發布日期：2026-07-26
幾個 AI 正在替你討論方案，第一輪卻發現大家把你的單機需求當成要做叢集，也就是讓多台主機一起運作；現在不用整場喊停重來。開啟「輪次間指引」開關後，討論工作階段（一次連續進行的討論）每一輪結束、下一輪開始前都會暫停，讓你補一句「只有單一主機，不要叢集」。若一直空白，最多等五分鐘便自動略過，討論照原本方式繼續，像主持會議時每回合按一下暫停鈴，及早把話題拉回來。主持人摘要下方的共識統計（列出每位參與者採取的立場）也會直接寫出名字；反對與替代意見用顏色標出，不必再翻紀錄猜是誰說的。第一次挑辯論語氣時，五種選項下方都有白話提示；例如「魔鬼代言人」會刻意站到相反邊，測試看似合理的結論是否站得住腳，因此不會再把它和「對抗式」混為一談。
Picture an AI council discussing a plan, then learning after round one that your single-host setup is being mistaken for a cluster. You no longer need to scrap the whole discussion. Turn on “guidance between rounds” and the session pauses before every later round, so you can add a correction such as “single host, no cluster.” A blank pause skips itself after five minutes, then the next round continues normally—like a chairperson’s bell that gets the meeting back on topic early. The consensus tally beneath the moderator’s summary now names each participant beside their position and colors dissenting or alternative views, so you need not hunt through the transcript. New to the five debate tones? Each picker option now has a plain-language hint: devil’s advocate deliberately argues the other side to test whether a plausible conclusion actually holds up, rather than merely being adversarial.
**Steer the council mid-discussion**: an opt-in "guidance between rounds" toggle pauses the session before every round after the first, so you can react to what you just heard — "we're single-host, no cluster" — instead of watching a debate run off in the wrong direction or stopping and restarting from scratch. The pause waits up to 5 minutes and is skipped automatically if left blank; the round then proceeds exactly as it did before.
**See who disagreed, not just how many**: the consensus tally under the moderator's summary now lists each participant by name next to their stance, with dissenting and alternative positions colored so they stand out, instead of a bare "2 agree, 1 alternative" that meant scrolling back through the transcript to find out who.
**The five debate tones now explain themselves**: a one-line plain-language hint appears under the tone picker — devil's advocate, for instance, now reads "the AIs will deliberately argue the opposite side to test whether a plausible conclusion holds up" — so a first-time user isn't left guessing how "adversarial" differs from "devil's advocate."