[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-claude-code-harness-engineering-design-zh":3,"tags-claude-code-harness-engineering-design-zh":34,"related-lang-claude-code-harness-engineering-design-zh":51,"related-posts-claude-code-harness-engineering-design-zh":55,"series-tools-44b25589-877f-4d11-9987-9b29f2bdf2e7":92},{"id":4,"title":5,"content":6,"summary":7,"source":8,"source_url":9,"author":10,"image_url":11,"keywords":12,"language":22,"translated_content":10,"views":23,"is_premium":24,"created_at":25,"updated_at":25,"cover_image":11,"published_at":26,"rewrite_status":27,"rewrite_error":10,"rewritten_from_id":28,"slug":29,"category":30,"related_article_id":31,"status":32,"google_indexed_at":33,"x_posted_at":10,"tweet_text":10,"title_rewritten_at":10,"title_original":10,"key_takeaways":10,"topic_cluster_id":10,"embedding":10,"is_canonical_seed":24},"44b25589-877f-4d11-9987-9b29f2bdf2e7","Claude Code 的 Harness 工程思路","\u003Cp>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fanthropics\u002Fclaude-code\" target=\"_blank\" rel=\"noopener\">Claude Code\u003C\u002Fa> 很有意思。它不是單純的聊天框。它把怎麼讓模型幹活，直接攤開來給你看。\u003C\u002Fp>\u003Cp>Anthropic 把 \u003Ca href=\"https:\u002F\u002Fmodelcontextprotocol.io\u002F\" target=\"_blank\" rel=\"noopener\">MCP\u003C\u002Fa>、Skills、Hooks、Subagents 一起放進產品。這等於把 Harness 設計公開了。說白了，它在示範一件事：模型再強，外層執行系統爛掉，結果還是會亂。\u003C\u002Fp>\u003Cp>我覺得這點很值得看。因為現在很多 AI 工具，還停在「會講」這一層。\u003Ca href=\"\u002Fnews\u002F8-hidden-claude-code-features-leaked-source-zh\">Clau\u003C\u002Fa>de Code 則把工具接入、上下文管理、權限控制、任務拆分，做成一套能跑的工作流。\u003C\u002Fp>\u003Ch2>Claude Code 為什麼值得單獨看\u003C\u002Fh2>\u003Cp>\u003Ca href=\"\u002Fnews\u002Fclaude-code-architecture-governance-practice-zh\">Clau\u003C\u002Fa>de Code 不是把 Claude 塞進終端機而已。它更像一個圍繞模型搭起來的執行系統。模型負責理解和生成，外層 Harness 負責拆任務、接工具、壓風險，再把結果送回去繼續迭代。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1775113337940-qsjx.png\" alt=\"Claude Code 的 Harness 工程思路\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>這種做法，和很多 IDE 外掛不太一樣。很多工具只是把提示詞丟給模型，再把輸出貼回編輯器。\u003Ca href=\"\u002Fnews\u002Fclaude-code-source-leak-hidden-internals-zh\">Clau\u003C\u002Fa>de Code 則更像模型驅動的操作外殼。它把上下文、工具、權限、子任務，放在同一條流程裡。\u003C\u002Fp>\u003Cp>從產品角度看，這種設計很實際。第一，使用者更清楚模型做了什麼。第二，團隊可以把最佳做法固定下來。你不用期待每個人都會寫超強提示詞。\u003C\u002Fp>\u003Cp>這也是它會被反覆討論的原因。它不只是 AI 編程工具。它更像 Anthropic 在公開展示自己的 Harness Engineering 思路。\u003C\u002Fp>\u003Cul>\u003Cli>\u003Ca href=\"https:\u002F\u002Fmodelcontextprotocol.io\u002F\" target=\"_blank\" rel=\"noopener\">MCP\u003C\u002Fa> 讓 Claude Code 接外部工具更標準。\u003C\u002Fli>\u003Cli>Skills 把常見任務包成可重用能力。\u003C\u002Fli>\u003Cli>Hooks 可以插入檢查、確認、審計動作。\u003C\u002Fli>\u003Cli>Subagents 能把大任務拆成小角色。\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>Harness Engineering 到底在解什麼\u003C\u002Fh2>\u003Cp>Anthropic 的重點，一直不是只讓模型會回話。