让 grok bot 方案用久一点：1）能用事件触发就用...

让 grok bot 方案用久一点： 1）能用事件触发就用。slack、github、一个 reaction。真的发生事情 bot 才跑（cron 就算没变也会叫醒它） 2）指示 bot，一进展不了或卡住就通知你。这样你作为人可以再把它拉回路上 3）能用连接器就别用浏览器！在网站上点来点去是一堆截屏和步骤。连接器一次调用给你同样数据，而且舒服／快／顺很多 4）技能保持紧，但把真正会带跑的细节放进去。太含糊 bot 可能乱走、重试，比稍长一点更贵。太具体你又得确保它真的正确 5）不要在 bot 聊天里写代码！对 repo 踢一个云端 agent，或叫它用 cursor/grok build cli 做 6）一次性监看用完就删（grok bot 会提醒你这件事）

原文：how to make your grok bot plan last longer: 1) use an event trigger when you can. slack, github, a reaction. the bot only runs when something actually happened (a cron wakes it even if nothing changed) 2) instruct the bot to notify you as soon as it can’t make further progress or is stalled. this way you as human can help putting it on the right course again 3) connectors > browser when possible! clicking around a site is a pile of screenshots and steps. a connector gives you the same data in one call, and it's so much nicer/faster/smoother 4) keep skills tight, but put in the details that actually steer the run. too vague and the bot might wander, retry, and costs more than a slightly longer one. too specific and you have to make sure its really correct 5) don't do the coding in the bot chat! kick a cloud agent at the repo, or ask it to use cursor/grok build cli to do it 6) delete one-shot watches when they're not needed anymore (grok bot will remind you about this)

集成