同時開五個 AI 對話之後，你就失去對它們的掌握了——哪個還在跑、哪個十分鐘前就卡住、哪個早就結束但忘記關。 這支開一個本機儀錶板把 Claude Code 和 Codex 的對話並列顯示。在線狀態用實際存活的 process 驗證，不是只讀設定檔；超過十分鐘沒活動的標橘色。 還有一顆服務狀態燈，分辨「我的對話卡住」和「服務整個掛了」。
Prevents AI agents from declaring tasks complete until every requirement is verified with a one-line evidence anchor — when a gate fails, the agent stops and reports rather than proceeding, and only a human can unlock the gate to continue.
AI Agent Skill：根據你主動提供的作息、當下情境與修正，提出可解釋的下一步假設。跨對話分層記憶與提醒都需要相容的宿主工具、欄位揭露及你的確認；沒有工具時只使用目前對話或提供未排程計畫。
我長期處理中英夾雜的技術型影音內容（教學影片、會議與簡報錄音），很清楚這類素材最頭痛的就是專有名詞和英文縮寫被聽錯。這個 Skill 把我實際在用的整套做法工具化：以 faster-whisper 在本機 CPU 就能轉錄（免 GPU、免雲端 API、免費），並用「轉錄前餵專有名詞提示＋轉錄後批次修正同音錯字」兩道工序把準確率拉高；YouTube 連結則走「先抓現成字幕、失敗自動改抓音訊轉錄」的穩定路線，最後一併產出 SRT 字幕、逐字稿與帶時間戳的重點摘要。全程本機運算、檔案不上雲，適合在意內容隱私又要快速消化長影音的人。 I work daily with mixed Chinese-English technical audio, where jargon and acronyms get mis-heard most. This Skill packages my workflow: faster-whisper runs locally on CPU (no GPU, no cloud API)
用 AI 做事最容易蒸發的不是成果，是決策理由。檔案還在，但「為什麼這樣設計、為什麼放棄另一個方案」隨著對話關閉一起消失，三週後你會完全想不起來。這支把一場對話收斂成四類寫進每日紀錄，封存前再逐檔回查驗證真的寫進磁碟——AI 說「已寫入」跟磁碟上真的有那個檔是兩件事，更陰險的是索引寫了但檔案本體從未落盤。
Forces AI agents to prove a task is actually done before declaring it complete. 4-phase gate: instruction reconciliation, evidence anchoring (values only readable from the actual output), self-scoring (A/B/C/D), and reverse witness check. Eliminates "done" declarations that aren't.
Stop the assistant from refusing work it could actually have done. "I can't" is an expensive sentence - if it is wrong, people stop asking, including for the things you could have done. This is the four-step check that catches a wrong no before it reaches the person waiting: has this already been solved here, is the wall real or just the first path, is this genuinely theirs to do, and have you actually asked the tool instead of guessing at its limits.
