[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-anthropic-model-task-persistence-tuning-zh":3,"article-related-anthropic-model-task-persistence-tuning-zh":30,"series-ai-agent-02bfe363-1ca6-4c20-be08-34d4ad532b71":83},{"id":4,"slug":5,"title":6,"content":7,"summary":8,"source":9,"source_url":10,"author":11,"image_url":12,"cover_image":12,"category":13,"language":14,"translated_content":11,"related_article_id":15,"keywords":16,"key_takeaways":22,"views":26,"created_at":27,"published_at":28,"topic_cluster_id":29},"02bfe363-1ca6-4c20-be08-34d4ad532b71","anthropic-model-task-persistence-tuning-zh","Anthropic 任務耐力調校指南","\u003Cp>你可能已經看過\u003Ca href=\"\u002Fnews\u002Fclaude-j-space-not-a-black-box-zh\">模型\u003C\u002Fa>一路做到底，卻在缺少關鍵資訊時不肯停下來，最後把錯誤放大成一串無效操作。這份指南適合正在做 \u003Ca href=\"\u002Ftag\u002Fagent\">Agent\u003C\u002Fa>、自動化腳本或長流程檢索的開發者。\u003C\u002Fp>\u003Cp data-speakable=\"summary\">本文\u003Ca href=\"\u002Fnews\u002Fterrazero-zero-demo-self-play-driving-zh\">示範\u003C\u002Fa>如何用提示詞、工具調用和停止條件，將 \u003Ca href=\"\u002Ftag\u002Fanthropic\">Anthropic\u003C\u002Fa> \u003Ca href=\"\u002Fnews\u002Fseriality-gap-video-diffusion-models-zh\">模型的\u003C\u002Fa>任務耐力調成可控的執行能力。\u003C\u002Fp>\u003Ch2>開始之前\u003C\u002Fh2>\u003Cul>\u003Cli>一個 Anthropic 帳號與可用的 API key，文件見 \u003Ca href=\"https:\u002F\u002Fdocs.anthropic.com\u002F\">Anthropic Docs\u003C\u002Fa>，SDK 範例可參考 \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fanthropics\u002Fanthropic-sdk-python\">anthropic-sdk-python\u003C\u002Fa> 或 \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fanthropics\u002Fanthropic-sdk-typescript\">anthropic-sdk-typescript\u003C\u002Fa>\u003C\u002Fli>\u003Cli>Node 20+ 或 Python 3.10+\u003C\u002Fli>\u003Cli>一個支援 HTTPS 的本地開發環境\u003C\u002Fli>\u003Cli>可使用工具調用或函式調用能力的 Claude 模型\u003C\u002Fli>\u003Cli>一個可重複測試的任務場景，例如修復程式碼、整理資料或多步檢索\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>Step 1: 寫出任務邊界\u003C\u002Fh2>\u003Cp>先把「要完成什麼」和「什麼時候必須停下」寫成固定規則。這一步的產出是讓模型知道自己的執行範圍，而不是靠臨場猜測。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784102587342-bz4a.png\" alt=\"Anthropic 任務耐力調校指南\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cpre>\u003Ccode>System prompt example:\n\nYou are an execution agent.\nGoal: finish the task only within the provided scope.\nStop and ask the user when:\n- required input is missing\n- a tool fails twice\n- the task may change external state\n- confidence is below 0.7\n\nNever continue with speculative work after a stop condition.\u003C\u002Fcode>\u003C\u002Fpre>\u003Cp>你應該看到模型在缺少關鍵輸入時主動暫停，而不是繼續試錯或自行補完。\u003C\u002Fp>\u003Ch2>Step 2: 拆成可驗收子目標\u003C\u002Fh2>\u003Cp>把單一句目標拆成一串小步驟，讓模型每做一步都能被檢查。這樣它可以持續推進，但不會因為目標太大而陷入無限循環。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784102588589-gbmb.png\" alt=\"Anthropic 任務耐力調校指南\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cpre>\u003Ccode>Task format:\n1. Restate the goal in one sentence.\n2. List required inputs.\n3. Propose 3 substeps.\n4. Execute one substep.\n5. Report status and ask for approval if blocked.\u003C\u002Fcode>\u003C\u002Fpre>\u003Cp>你應該看到輸出先列出計畫，再執行一小步，最後回報狀態與阻塞原因。\u003C\u002Fp>\u003Ch2>Step 3: 設定工具闸門\u003C\u002Fh2>\u003Cp>把模型的「想做」和真正的外部操作隔開。凡是會改檔、發請求或寫入資料庫的動作，都先經過工具闸門與確認邏輯。\u003C\u002Fp>\u003Cpre>\u003Ccode>Tool policy:\n- Read-only tools: auto-approve\n- Write tools: require explicit confirmation\n- Network tools: require destination allowlist\n- Destructive tools: always ask before use\u003C\u002Fcode>\u003C\u002Fpre>\u003Cp>你應該看到模型在觸發寫入或外部請求前先要求確認，而不是直接執行高風險動作。