Most automation tools ask you to learn something new before they save you any time. Codex’s latest feature flips that completely — you just do the task once while talking through it, and Codex learns the workflow from watching you.
It’s called record and replay (OpenAI calls the output a “skill”), and after testing it against similar features from competitors, it’s the closest any AI tool has gotten to actually being useful for this kind of thing.
What the Feature Actually Does
You open the Codex desktop app, hit record, and you’ve got up to 30 minutes to perform any workflow on your computer. Codex captures your screen activity and your microphone. As you work, you narrate each step — the same way you’d walk a new hire through a process.
Say you’re a freelance designer who invoices clients every two weeks. You open your time-tracking spreadsheet, copy hours into an invoice template, rename the file with the client name and date, then drop it into a client-specific folder before emailing a PDF. That entire sequence, narrated step by step, becomes a reusable skill. Next billing cycle: type a slash command, select the skill, let Codex run it.
The narration matters more than it might seem. Codex uses your spoken context to understand intent, not just mechanics. If you say “the order of these files doesn’t matter, just grab the most recent one,” that flexibility gets baked into the skill. A silent screen recording wouldn’t capture that.
How It Compares to Similar Tools
Record-and-replay AI isn’t a new concept. Anthropic’s Claude has offered a “Teach Claude” button in its Chrome extension for a while — same basic idea, microphone on, demonstrate the task, let it repeat. The problem is reliability. Claude’s version works often enough to feel promising but not often enough to trust with anything you’d rather not babysit.
Codex’s implementation clears that bar more consistently. It’s not perfect — more on that in a moment — but it crosses the threshold where you’d actually build a habit around it rather than test it once and forget it exists.
The catch: right now it’s Mac-only and requires a paid plan. That limits who can use it today, though those restrictions will likely loosen as the feature matures.
Where It Still Falls Short
Before you start dreaming about a four-hour workweek powered by recorded skills, here’s the honest version.
First runs need hand-holding. The first time you execute a newly recorded skill, Codex will almost certainly pause to ask for permissions — browser access, folder access, that kind of thing. Expect to run it two or three times before it moves through cleanly.
It doesn’t self-correct like a human would. If something unexpected happens mid-task — a file has a slightly different name than expected, or a webpage loads slower than usual — a person adapts without thinking. Codex can get stuck, or worse, take a wrong turn confidently. In testing, it once moved an unrelated file into the wrong folder and then politely asked for permission to undo its own mistake.
Complexity has a ceiling. A three-step workflow involving a folder, a browser, and a file move runs well. A ten-step workflow with branching conditions? The failure rate climbs. For now, the sweet spot is linear, repeatable tasks with predictable inputs.
The Right Mental Model for Using This
The mistake is thinking of Codex skills as a replacement for doing the task. The better frame: you’re shifting from worker to supervisor.
Instead of spending 15 minutes on a repetitive task, you spend 2 minutes checking whether Codex ran it correctly. That’s still a meaningful win, especially if you can queue up several skills in parallel while you focus elsewhere.
Think of tasks like:
- Pulling weekly data from a web dashboard into a local spreadsheet
- Batch-resizing and renaming image exports from a design tool
- Moving completed project files through a folder archive structure
- Filling in a recurring form with information from a reference document
None of these are glamorous. That’s the point. The dull, clicking-through-the-same-screens tasks are exactly where Codex skills earn back time without requiring you to write a single line of code or configure a single automation rule.
Getting Started
If you have Codex on Mac with a paid subscription, look for “Record Skill” under the create menu in the desktop app — it’s separate from the standard plugins list. Before you record, jot down the steps in plain language. Narrate generously as you go, especially at decision points. Run the skill immediately after recording to catch permission prompts early, then run it two or three more times on real data before you start relying on it.
The feature isn’t magic yet. But “more reliable than anything else that does this” is a meaningful place to be.
