Record a Skill: Who Will Inherit the Baby Boomers' Knowledge?
Anthropic has introduced a new feature for Claude Cowork: “Record a Skill.” The idea behind it is very simple. Instead of describing a step in a prompt, you record your screen, perform the task once, and comment aloud on what you’re paying attention to and why. Claude analyzes clicks, keystrokes, and voice commentary to build a reusable skill that it can later execute on its own. The feature is currently available to Pro, Max, and Team subscribers in the Cowork environment of the Claude desktop app for Mac. (Source: BornCity, Search Engine Journal)
The concept isn’t entirely new: OpenAI had already introduced something very similar for Codex on June 18 with “Record and Replay,” though that feature wasn’t available in the European Economic Area, Switzerland, or the United Kingdom. One day after Anthropic’s announcement, Loom followed suit and introduced video prompts that convert screen recordings into structured action plans for agents like Claude or Cursor. (Source: BornCity)
Three providers, one week, one direction: AI should no longer learn solely from text, but from real, observed practice. That’s exactly what makes this news more interesting than it seems at first glance.
Why this matters now, of all times
At the same time, a recent forecast by the German Economic Institute (IW) shows just how acute the problem in Germany is already becoming: By 2036, the economy will face a shortage of approximately 4.3 million workers. The reason: Significantly more people are retiring than young people are entering the workforce—a loss of about half a million workers per year. “Germany is not facing demographic change; it is already in the thick of it,” says IW labor market expert Holger Schäfer. (Source: IW Cologne, Tagesschau)
.png)
At first glance, this is a capacity issue. But the real cost lies elsewhere: as the baby boomers retire, knowledge that was never properly documented is disappearing. Not what’s on the organizational chart, but what matters in day-to-day operations—typical special cases involving customers, lessons learned from handling complaints, risks that you “just see,” and priorities that were never formally established. When this kind of knowledge leaves the company along with an employee, it’s not just manpower that’s lost. It’s a sense of direction.
Why Traditional Approaches Fail
Companies have been trying to get a handle on this for years through interviews and documentation—with limited success. Interviews are only of limited help because people are rarely aware of their own intuition. You can ask someone how they make a decision, but you often only get half the answer because the rest takes place in the unconscious mind. Documentation, on the other hand, usually captures only the obvious, not the actual thought process behind it.
A second approach that is currently being pursued with great vigor is skill inference: an AI analyzes the digital footprints someone leaves behind to deduce how decisions are made and quality standards are set. A third approach—which we believe is currently the most promising—is emerging more incidentally: Anyone who already works with AI agents automatically passes on their implicit knowledge every time they provide feedback. For example, if an agent is repeatedly corrected on phrasing or formatting, it learns not only the explicit guidelines but, over time, also the underlying quality standards.
The Hayek Problem
This is precisely where a concept comes into play that is currently resurfacing in the debate about AI and organizations: the Hayek problem. It describes how knowledge in organizations is largely implicit and localized—tied to individual people and contexts, and, in many cases, inaccessible to anyone else. Many researchers argue that no matter how intelligent AI models become, they cannot overcome this problem because they need the full context to function effectively.
Wharton researcher Ethan Mollick recently summed it up in a public exchange with a researcher from OpenAI. He believes it’s a valid argument that “curing cancer will be easier than replacing Accenture”—in other words, curing cancer will be easier than replacing Accenture. (Source: Ethan Mollick on X, via Blockchain.News)
His point: Cancer is a well-defined problem with well-defined data. Organizations, with their distributed, informal knowledge, are not.
Against this backdrop, “Record a Skill” is more than just a nice feature. Until now, tacit knowledge first had to be described and abstracted before an AI could work with it—precisely the hurdle that interviews and documentation fail to overcome. An annotated screen recording gets around this, at least in part: It conveys not only the sequence of clicks, but also, through the spoken explanation, some of the logic behind it.
The Downside: Knowledge That Ages Along With AI
Before this leads to too much euphoria, it’s worth taking a look at a study published in February in the *Academy of Management Review*. In it, Jin Gerlach of the University of Passau and Don Lange of Arizona State University describe a self-reinforcing cycle they call “Fading Memories,” which unfolds in three steps: First, AI takes over a task, such as a quality check. Second, employees use the necessary expertise less frequently as a result, lose it, or leave the company—while new employees don’t even build it up in the first place because the AI is already handling the task. Third, the AI models themselves become outdated because they were trained on historical data and would need to be updated regularly—a process for which, ironically, the very human expertise that has been lost due to the use of AI is lacking. Gerlach warns that lost human expertise could impair the quality of AI models over time—in the worst case, “insidiously and unnoticed.” (Source: University of Passau)
Once a skill has been acquired, it is not a finished product but must be actively maintained—otherwise, there is a risk of the very gradual loss of knowledge that one was actually trying to prevent.
What Companies Should Be Doing Right Now
If you take this issue seriously, you don’t have to launch a major project right away. It makes more sense to start small and focused: Look for processes that are heavily tied to specific individuals, where new employees take a long time to feel confident, and that immediately become unstable in the event of vacation, illness, or resignation. That’s exactly where knowledge-critical work usually lies—and that’s exactly where it’s worth demonstrating and recording this work rather than just describing it.
There are two things that should be considered from the outset, especially in light of the “Fading Memories” warning: first, a regular review cycle to verify whether a recorded skill still aligns with current practice. Second, a deliberate minimum of human practice, so that there is still someone in the company who can recognize incorrect or outdated AI behavior.
Conclusion
The baby boomers are retiring. This is well known and well documented. What receives less attention is what is being lost along with them: not just the workforce, but a wealth of practical knowledge that has never been fully documented.
“Record a Skill” doesn’t solve this problem overnight. But it’s a sign of where agent-based systems are headed—away from isolated prompts and toward systems that can absorb practical knowledge directly from real-world experience. The more exciting question for companies is therefore no longer just: What tasks can AI take on? But rather: What knowledge do we need to pass on before it disappears—and how do we keep it alive instead of just freezing it in time?
Those who want not only to understand this topic but also to implement it in their own companies will find two suitable starting points at Leaders of AI: The AI Integration Expert program focuses specifically on how to systematically integrate AI agents into real-world work processes—including tool selection, governance, and performance measurement. Those who prefer to focus on the strategic level—such as how to build an entire team of AI agents and how to ensure knowledge transfer across the entire company—will find what they’re looking for in the Master Business with AI MBAI®) university certificate program .
Sources
- BornCity: “Record a Skill” – Claude Learns Work Steps by Watching
- BornCity: Claude “Record a Skill” – Screen Recordings Become Automation
- Search Engine Journal: Anthropic's Claude Can Now Watch a Video and Learn Your Job
- IW Cologne: Baby Boomers Retiring – 4.3 Million Workers Will Be Missing by 2036
- Tagesschau: IW Study – Baby Boomers’ Retirement Leaves a Major Gap in the Labor Market
- University of Passau: Study — How Artificial Intelligence Can Lead to a Loss of Knowledge in Companies
- Ethan Mollick on X, as reported by Blockchain.News: OpenAI Debate Highlights 2035 White-Collar Risk