\u003C\u002Fp>\u003Ch2>Step 4: 定義停止條件與重試上限\u003C\u002Fh2>\u003Cp>讓模型知道什麼情況下要停，不要把失敗當成無限續命的理由。這一步的產出是明確的重試次數、超時規則與退出方式。\u003C\u002Fp>\u003Cpre>\u003Ccode>Retry policy:\n- transient tool error: retry up to 2 times\n- missing user input: stop immediately\n- repeated parsing failure: stop after 1 repair attempt\n- timeout: stop and summarize partial progress\u003C\u002Fcode>\u003C\u002Fpre>\u003Cp>你應該看到模型在工具失敗後最多重試到上限，然後輸出失敗摘要與下一步建議。\u003C\u002Fp>\u003Ch2>Step 5: 記錄軌跡並回放\u003C\u002Fh2>\u003Cp>把每次計畫、工具調用、失敗原因與停止點都記下來，才能分辨模型是在有效堅持，還是在重複繞圈。這一步的產出是可回放的任務紀錄。\u003C\u002Fp>\u003Cpre>\u003Ccode>Log fields:\n- task_id\n- goal\n- substep\n- tool_name\n- result\n- retry_count\n- stop_reason\n- user_intervention_needed\u003C\u002Fcode>\u003C\u002Fpre>\u003Cp>你應該能用同一批任務回放結果，並看到高價值任務完成率更穩定，無效循環次數下降。\u003C\u002Fp>\u003Ctable>\u003Cthead>\u003Ctr>\u003Cth>指標\u003C\u002Fth>\u003Cth>基準／優化前\u003C\u002Fth>\u003Cth>結果／優化後\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd>無效重試次數\u003C\u002Ftd>\u003Ctd>未設定上限時持續增加\u003C\u002Ftd>\u003Ctd>最多 2 次後停止\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>使用者介入次數\u003C\u002Ftd>\u003Ctd>模型盲目繼續時偏高\u003C\u002Ftd>\u003Ctd>缺少輸入時立即請求確認\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>任務完成率\u003C\u002Ftd>\u003Ctd>計畫不清時波動大\u003C\u002Ftd>\u003Ctd>拆分子目標後更穩定\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003Ch2>常見錯誤\u003C\u002Fh2>\u003Cul>\u003Cli>把「耐力」當成「永不停止」。修法是把 stop condition 寫進系統提示詞，並在流程中強制執行。\u003C\u002Fli>\u003Cli>讓模型直接操作高風險工具。修法是把寫入、刪除與外部請求放進確認闸門。\u003C\u002Fli>\u003Cli>只看最終答案，不看過程紀錄。修法是保存子步驟、重試次數與停止原因，再做回放評估。\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>接下來可以看什麼\u003C\u002Fh2>\u003Cp>下一步可以把這套控制層接到你的 Agent 框架，加入權限分級、人工審批與自動回滾，讓模型能持續推進，但不會越界。\u003C\u002Fp>","用提示詞、工具闸門和停止條件，把 Anthropic 模型的任務耐力變成可控的執行能力。","www.zhihu.com","https:\u002F\u002Fwww.zhihu.com\u002Fquestion\u002F2059361484203996542",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784102587342-bz4a.png","ai-agent","zh","5115e111-8a16-41df-a00b-529be3f2c3c9",[17,18,19,20,21],"Anthropic","Claude","tool calling","prompt engineering","Agent",[23,24,25],"先定義任務邊界與停止條件，再談模型耐力。","把大任務拆成可驗收子目標，才能穩定推進。","工具闸門、重試上限與回放紀錄是控制失控的關鍵。",0,"2026-07-15T08:02:31.959666+00:00","2026-07-15T08:02:31.949+00:00","077b933d-cba4-4b6b-9f20-a23c3519c683",{"tags":31,"relatedLang":42,"relatedPosts":46},[32,34,36,38,40],{"name":33,"slug":33},"agent",{"name":20,"slug":35},"prompt-engineering",{"name":19,"slug":37},"tool-calling",{"name":17,"slug":39},"anthropic",{"name":18,"slug":41},"claude",{"id":15,"slug":43,"title":44,"language":45},"anthropic-models-build-better-agent-endurance-en","Anthropic Models: Build Better Agent Endurance","en",[47,53,59,65,71,77],{"id":48,"slug":49,"title":50,"cover_image":51,"image_url":51,"created_at":52,"category":13},"7acb5b4c-de5e-4ded-83cf-82cf93f47a00","google-gemini-enterprise-agent-platform-cloud-service-zh","Google Gemini Enterprise 代理平台把 AI 代理變成雲…","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784059372531-sfab.png","2026-07-14T20:02:24.962585+00:00",{"id":54,"slug":55,"title":56,"cover_image":57,"image_url":57,"created_at":58,"category":13},"1f24f862-f37b-4bd9-b78d-434713905348","workbuddy-harness-engineering-agent-reliability-zh","WorkBuddy 證明了 Agent 