這些不是盲目歌頌 AI 的「好用法」，而是經過 evidence screen 的 agent operating plays：每個 play 都要有公開研究、企業領袖方向或安全治理依據，否則不放入清單。 These are not blindly optimistic AI tricks. They are evidence-screened agent operating plays: each needs public research, leadership direction or governance support before it stays on the list.
判斷一個 play 是否聰明，不看它聽起來多有未來感，而看它是否有 evidence、清楚輸入、負責人、SSOT、審核點與停止方法。A play is smart only if it has evidence, 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.
一個 play 要留在清單上，必須有清楚 evidence basis、SSOT、負責人、權限、審核點同停止方法。聽起來聰明但無證據、無 owner、無控制，就不應該進入營運系統。A play stays on the list only when it has an evidence basis, SSOT, ownership, permissions, review points and stop conditions. If it sounds clever but lacks proof, ownership or control, it should not enter the operating system.
把 Smart Plays 升級成 Build 安全控制Upgrade these plays into Build safety controls→