重點是讓它在真實任務裡穩定做事。這也是 Harness Engineering 的核心。模型能力只是起點，外層執行系統才決定能不能進生產環境。\u003C\u002Fp>\u003Cp>你可以把它想成兩種工作流。第一種是純提示詞流程，輸入問題，拿到答案。第二種是完整 Harness，把工具呼叫、狀態保存、錯誤恢復、權限邊界都放進流程。前者適合 demo，後者才適合長期用。\u003C\u002Fp>\u003Cblockquote>\"The most important thing we can do is make AI systems that are helpful, harmless, and honest.\" — Dario Amodei\u003C\u002Fblockquote>\u003Cp>這句話來自 \u003Ca href=\"https:\u002F\u002Fwww.anthropic.com\u002F\" target=\"_blank\" rel=\"noopener\">Anthropic\u003C\u002Fa> CEO Dario Amodei。放在 Claude Code 上看，很貼切。因為 helpful、harmless、honest，不是靠一句系統提示就能搞定。\u003C\u002Fp>\u003Cp>它需要工具層、任務層、權限層一起配合。講白了就是，模型要做事，但不能亂做事。這就是 Harness Engineering 真正在管的東西。\u003C\u002Fp>\u003Cp>Claude Code 把這種想法直接產品化。它沒有把複雜性藏起來，而是拆成幾個可以配置的接口。你可以按任務，選擇要不要開啟某些能力。\u003C\u002Fp>\u003Ch2>MCP、Skills、Hooks、Subagents 怎麼分工\u003C\u002Fh2>\u003Cp>把 Claude Code 拆開看，四個機制各有角色。MCP 負責外部連接。Skills 負責能力封裝。Hooks 負責流程控制。Subagents 負責任務分解。它們不是重複，而是分層處理同一件事。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1775113352567-wkrr.png\" alt=\"Claude Code 的 Harness 工程思路\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>MCP 的價值很明確。它把接什麼工具、讀什麼資料，從應用內部邏輯裡抽出來，變成標準協議。對開發者來說，接 Git、資料庫、內部知識庫、部署系統時，不用每次都重寫一套接口。\u003C\u002Fp>\u003Cp>Skills 比較像可重用的任務模板。像程式碼審查、文件整理、測試生成、遷移腳本，這些高頻動作都能包成技能包。這樣做的好處，是輸出風格比較穩，團隊也比較好統一標準。\u003C\u002Fp>\u003Cp>Hooks 則是安全和治理的抓手。像高風險操作前確認，或是在輸出前做靜態檢查，都可以交給 Hook。它讓模型不只是會做，還要照規則做。\u003C\u002Fp>\u003Cp>Subagents 解的是上下文膨脹。當一個代理同時做規劃、搜尋、編輯、驗證，很容易亂掉。拆成多個子代理後，每個角色只盯一小塊，通常會更穩。\u003C\u002Fp>\u003Cul>\u003Cli>\u003Ca href=\"https:\u002F\u002Fmodelcontextprotocol.io\u002F\" target=\"_blank\" rel=\"noopener\">MCP\u003C\u002Fa> 解決工具接入標準化。\u003C\u002Fli>\u003Cli>Skills 解決任務重用。\u003C\u002Fli>\u003Cli>Hooks 解決流程控制和審計。\u003C\u002Fli>\u003Cli>Subagents 解決複雜任務拆分。\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>和其他 AI 編程工具比，差在哪\u003C\u002Fh2>\u003Cp>把 Claude Code 和常見 AI 編程工具放一起看，差異就很明顯。很多工具比較像增強版聊天視窗，重點在生成速度和互動順手。Claude Code 則更像可編排的執行層，重點在任務治理和系統整合。\u003C\u002Fp>\u003Cp>這種差異會直接反映在實際使用上。第一，它更適合接企業內部工具鏈。第二，它更容易把流程固定住。第三，它更適合做可審計的自動化動作，而不是只做一次性問答。\u003C\u002Fp>\u003Cp>如果只看表面體驗，你可能會覺得它比較工程化。沒那麼輕巧。可是從長期維護看，工程化反而是優點。模型工具一旦進到團隊協作和生產流程，最怕的就是不可預測。\u003C\u002Fp>\u003Cp>可以直接這樣比：\u003C\u002Fp>\u003Cul>\u003Cli>一般聊天式工具：適合探索和草稿。\u003C\u002Fli>\u003Cli>IDE 插件：適合局部補全和改寫。\u003C\u002Fli>\u003Cli>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fanthropics\u002Fclaude-code\" target=\"_blank\" rel=\"noopener\">Claude Code\u003C\u002Fa>：適合把模型放進真實工作流。\u003C\u002Fli>\u003Cli>企業代理平台：適合跨系統自動化和權限管理。\u003C\u002Fli>\u003C\u002Ful>\u003Cp>還有一個很現實的點。當模型能力越來越接近時，真正拉開差距的，往往是 Harness。誰能更好地組織上下文、接工具、控風險，誰就更容易把模型能力變成可交付結果。\u003C\u002Fp>\u003Ch2>這背後其實是產業方向\u003C\u002Fh2>\u003Cp>Claude Code 透露的訊號，不只關於一個產品。它也在說，AI 工具競爭會越來越像執行系統競爭。