This summer, two of the biggest AI companies shipped almost the same product within weeks of each other. OpenAI added a feature called Record & Replay to its coding assistant Codex on June 18. You do a task once on your Mac, like filing an expense report, and Codex watches, then turns what it saw into a set of steps it can run again on its own. Weeks later, on July 21, Anthropic launched Record a Skill inside Claude Cowork, which does nearly the same thing but also lets you talk out loud while you work, so Claude picks up your reasoning, not just your clicks.
Both companies are admitting something without saying it directly: typing instructions to an AI only gets you so far. People are bad at describing their own jobs. Ask someone how they process an invoice and they will skip the part where they always double check anything from a new vendor, because that habit is so automatic they do not think to mention it. Researchers call this tacit knowledge, meaning knowledge you have but cannot fully put into words. Recording a task, rather than describing it, captures that hidden judgment along with the visible steps.
This is not a new idea. Computer scientists have tried to build software that learns by watching since the 1980s, under the name programming by demonstration. It kept failing because turning one example into something that also works on slightly different situations was too hard for older software to handle. Today's AI models can reason about what you were trying to do, not just copy your clicks, which is why this old idea finally works well enough to ship.
Business operators should recognize this shape, because it looks like software many of them already own. Robotic process automation, sold by companies like UiPath and Automation Anywhere, has spent two decades automating this kind of back office paperwork: claims intake, invoice matching, approval routing. The catch is that setting it up usually needs an IT project, a consultant, and weeks of configuration. Record & Replay and Record a Skill let any employee build a similar automation just by doing their job once in front of the computer. That is a real problem for the traditional automation sales pitch, especially for smaller companies that could never justify a full automation project but already pay for ChatGPT or Claude.
There is also a real cost this could fix. A recent Glean's Work AI Institute report found office workers spend more than six hours a week correcting AI tools and re-explaining context those tools should already know, a habit researchers call botsitting. If a workflow only needs to be shown once instead of explained every time, a good chunk of that wasted time goes away.
None of this is fully ready yet. Both features currently only work on Mac computers, need a paid subscription, and are not available yet in the European Union, the UK, or Switzerland. There is also a privacy question underneath it: a screen recording captures everything on your screen, including client names, salaries, or medical details, not just the steps of the task. Any company that adopts this needs a clear policy on what can be recorded before employees start teaching their AI by demonstration.
The bigger signal is where this is heading. If showing an AI your job becomes as normal as training a new hire, the companies that write this down first, task by task, build a real advantage. Everyone else keeps re-explaining the same thing to a machine that should have learned it months ago.