Why You Need One
把重複判斷、整理、草擬變成自己的工作槓桿。Turn repeated judgement, organising and drafting into personal leverage.
Build It Live
現場建立 agent：角色、知識、技能、工具、批准。Build the agent live: role, knowledge, skill, tools and approval.
Secure By Default
先定安全邊界，再加能力；避免一開始就過度授權。Set safety boundaries before adding capability; avoid granting too much authority from day one.
Live Build Path
帶走一個安全 personal agent 的 9 個步驟Nine steps to leave with a safe personal agent
寫清楚它幫誰、做什麼、為何值得有一個 agent。State who it helps, what it does and why an agent is worth having.
設定名稱、任務、輸入、拒絕規則、升級規則與成功指標。Configure name, mission, inputs, refusal rules, escalation and success metric.
只放入已批准知識；清楚標明不可使用的敏感資料。Use approved knowledge only; name the sensitive data it must not use.
把一個重複流程寫成 reusable skill，而不是每次重新 prompt。Turn one repeated workflow into a reusable skill, not a one-off prompt.
決定它可讀、可草擬、不可直接執行的工具範圍。Decide what it may read, draft and never execute directly.
任何發送、分享、改資料、承諾或花錢前都要先批准。Require approval before sending, sharing, changing data, committing or spending.
用一個故意刁鑽情境測試它會拒絕、追問或升級。Use one adversarial scenario to test refusal, clarification or escalation.
完成 My Safe Agent Card：用途、邊界、工具、批准、成功指標。Finish the My Safe Agent Card: purpose, boundaries, tools, approval and success metric.
Takeaway Artifact
離場時，你應該有一個 Agent，以及一張安全設定卡。You should leave with an agent and a security setup card.
My Safe Agent Card
用途、不可做事項、知識來源、工具範圍、批准點、成功指標。Purpose, must-not-do rules, knowledge sources, tool scope, approval points and success metric.
Starter Prompts用提示庫加速第一個 skill，但要按你的資料與風險修改。Use the prompt library to speed up the first skill, then adapt it to your data and risk.
Safe Next Step先試低風險 loop，再把涉及客戶、金錢、權限的用法升級到 Build。Pilot a low-risk loop first, then move customer, money and permission workflows to Build controls.
Product Lead Lens
今日做 Agent，不是先追求自治，而是先追求可觀察、可評估、可限制。Building agents today is not about autonomy first; it is about observability, evaluation and constraints first.
Context Pack
先定義它可用的知識、語氣、範例與禁區，再談工具。Define knowledge, tone, examples and exclusions before adding tools.
Mini Evals
準備 3 個成功例、2 個拒絕例、1 個攻擊例，先測再擴權。Prepare three success cases, two refusal cases and one attack case before granting more authority.
Trace Review
看得見它用了什麼資料、叫了什麼工具、為何需要批准。See what data it used, which tool it requested and why approval was needed.