Agent / 自動化

核心觀念

• Skill 即 SOP：AI 的「技能 (Skill)」本質上是一份給 AI 看的、極其詳細的標準作業流程 (SOP)。你把做事方法教給它一次，它就能無限次重複執行。
• 一次給足指令：與其分次對話、逐步下指令，不如一次性將包含所有步驟、條件與預期產出的複雜指令完整交付給 Agent，這樣最節省 Token 且效率最高。
• 人機協作新模式：人負責定義目標與最終決策（例如，確認刪除哪些重複檔案），AI Agent 負責所有繁瑣的執行過程。
• 遠端分身：透過 Dispatch 功能，手機可以成為遠端遙控器，隨時隨地命令家裡或辦公室的電腦執行複雜任務，實現真正的「數位分身」。

整合工作流

1. 任務識別：找出日常工作中重複性最高、最耗時的任務（例如：每週整理下載資料夾、分析報表並製圖、搜尋特定主題的新聞）。
2. 首次手動執行與紀錄：
   • 在 Claude Code 或 Codex 中，首次手動、逐步地引導 AI 完成這個任務。
   • 例如，引導它「分析這個資料夾」、「找出重複檔案」、「將圖片檔移到…」、「將文件檔依內容分類」、「最後生成一份整理報告」。
3. 封裝為 Skill：
   • 當整個流程被成功驗證後，對 AI Agent 下達指令：「把剛剛我們做的事情打包成一個技能，命名為『智慧檔案整理』」。
   • AI 會使用內建的「Skill Creator」，將剛才的對話與操作流程，轉換成一個 .md 格式的 Skill 檔案。
4. 本地自動化 (Cowork)：
   • 之後，當需要整理檔案時，只需對 AI 說「使用『智慧檔案整理』技能來整理這個資料夾」，AI 就會自動執行所有步驟。
5. 遠端自動化 (Dispatch)：
   • 在電腦端的 Claude 啟用 Dispatch 功能。
   • 在手機 App 上，可以直接傳訊息給電腦：「幫我用『智慧檔案整理』技能，整理一下桌面」。
   • 電腦端的 AI 就會開始工作，並將報告回傳到手機。
6. 背景自動化 (Computer Use)：
   • 對於需要上網搜尋、下載資料的任務，可指示 Agent：「請使用 Computer Use 功能，去網路上抓…資料，下載到…目錄」。
   • AI 會在背景執行，不干擾你當前的工作。

最佳金句

「Skill就是你交給Claude一套完整的做事方法，遇到相關任務，他就會自動套用這個方法，你不需要重複去教他。」 「AI時代真的是我們需要動手部分越來越少了。」 「你不要一次叫他做一件，你要一次把你要他做的事情一次把他送出這樣子最省 Token 的數量。」

教學切入建議

常見誤區

• 指令過於零碎：習慣像聊天一樣，一次只叫 AI 做一件事，導致 Token 浪費且 AI 容易忘記上下文。
• 權限問題：AI Agent 在執行檔案操作（如刪除、移動）或遠端控制時，被系統的安全性設定擋下，使用者不知如何授權。
• 忘記預先登入：使用 Dispatch 或 Computer Use 操控瀏覽器時，忘記在電腦上預先登入相關網站（如 Gmail、Google Drive），導致 AI 無法操作。
• 對 Skill 的誤解：以為 Skill 是什麼神奇的程式碼，實際上它就是一份寫得非常詳細、結構化的 Markdown 指示文件。

---en---

Core Concepts

• Skill as SOP: An AI’s “Skill” is essentially an extremely detailed Standard Operating Procedure (SOP) for the AI. Once you teach it a method, it can perform it repeatedly without further instruction.
• Deliver Complete Instructions at Once: Rather than engaging in fragmented conversations and issuing commands step-by-step, it’s more efficient and token-saving to deliver a single, complex instruction that includes all steps, conditions, and expected outcomes.
• New Human-Machine Collaboration Model: Humans define objectives and make final decisions (e.g., confirming which duplicate files to delete), while the AI Agent handles all tedious execution processes.
• Remote Digital Twin: Through the Dispatch function, a smartphone can act as a remote control, commanding a home or office computer to perform complex tasks anytime, anywhere, realizing a true “digital twin.”

Integrated Workflow

1. Task Identification: Identify the most repetitive and time-consuming tasks in daily work (e.g., weekly organization of download folders, report analysis and charting, searching for news on specific topics).
2. First Manual Execution and Recording:
   • In Claude Code or Codex, manually guide the AI through the task step-by-step for the first time.
   • For example, instruct it to “analyze this folder,” “find duplicate files,” “move image files to…”, “classify document files by content,” and “finally generate an organization report.”
3. Encapsulate as Skill:
   • Once the entire process is successfully validated, command the AI Agent: “Package what we just did into a skill, named ‘Smart File Organizer’.”
   • The AI will use the built-in “Skill Creator” to convert the conversation and operational flow into a .md formatted Skill file.
4. Local Automation (Cowork):
   • Subsequently, when files need organizing, simply tell the AI: “Use the ‘Smart File Organizer’ skill to organize this folder,” and the AI will automatically execute all steps.
5. Remote Automation (Dispatch):
   • Enable the Dispatch function in Claude on your computer.
   • On the mobile App, you can directly message the computer: “Use the ‘Smart File Organizer’ skill to organize the desktop.”
   • The AI on the computer will start working and send a report back to the phone.
6. Background Automation (Computer Use):
   • For tasks requiring online search and data download, instruct the Agent: “Please use the Computer Use function to retrieve… data from the internet, and download it to… directory.”
   • The AI will execute in the background without interrupting your current work.

Best Quotes

“A Skill is a complete method of doing things that you give to Claude; when it encounters related tasks, it will automatically apply this method, so you don’t need to teach it repeatedly.” “In the AI era, we truly need to do less and less manual work.” “Don’t ask it to do one thing at a time; you should send all your instructions at once, which saves the most tokens.”

Teaching Entry Points

Common Pitfalls

• Fragmented Instructions: A habit of conversing, asking AI to do one thing at a time, leading to token waste and AI losing context easily.
• Permission Issues: AI Agents are blocked by system security settings when performing file operations (e.g., deleting, moving) or remote control, and the user doesn’t know how to grant permissions.
• Forgetting to Pre-login: When using Dispatch or Computer Use to control a browser, forgetting to pre-login to relevant websites (e.g., Gmail, Google Drive) on the computer, resulting in the AI being unable to operate.
• Misunderstanding of Skills: Believing that Skills are some magical code, when in reality, they are very detailed, structured Markdown instruction documents.