不是誰回得更會講，而是誰能穩定做完事情。\u003C\u002Fp>\u003Cp>這點在企業場景特別重要。你做內部自動化，最怕不是模型不夠聰明，而是流程失控。像誤刪資料、亂發指令、權限越界，這些都不是單靠 prompt 能壓住的。\u003C\u002Fp>\u003Cp>所以我會把 Claude Code 看成一個樣板。它把工具接入、任務拆分、審計、上下文管理，直接做成產品的一部分。這種思路，比單純追求更長上下文或更大參數，來得更接近實戰。\u003C\u002Fp>\u003Cp>如果你在做 AI 軟體，接下來該問的問題很直接：你的 Harness 有沒有設計好？你有沒有定義哪些動作要確認？哪些資料能讀？哪些任務要拆？這些問題，比模型分數更接近生產現場。\u003C\u002Fp>\u003Ch2>結尾：我會怎麼看這件事\u003C\u002Fh2>\u003Cp>我自己的判斷很簡單。接下來真正有價值的 AI 編程工具，會越來越像 Claude Code。它們不會只賣模型能力，而是把 Harness 當成產品核心。\u003C\u002Fp>\u003Cp>如果你現在還在用「一個提示詞解全部」的思路做工具，遲早會碰到穩定性和治理問題。比較實際的做法，是先把工具接入、權限邊界、任務拆分、審計機制做好，再談模型要多強。\u003C\u002Fp>\u003Cp>你可以先問團隊一個問題：如果模型今天答對 80%，剩下 20% 的錯誤，誰來接？這個答案，往往就是你產品能不能上線的分水嶺。\u003C\u002Fp>","Claude Code 把 MCP、Skills、Hooks 和 Subagents 直接端上檯面，讓人看到 Anthropic 怎麼把 Harness Engineering 做進產品。","zhuanlan.zhihu.com","https:\u002F\u002Fzhuanlan.zhihu.com\u002Fp\u002F2021603278606087058",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1775113337940-qsjx.png",[13,14,15,16,17,18,19,20,21],"Claude Code","Harness Engineering","Anthropic","MCP","Skills","Hooks","Subagents","LLM","AI 編程","zh",2,false,"2026-04-02T05:21:31.088419+00:00","2026-04-02T05:21:31.062+00:00","done","79498b22-6c42-4368-82a9-373b651522be","claude-code-harness-engineering-design-zh","tools","da4a873e-dd38-48df-acce-2ddcc49ed456","published","2026-04-09T09:00:51.815+00:00",[35,37,39,41,43,45,47,49],{"name":36,"slug":36},"hooks",{"name":14,"slug":38},"harness-engineering",{"name":16,"slug":40},"mcp",{"name":13,"slug":42},"claude-code",{"name":15,"slug":44},"anthropic",{"name":20,"slug":46},"llm",{"name":48,"slug":48},"subagents",{"name":21,"slug":50},"ai-編程",{"id":31,"slug":52,"title":53,"language":54},"claude-code-harness-engineering-design-en","Claude Code 里的 Harness Engineering 思路","en",[56,62,68,74,80,86],{"id":57,"slug":58,"title":59,"cover_image":60,"image_url":60,"created_at":61,"category":30},"d058a76f-6548-4135-8970-f3a97f255446","why-gemini-api-pricing-is-cheaper-than-it-looks-zh","為什麼 Gemini API 定價其實比看起來更便宜","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778869845081-j4m7.png","2026-05-15T18:30:25.797639+00:00",{"id":63,"slug":64,"title":65,"cover_image":66,"image_url":66,"created_at":67,"category":30},"68e4be16-dc38-4524-a6ea-5ebe22a6c4fb","why-vidhub-huiyuan-hutong-bushi-quan-shebei-tongyong-zh","為什麼 VidHub 會員互通不是「買一次全設備通用」","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778789450987-advz.png","2026-05-14T20:10:24.048988+00:00",{"id":69,"slug":70,"title":71,"cover_image":72,"image_url":72,"created_at":73,"category":30},"7a1e174f-746b-4e82-a0e3-b2475ab39747","why-buns-zig-to-rust-experiment-is-right-zh","為什麼 Bun 的 Zig-to-Rust 