This summer, OpenAI and Anthropic both announced features that let their AI systems learn a task by watching someone do it once, instead of being told how in a prompt. In June, OpenAI launched Record & Replay, a feature that lets ChatGPT and Codex users demonstrate a workflow and turn it into a reusable skill. Weeks later, Anthropic unveiled Record a Skill inside Claude Cowork: Record your screen doing a task, narrate your reasoning, and Claude turns it into a skill it can run again. Two competitors, converging on the same solution within weeks of each other. To me, that’s an admission that prompting alone was never going to get AI where it needs to go.
I’ve been thinking about and working toward this milestone since my time at Apple, where I worked for 12 years, spending my formative years building the Chinese version of Siri. As a founding engineer, I was excited to watch Siri break new ground as a voice-activated assistant, which let people use natural language instead of tapping through menus. The next challenge was context awareness: building a system that could complete the first request correctly, as well as carry that understanding into the ensuing command. The gap was never linguistic. It was that the system needed to retain the memory of someone’s preferences or habits across interactions.
For example, tell a voice assistant to set an alarm for 6 every day, and it’s unclear whether the time is for morning or night. Most people don’t think of that as ambiguous, because they know their own schedule, and they expect whoever’s listening to know it too. That disconnect between what people say and what they actually mean is the same one that OpenAI and Anthropic are now aiming to close.
Show, don’t tell
How someone works is more idiosyncratic than what software assumes. Even if two people have the same job and responsibilities, they’ll use different tools, follow distinct sequences, and rely on a lot of unspoken context. Researchers call this tacit knowledge: what people know but can’t quite articulate, a term coined back in 1966 by the philosopher Michael Polanyi. One study estimated that 40% of a company’s valuable knowledge is inside individual employees’ heads, never written down anywhere.
We often experience this limitation when using AI prompts. If you ask someone to describe how they complete an expense report, they’ll say that they upload a receipt, categorize it, and then submit it. But what they’ll likely omit is that they ask their manager to review meals over $75, or that they classify their client dinners differently from team lunches. That’s the challenge that prompt-based AI can’t address, because it can’t correct its way to context that hasn’t been stated.
In contrast, when you demonstrate the same task to an AI system, it captures the sequence, the decision points, and the small judgment calls that are intrinsic to how someone gets the job done. A demonstration observes all the nuances of a workflow, because the context and the action go hand in hand from the very beginning.
Build a demo library
Record & Replay and Record a Skill are real progress in understanding and mimicking how we work. Paired with a scheduled task, a recorded skill doesn’t need someone to reopen it, because it can run autonomously and in the background while the user works.
As more people use demonstrations to automate their workflows, the next obstacles are ensuring that people recognize what they can hand off, and making this process more intuitive. A study from the AI company Glean found that employees spend roughly 6.4 hours a week “botsitting”: babysitting AI tools, correcting their output, and reexplaining things the tool should already know. Imagine how much of that wasted time could be recouped once people get into the habit of recording what they do, so a system can learn the workflow once and then take over.
There’s also an emerging opportunity for workflow mining, allowing technology to observe how work is done and extract reusable knowledge from it. Workflow data will become invaluable, not only because it automates multistep tasks from end to end, but also because it can create a library of workflows that people can select and personalize. Someone can build a workflow for closing out expense reports, while someone else adapts it for their own approval chain. As a result, the library gets sharper with every version added, the same way open-source code improves as more developers build upon it.
The real test
OpenAI and Anthropic’s launching demonstration features within weeks of each other is a clear signal that the biggest gains will be in systems that learn how people actually work. Most of us still default to prompting because we’ve become accustomed to it, but one demonstration will save the back-and-forth prompts and iterations every time.
We’ve spent years teaching AI to understand what we say, and we can now treat a demonstration the way we’d treat training a new hire: something you do once, carefully, so you don’t have to explain it again. Judge these tools less by what they can already do, and more by whether you’ve shown them how you work.