可靠性不靠大模型本身","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783987377305-yt32.png","2026-07-14T00:02:31.385926+00:00",{"id":60,"slug":61,"title":62,"cover_image":63,"image_url":63,"created_at":64,"category":13},"415418c7-749e-4352-a21d-d2fa62d8b96b","perplexity-teammate-coding-agent-strategy-zh","Perplexity 應把 Teammate 做成 coding agent，…","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783589576714-n7gs.png","2026-07-09T09:32:25.568287+00:00",{"id":66,"slug":67,"title":68,"cover_image":69,"image_url":69,"created_at":70,"category":13},"884f1bb8-4bae-4cfa-87d1-b323be1d6166","hp-adopts-openai-frontier-global-operations-zh","HP 將 Frontier 送進全球營運","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783362776958-1e1n.png","2026-07-06T18:32:30.897068+00:00",{"id":72,"slug":73,"title":74,"cover_image":75,"image_url":75,"created_at":76,"category":13},"b97b8932-56a9-431f-8270-3f892f8feb94","build-production-vector-db-rag-pipeline-zh","用 n8n 建出可上線的向量資料庫","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783258377362-cg5x.png","2026-07-05T13:32:23.466634+00:00",{"id":78,"slug":79,"title":80,"cover_image":81,"image_url":81,"created_at":82,"category":13},"72b05d2a-7461-4885-a57f-506fd42d714d","ornith-1-agent-coding-server-template-zh","Ornith-1 把代理寫碼變成伺服器","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783085625241-1df8.png","2026-07-03T13:33:20.199747+00:00",[84,89,94,99,104,109,114,119,124,129],{"id":85,"slug":86,"title":87,"created_at":88},"4ae1e197-1d3d-4233-8733-eafe9cb6438b","claude-now-uses-your-pc-to-finish-tasks-zh","Claude 開始幫你操作電腦","2026-03-26T07:20:48.457387+00:00",{"id":90,"slug":91,"title":92,"created_at":93},"5bede67f-e21c-413d-9ab8-54a3c3d26227","googles-2026-ai-agent-report-decoded-zh","Google 2026 AI Agent 報告解讀","2026-03-26T11:15:22.651956+00:00",{"id":95,"slug":96,"title":97,"created_at":98},"2987d097-563f-46c7-b76f-b558d8ef7c2b","kimi-k25-review-stronger-still-not-legend-zh","Kimi K2.5 評測：更強，但還不是神作","2026-03-27T07:15:55.277513+00:00",{"id":100,"slug":101,"title":102,"created_at":103},"95c9053b-e3f4-4cb5-aace-5c54f4c9e044","claude-code-controls-mac-desktop-zh","Claude Code 也能操控 Mac 了","2026-03-28T03:01:58.58121+00:00",{"id":105,"slug":106,"title":107,"created_at":108},"dc58e153-e3a8-4c06-9b96-1aa64eabbf5f","cloudflare-100x-faster-ai-agent-sandbox-zh","Cloudflare 的 AI 沙箱跑超快","2026-03-28T03:09:44.142236+00:00",{"id":110,"slug":111,"title":112,"created_at":113},"1c8afc56-253f-47a2-979f-1065ff072f2a","openai-backs-isara-agent-swarm-bet-zh","OpenAI 挺 Isara 的 agent swarm …","2026-03-28T03:15:27.513155+00:00",{"id":115,"slug":116,"title":117,"created_at":118},"7379b422-576e-45df-ad5a-d57a0d9dd467","openai-plan-automated-ai-researcher-zh","OpenAI 想做自動化 AI 研究員","2026-03-28T03:17:42.090548+00:00",{"id":120,"slug":121,"title":122,"created_at":123},"48c9889e-86df-450b-a356-e4a4b7c83c5b","harness-engineering-ai-agent-reliability-2026-zh","駕馭工程：從「馬具」到「作業系統」，AI Agent 可靠性的終極密碼","2026-03-31T06:42:53.556721+00:00",{"id":125,"slug":126,"title":127,"created_at":128},"96d8e8c8-1edd-475d-9145-b1e7a1b02b65","mcp-explained-from-prompts-to-production-zh","MCP 怎麼把提示詞變工作流","2026-04-01T09:24:39.321274+00:00",{"id":130,"slug":131,"title":132,"created_at":133},"f2ca7720-b471-4ce5-9336-2a9ac2a876fd","amazon-bedrock-agents-multi-agent-workflows-zh","Amazon Bedrock Agents 進入多代理工作流","2026-04-01T09:30:29.945429+00:00"]