這些不是盲目歌頌 AI 的「好用法」，而是經過 reality check 的 agent operating plays：每個 play 都標明何時聰明、何時會變蠢，以及必須加上的護欄。 These are not blindly optimistic AI tricks. They are reality-checked agent operating plays: each names when it is smart, when it becomes dumb, and which guardrail keeps it useful.
判斷一個 play 是否聰明，不看它聽起來多有未來感，而看它有沒有清楚輸入、負責人、SSOT、審核點與停止方法。A play is smart only if it has clear inputs, ownership, SSOT, review points and a way to stop.
最常見失敗不是 AI 不夠聰明，而是 loop 太模糊、權限太大、資料來源太舊，或把建議當成自動決策。The usual failure is not model stupidity. It is vague loops, broad permissions, stale sources or treating advice as automatic decision.
Spark 可以輕巧；但任何會改資料、通知人、花錢或影響客戶的 play，都要升級到 Build 的控制標準。Spark can stay lightweight; any play that changes data, notifies people, spends money or affects customers graduates to Build controls.
每個聰明用法都可以變成一個 agent loop；但只有加上 SSOT、負責人、權限、審核點同停止方法，先會由「聽起來聰明」變成真正可營運的 operating system。Each smart play can become an agent loop; only with SSOT, ownership, permissions, review points and stop conditions does it become an operating system worth running.
把 Smart Plays 升級成 Build 安全控制Upgrade these plays into Build safety controls→