these little things and I was creating skills. I was already doing this. But again, I'm a little dorky. I always try to break these things. But this is pretty big. But keep in mind, this skills.md file, this is an asset that doesn't just have to live inside of Codex. Obviously, it's going to work best in Codex because everything that it writes, you know, if it's using different skills, if it's using different apps, it's going to be tagging those in a Codex manner that Codex can read. However, you can use this in Claw Desktop, right? You'll have to go in, but you can import this skill into Claw Desktop. You can change, you know, if it does call a Codex specific skill, you can change that over. if it's, you know, using Claude's or sorry, if it's using Codex computer use or Codex Chrome, you know, you can go and update those things accordingly. But essentially you are using this record feature to literally do anything right And this is why I think Codex and Claude can actually become more connected now This is you know don think of this as just locking yourself into Codex Although I do think the harness and Codex is much better than Claude desktop and all the benchmarks show that

But if you are a heavy Claude desktop user and that's where you get the majority of your work done, but you still also have a Chatech PD account, you should probably go in and give this a try because that shared structure matters because now these replays, the recording replays can be used across many different surfaces. And remember, it works through your screens, your browsers and your plugins. So you can use other saved skills. You can use other plugins because this is an editable feature. So anything that I would have typed in codex, right? All I did when I started, I just said, let's start, right? I could have typed something out and said, let's use this skill. Let's use these apps. You know, you're going to watch me work on my computer. I just did something simple that would hopefully work. All right. Also, scripts and connectors can increase the reliability, but screen control, you know, it does increase those trust requirements. So if you are rolling this out, you know, in your enterprise, make sure that you get all those proper permissions that you need. And the limits right now are privacy and scope.

So, you know, not everyone has this. You do have to be on a Mac. If you are in the EU, UK, Switzerland, you don't have access to this. And I would always start off with low stakes workflow. So what I just showed, right, this isn't something that's being published live. This is something that I'm using internally. This is still a very new feature. Even for me, I've been using it a lot. I like it. It's still a new feature. So I would start by using this with low stakes things. So, you know, I would go through list your weekly computer tasks with all the steps and clear success criteria. So for me, the success criteria for this is, okay, is it going to accurately find those five things? And I open my notebook LM app every morning. If I go take a walk, I can listen to this and be better up to date on probably, hopefully what all of you are thinking. So I would start just recording one short workflow using realistic inputs, no secrets, no passwords, nothing like that. Read each skill when it's done before rerunning it or setting it up as a scheduled automation.

And you do that in Codex literally. So it looks like this is done recording. It called this skill just the Beehive Notebook LM debate. All right. So now I can say run this every day at 6 a.m. All right. That's all you have to do. And then every day at 6 a.m., I can open my Notebook LM app and it's that going to the studio section and it's waiting for me. All right. So we can read each skill file, refine preferences and build your automation library. All right. Let's check in. I don't think I'm going to have time to rerun this skill, but I do want to check it. All right. So it says done. So it took three minutes. Oh, wow. Impressive here. That was 167 events capture. And I would go through, always encourage people, read the chain of thought. So I'm going through here. I can see exactly what it did. Make sure everything is correct. I'm looking through this. It looks like it's doing a pretty good job. I don't see anything that's wrong. And then it says validation passed with official skill validator.

The only hiccup was something with PyYAML, so something with a programming language there. And instead, it kind of wrote its way around that. So pretty good. And then it says that any time I can invoke this later using the Beehive Notebook LM debate skill. So as an example, let me just open this. I'm going to say run the Dehive notebook skill right now. All right. So we'll just see, make sure it's going to kick off. It should start taking over my browser here. So we'll probably see that a little bit in the background just to show everyone that things are, well, hopefully running and working. And there we go. So it's running it now. I'm not going to sit around for another three to 10 minutes and blab on because I wanted to keep this one tight. Hey, under 30 minutes. There we go. But it's going. It's working. And now from here, I would validate it. I would run it once. I would make sure everything is correct. I would go listen to it on my Notebook LM app. I would see what I wanted to tweet. Once I get it exactly how I want

it, then I schedule it. You schedule it and forget it. But as we wrap up today's show, I want you to really understand what this unlocks because a couple of things. Number one, it is that I think now the easiest way to create skills, skills still to this day, skills and, you know, plugins, which is, you know, Codex's version of using skills and apps. So skills and plugins are, I think, probably the new measure, right? It's not about how many tokens you and your teams use. It's about how often is your agent running a skill or a scheduled skill. I think that's huge because that's where you are starting to actually win back time by running repeatable workflows without you, the human, having to go in, type something, find some, oh, where's this Google Doc? Where's this Obsidian Vault? Where's all this at, right? So, you know, really recording your mundane workflows across different browser tabs. Obviously, you can do things where you're

logged in like this, right? Obviously, I was logged into my Gmail to reach Notebook LM. I was logged into my Beehive account. So I told it to use my Chrome profile because I keep those logged in. So think about all of those things that you do over and over and over, right? And not having to write something out by hand, although I kind of did a little bit of that just to ensure that this would hopefully work zero shot. But think of what this unlocks and what this changes. And it doesn't require literally anything. It doesn't require any technical skill. I'm not an overly technical person. You go in, if you have a ChatGPT account, download codecs, go through again, assuming you have all permissions, record these workflows and it's done. Right. And now you have a worker working for you that you just showed it what to do and it did it. And I think this is a breath of fresh air, a very small feature that is going to make a big difference. So thanks for tuning in. I hope this was helpful doing our AI Wednesday at Work series. Like I said, I've done a couple other Codex walkthroughs.

So make sure you go check out 781, episode 781, where we went over Codex sites. Another great one with, I think, a lot of value there. So make sure you go check that out. And then go to youreverydayai.com. Sign up for the free daily newsletter. We're going to be recapping the highlights from today's show and a whole lot more. Thanks for tuning in. Hope to see you back tomorrow and every day for more Everyday AI. Thanks, y'all.