How Codex Record & Replay Turns a Manual Workflow Into a Reusable Skill

There is a new plugin in the Codex app called Record & Replay, and the pitch is simple: do a repetitive task on your computer once, narrate it out loud while Codex watches, and it builds a reusable skill that can run that workflow again on its own.

In this demo I tested it on one of the most tedious jobs I do every week, uploading and configuring a YouTube video draft. Here is exactly how it went, including the setup, the permissions, the part that worked, and the part where the sandbox stopped and asked permission.

Record & Replay is not on by default

First thing worth knowing: when you type "record" into the Codex app, nothing happens. The plugin is not enabled out of the box.

You have to go to Plugins, search "record," and add it. Once it is in, you invoke it with @record and replay. From there you describe what you want, for example: "please record this workflow and create a skill from it. The purpose is to create a draft of a YouTube video once we are ready to post it."

Grant the OS permissions first

Before Codex can observe or act on your machine, it needs explicit operating system permissions. On macOS that means going into Privacy & Security and enabling two things for Codex:

1. Screen recording, so it can see your actions.
2. Computer use, so it can act on its own later.

If those are not granted, recording and replay simply will not work. Turn them both on up front.

Recording the actual workflow

I had not finished this week's videos, so I used last week's instead. With the recording running, I opened Arc browser and walked through the whole YouTube Studio process out loud:

• Create, upload videos, select files, and grab the full final cut.
• Paste in the title and the description.
• Add the three thumbnails for A/B testing. I always label the final three with Fn_final_01, 02, and 03, so I selected those and set the test.
• Open "show more," paste in the optimized tags. I do not write these by hand; I have AI pre-generate them.
• Mark it as not made for kids.
• Turn monetization on.
• Manage ad slots manually, placing them at roughly 3:13, 4:43, and 7:41.
• Add an end screen with a top-right and bottom-right element, leaving the rest default.
• Save the whole thing as private.

The reason I narrated every step is intentional. As I said during the demo, I was talking out loud so the model captures the intent behind each click, not just the raw clicks themselves, and so you are not stuck watching me click in silence. Note: this is narration for context, not voice-to-text.

How Codex turns the recording into a skill

When you hit "done recording," Codex explains what it is doing with the data. The raw log is huge because browser accessibility trees include whole pages. Instead of keeping that mess, it summarizes the events into a compact sequence of clicks, selections, typed text, windows, and URLs.

In its own words, the skill captures intent without leaking long private content. That is a deliberate design choice: it avoids dragging your full private page contents into the skill.

In this run it judged the workflow clear enough to build the skill without asking me follow-up questions, then validated its work by reading the skill back from disk. If it had needed more detail, it would have asked.

The headline number: it generated the full YouTube-draft skill in 2 minutes 9 seconds, capturing upload, metadata, AI disclosure, monetization and ad suitability, ad slots, the end screen, and saving as private with a no-unsaved-changes verification step.

Testing the replay on a brand-new video

A skill is only useful if it replays. So, treating this as the beginners we are, I just asked straight up how to test it, then pointed Codex at a different video in the same directory, the most recent one, a funnel video I had built toward an $85,000 goal.

This is where it got interesting. The replay is sandboxed and permission-gated. Rather than guessing the file from the upload dialog, the sandbox blocked reading the desktop video file and asked permission to inspect the directory first. That is the right behavior. You approve the prompt instead of letting it guess.

Two practical takeaways from the replay:

• Give it a clear target. Telling it to organize by date added and take the most recent file helps it pick correctly.
• Expect permission prompts and approve them. The gating is a feature, not a bug.

Computer use is flagged as elevated risk

Codex itself labels computer use as elevated risk. This is early and experimental, so use it deliberately and at your own risk. I demo things like this on the channel so you do not have to be the first one to find the sharp edges. Selfishly I am doing it for me too, but mainly I am doing it for you.

Recording is also a discovery tool

Something I did not expect: the act of recording one workflow made me immediately spot more tasks worth automating.

The big one is timestamps. My current ritual is to place green markers on my DaVinci Resolve timeline after the intro and at each topic break, manually note the exact minute and second, then type them into Clearmud OS, the tool I built for myself, when I create a new script. It is the part I would love, love, love to stop doing by hand.

My hunch on the real fix: generate captions and let the captions dictate the timestamps, rather than transcribing marker positions manually. That is a future automation, but I would never have framed it that clearly if recording this one had not put the tedium right in front of me.

The bottom line

Record & Replay lets you teach Codex by demonstration. Add the plugin, grant screen recording and computer use, record a complete real run while narrating your intent, and let the model compress it into a skill. Replay is sandboxed and will ask before it reaches for files. The click-heavy parts of a creator's workflow, metadata, ad slots, end screens, are now automatable without writing a line of code.