Robert Ta demos OpenAI Codex's Record-and-Replay plugin, which turns recorded desktop workflows into reusable agent skills. Record a multi-step process on your Mac, and the plugin captures screen actions, metadata, and your spoken narration to build a skill that can classify statements as hallucinated or factual. It replaces long prompt engineering with short video demos.
OpenAI Codex lets you record your screen once and turn it into a reusable skill. And here's what it means for anyone building agent workflows that touch desktop apps, browsers, or manual processes. You could stop writing your 50-line prompts to describe a workflow your intern could learn by watching you do it once and start recording a 30-second demo that becomes a permanent agent skill that you can make better. So how Record and Replay works, you show Codex a workflow once, it literally watches you and it, you turn it into a reusable skill. Record and Replay lets you demonstrate a workflow on your Mac and turn it into a reusable skill. Use it when the workflow is repetitive, depends on your preferences, or is easier to show than to describe in a prompt. For example, you might record how you file an expense, book a parking space, create a correctly configured issue, publish a video, or download a recurring report. Codex can package the pattern into a skill that you can use again with computer use, browser actions, connected plugins, or a combination of them. Okay, so this is a demo for AI builders on using the new plugin for Codex on how can you record your workflow manually on your computer, and Codex will codify it as a skill. Any manual process, browser work, desktop work. So I'm going to go ahead and just start the recording here after downloading and installing the plugin and giving all the Mac accessibility permissions for the file system. So what I'm gonna try to automate is in Cloud Code here, what I normally do is, uh, I'll create a bunch of threads, the platform by Meta for content. That's part of my top of funnel go-to-market strategy and content strategy. What I then do is I go here and I go to Obsidian. This is my preferred way of doing it. I kind of use Obsidian as just my markdown reader against my file system here. And I'll go here and I'll look at the threads and here I just do a human pass over this. It's already ran and evals. Check against an extensive rubric after about a month of running this Threads agent harness. And then I could read through this. I'm like looking for actually areas where I'm most concerned about is where it talks about my previous employment. And I would go and look for the ones that sound really specific, not talking in generalities. Like this, this makes sense. When I was lead product architect at Workday, we evaluated vendor reliability before deep integration. Simple question, will this company still report this product? So I'll go to here and I'll say, I evaluated threads. Here are examples of good, bad usage of my career, where good is generic enough to be an umbrella of work towards my facts I've given you, and bad being we've made up something super specific that is a lie and hallucination. Here's a good one. There's a bad one. Cool. So, what it's gonna do now is go and evaluate all of the ones around this particular criteria and then regenerate this. Let's go to Codex and stop. I'm done recording.
Cool. I really like that you also narrated it. I'm curious if it's gonna pick up on that and include your narration within the skill prompt that it creates.
Yeah, I'm really curious about that as well. Find review note in Obsidian, locate examples of career context claims, copy good and bad examples, then feed them into internal Codex prompt to calibrate against factual grounding versus hallucination. Checking the skill creation utilities next so the result is discoverable, not just a note. It seems like it did capture the audio context. I'm quite impressed. So they must have a really good transcription of vowels that performs really well within even just a normal conversation turn. For this kind of flow, I would have imagined as a user it would have taken me much longer, uh, if they're doing all of this synthesis of context on the backend to then fulfill the query.
Really impressive how deeply integrated it is with the operating system. And, you know, like their chronicling feature, you can see this too. They have a metadata event timeline of everything you're doing in the system. It's actually not intuitive to conceptualize what the system— what Codex has access to directly, you know, because it seems like it has system logs, it has metadata, It has screen share capability.
Use this skill to keep personal career context claims useful without letting the model invent precise biography, employer, client, title, timeline, metric, or project details. The goal is not to remove all personal authority, is to distinguish broad supporting framing from unsupported specificity. Oh, it even goes into some level of classification. It's a generic supported, needs citation, hallucinated. Yeah, for a one-shot, pretty damn good.
Zero-to-one skill like that, That is impressive. And it might not be amazing, right? But then there's also tools to improve it.
Yeah, I could easily see this being used to just save, even just 80/20, save you some time for a manual workflow. And then you compound the skill over time and improve your evals rubric.
