Codex Can Now Learn a Task by Watching You Do It Once

There is a particular kind of work chore that resists every productivity tool: the multi-step errand that lives across several apps. Filing an expense report. Submitting a time-off request. Pulling the same five numbers into the same weekly sheet. You know it cold, you do it on autopilot, and you have never once managed to automate it, because describing it to a computer was harder than just doing it again.

On June 18, OpenAI shipped a Codex feature aimed straight at that chore. It is called Record and Replay, and the pitch is almost suspiciously simple: show the assistant the task one time, and it figures out the rest.

Show it once, reuse it forever

Here is how it works. You turn on recording, then complete a repetitive task on your Mac the way you normally would. Codex watches the whole thing: the steps you take, the windows you move between, the buttons you click. When you stop recording, it takes that demonstration and writes itself a reusable "skill," which is just a plain instruction file describing what you did.

OpenAI's own examples are deliberately unglamorous. "Record and Replay lets you show Codex a recurring task, like filing an expense report or submitting a time-off request," the company wrote when it announced the feature. "Codex turns that demo into an inspectable, editable skill. You control when recording starts and stops." Greg Brockman, OpenAI's president, put it more bluntly: "You can now teach Codex by demonstration."

The file stays yours to edit

The detail that matters most is that the result is not a black box. The skill Codex generates is a readable file you can open, inspect, and change. If it misunderstood a step, you fix the line instead of re-recording the whole thing. If the task shifts, you edit the recipe rather than starting over.

It is also built to flex. The saved skill accepts inputs that change from run to run, so a single expense-report recording works again next month with a new date and a different receipt. That is the difference between a brittle macro that breaks the moment anything moves, and something closer to a coworker who has done the task before and just needs to know this month's numbers.

For anyone who has bounced off automation tools because the setup cost more time than the task itself, that is the whole game. The setup always cost more than the chore, and that is the part this removes.

A limited launch, and a skeptical crowd

The catch is that almost nobody can use this yet. Record and Replay runs only on macOS, it requires OpenAI's "Computer Use" capability to be switched on, and at launch it is unavailable across the European Economic Area, the United Kingdom, and Switzerland. There are no public details yet on pricing or a wider rollout.

Reaction has been split roughly down the middle. Plenty of early users were giddy: "Just show Codex what you want to do once and let it take care of it forever," wrote one, summing up the appeal. Others shrugged, or pointed out that the idea is not new. Matt Shumer, who builds AI agents, noted that his team demonstrated a version of teaching a model by recording a workflow "years ago," adding, "just a lil early." A louder camp simply wanted OpenAI to ship a smarter model instead of another convenience feature.

There is a privacy angle worth flagging too. Teaching an AI your workflow means handing it a recording of exactly how you do your job, which is a different kind of data than a typed prompt. That trade is worth being clear-eyed about before you record your way through a sensitive task.

Why it matters to you

This is the clearest step yet toward genuinely automating the repeating parts of your work. The automation you can set up today, the kind you build in a tool like Zapier or n8n, is rigid. Showing an assistant the task once and letting it write its own instructions is a different thing entirely.

I'm excited to put this one to the test. Who knows how well it really handles breaking a repeatable task into steps, but the chance to automate some of the boring parts of your job out of the way is well worth a shot. The shape of it is the thing to notice: handing a task to a computer is starting to look like just doing it once while it watches.

Sources

← Back to AI for Mortals