Record a skill
Ben's Bites
OpenAI's Codex can now watch you do a task once and turn it into a reusable, editable skill. No more re-explaining the same workflow every time.
Based on reporting by Ben's Bites — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Codex just picked up a feature that quietly changes how people will use it day to day: Record & Replay. Show it a repetitive task once, filing an expense report, submitting a time-off request, whatever tedious loop eats your afternoon, and Codex converts that single demo into a skill. Not a black box either. You can open it up, inspect the steps, edit them, and run it again whenever the task repeats.
This is a small feature on paper but it points at a bigger shift happening across the agent tooling space right now. Cursor shipped something similar with its /automate command, where you describe an automation in plain language and Cursor sets up the triggers, tools and instructions itself. Claude Code, meanwhile, is leaning into a related idea with its new steering guide, which is essentially a manual for where to park instructions, skills, hooks and subagents so an agent behaves consistently across a project. The common thread: teaching an AI a task once, then trusting it to repeat that task correctly, is becoming the actual product, not a demo trick.
It matters because the annoying part of agentic AI has never been getting a model to do something clever once. It's getting it to do something boring reliably, over and over, without you babysitting the process each time. Record & Replay is OpenAI betting that most valuable automation isn't exotic reasoning, it's capturing mundane office workflows and making them replayable. Expense reports and PTO requests are not glamorous use cases. They're exactly the kind of thing that eats hours across a company every month.
The skeptical read is that Codex still needs a human to perform the task correctly the first time, and any edge case outside that demo could trip the skill up. Editable steps help, but editability also means someone has to notice when a skill breaks and go fix it. Still, turning a screen recording into an inspectable, versioned automation is a more honest approach than a lot of the "just describe it and AI figures it out" pitches floating around this space. You show your work, and so does the agent.
My take — AI-written commentary, not fact-checked reporting
I like this one because it's boring in the right way. Half of agent hype is about frontier reasoning benchmarks nobody's job actually depends on, while the real unlock is Codex watching you file an expense report once and never making you explain it again. Give me ten more features like this over another leaderboard number any day.
Read more about this at: Ben's Bites
