Grok Bot’s skills guide describes teaching a browser task by demonstration where the control is available. A user performs the workflow once, reviews the generated skill and tests a safe example before scheduling. The recording covers visible computer interaction for up to ten minutes, without microphone audio. The guide calls the learned skill a draft and warns that availability may be gradual. Its October 12 revision does not establish a launch date; this is analysis of documented behavior.
The important gap is what one successful example leaves unsaid. The guide asks users to add decision rules, failure handling and approval boundaries. It also warns that a routine’s test run performs real work, potentially changing files or using connected tools. Reviewing a generated method and running it are therefore separate steps, with different consequences.
Editorial take: Before scheduling a learned task, write down one missing-input case, one ambiguous-result case and the point where it must stop for approval. Inspect the draft against those cases, then choose harmless test inputs. That turns a recording into an explicit operating procedure rather than assuming a single demonstration covered every exception. We have not tested the feature.