Claude Cowork is moving skill creation from written instructions toward observed work. The reported flow records a screen walkthrough while the user narrates the task, then turns that demonstration into a skill Claude can reuse later.
That is a different product boundary from “save this prompt.” The reusable unit is no longer only a paragraph of process advice. It can include the demonstrated sequence, the intent behind the sequence, and the operational context needed to repeat it.
What changed?
The important signal is not that Claude can watch a video. The signal is that a mainstream agent surface is treating human demonstration as a capability authoring path.
Recorded skills make the agent memory loop more concrete: watch work, extract procedure, package the procedure, and make it callable in later sessions. If the generated skill remains inspectable and editable, the result starts to look like lightweight process automation rather than a chat transcript.
New Runtime Read
Teams should treat recorded skills as draft automation, not as finished infrastructure.
The high-value loop is record, inspect, test, version, and then reuse. Without review, a recorded workflow can preserve accidental clicks, fragile UI assumptions, hidden credentials, or outdated business rules. With review, it becomes a fast way to turn tacit work into a portable agent capability.