Grok Bot 录一次任务即生成共享技能

rompasso @0xrompasso原推发布于 2026.10.18

在 X 上打开原推

Record yourself doing a task once, for up to ten minutes, and Grok Bot turns it into a skill every agent on your account can run. xAI’s own docs warn that one example rarely shows the skill where it needs to stop. This isn’t a dry-run sandbox for testing what you taught it. Teach a task watches one live walkthrough, drafts a skill from it, and files it where every agent on the roster can reach for it. What the documentation itself flags about that single demonstration: → A ten-minute cap on the recording, so whatever edge case never came up in that window never makes it into the skill → The draft needs manual work afterward, decision rules, failure handling and approval points added by hand, none of it inferred from the one run → Test runs execute for real, there’s no simulated pass, so checking the skill means letting it act again before you’ve finished reviewing it → Every saved skill lands in one private library shared across your whole roster, not scoped to the agent that learned it The docs’ own advice after all that: confirm the output format, check that an audit trail exists, and verify the agent actually stops where you expected, three things a single clean demo never proves on its own. That’s the gap between demonstrating and delegating. Showing an agent how to do something once was never the same as teaching it where not to. What xAI’s own checklist for testing a taught skill actually covers is in the writeup below 👇

这条是引用着另一条写的：x.com/0xwast3/status/2110312109783269475