Teaching the same task every time is waste I want out of the workflow. Codex Record & Replay, announced for macOS on June 18, 2026, proposed generating a reusable skill from a demonstration. I like starting by showing the work. The decisive part comes when the input changes and the original sequence no longer answers every question.
In August 2026, I requested delegated research into turning a capability into a skill. I wanted to reuse what was learned without starting the explanation from scratch on the next request. That later request explains my interest in the June announcement. A demonstration can supply material for an instruction I can maintain and call again.
Copying the exact sequence of one execution is limited, though. In a hypothetical example, I demonstrate a computer activity and then supply another equivalent input. The skill needs to preserve the objective while recognizing what changed. If a detail of the first example becomes a permanent requirement, I have automated a coincidence. The next task will expose the difference.
I want to read the generated instruction and understand its assumptions. When should it be called? What result should it produce? Which situations require a fresh decision? Those answers help me review the capability without watching a long recording again whenever something turns out differently. Reuse also means being able to maintain the thing afterward.
Demonstration has an advantage that interests me: it reveals decisions that are easy to show and tedious to describe in advance. During execution, I might discover that a detail changes the result. That detail belongs in the skill. The route chosen merely because a particular window happened to be open may not deserve to become a mandatory procedure.
My first repetition will use a deliberate difference in the input. Something small, sufficient to check whether the agent understood the work beyond the example. I want the results side by side so I can correct the instruction where generalization fails. Reproducing an identical demonstration tells me less about the next task I actually intend to delegate.
The announcement offers a way to teach; my August request defines what I want from the teaching. I will treat the generated skill as material to review and require a repetition with different input before incorporating it into a routine. If I still have to explain the entire task for it to work again, teaching remains unfinished. That extra work needs to be visible before I rely on repetition.