實驗是對的","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778767879127-5dna.png","2026-05-14T14:10:26.886397+00:00",{"id":75,"slug":76,"title":77,"cover_image":78,"image_url":78,"created_at":79,"category":30},"e742fc73-5a65-4db3-ad17-88c99262ceb7","why-openai-api-pricing-is-product-strategy-zh","為什麼 OpenAI API 定價是產品策略，不是註腳","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778749859485-chvz.png","2026-05-14T09:10:26.003818+00:00",{"id":81,"slug":82,"title":83,"cover_image":84,"image_url":84,"created_at":85,"category":30},"c757c5d8-eda9-45dc-9020-4b002f4d6237","why-claude-code-prompt-design-beats-ide-copilots-zh","為什麼 Claude Code 的提示設計贏過 IDE Copilot","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778742645084-dao9.png","2026-05-14T07:10:29.371901+00:00",{"id":87,"slug":88,"title":89,"cover_image":90,"image_url":90,"created_at":91,"category":30},"4adef3ab-9f07-4970-91cf-77b8b581b348","why-databricks-model-serving-is-right-default-zh","為什麼 Databricks Model Serving 是生產推論的正確預設","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778692245329-a2wt.png","2026-05-13T17:10:30.659153+00:00",[93,98,103,108,113,118,123,128,133,138],{"id":94,"slug":95,"title":96,"created_at":97},"de769291-4574-4c46-a76d-772bd99e6ec9","googles-biggest-gemini-launches-in-2026-zh","Google 2026 最大 Gemini 盤點","2026-03-26T07:26:39.21072+00:00",{"id":99,"slug":100,"title":101,"created_at":102},"855cd52f-6fab-46cc-a7c1-42195e8a0de4","surepath-real-time-mcp-policy-controls-zh","SurePath 推出即時 MCP 政策控管","2026-03-26T07:57:40.77233+00:00",{"id":104,"slug":105,"title":106,"created_at":107},"9b19ab54-edef-4dbd-9ce4-a51e4bae4ebb","mcp-in-2026-the-ai-tool-layer-teams-use-zh","2026 年 MCP：團隊真的在用的 AI 工具層","2026-03-26T08:01:46.589694+00:00",{"id":109,"slug":110,"title":111,"created_at":112},"af9c46c3-7a28-410b-9f04-32b3de30a68c","prompting-in-2026-what-actually-works-zh","2026 提示工程，真正有用的是什麼","2026-03-26T08:08:12.453028+00:00",{"id":114,"slug":115,"title":116,"created_at":117},"05553086-6ed0-4758-81fd-6cab24b575e0","garry-tan-open-sources-claude-code-toolkit-zh","Garry Tan 開源 Claude Code 工具包","2026-03-26T08:26:20.068737+00:00",{"id":119,"slug":120,"title":121,"created_at":122},"042a73a2-18a2-433d-9e8f-9802b9559aac","github-ai-projects-to-watch-in-2026-zh","2026 必看 20 個 GitHub AI 專案","2026-03-26T08:28:09.619964+00:00",{"id":124,"slug":125,"title":126,"created_at":127},"a5f94120-ac0d-4483-9a8b-63590071ac6a","claude-code-vs-cursor-2026-zh","Claude Code 與 Cursor 深度對比：202…","2026-03-26T13:27:14.279193+00:00",{"id":129,"slug":130,"title":131,"created_at":132},"0975afa1-e0c7-4130-a20d-d890eaed995e","practical-github-guide-learning-ml-2026-zh","2026 機器學習入門 GitHub 實用指南","2026-03-27T01:16:49.712576+00:00",{"id":134,"slug":135,"title":136,"created_at":137},"bfdb467a-290f-4a80-b3a9-6f081afb6dff","aiml-2026-student-ai-ml-lab-repo-review-zh","AIML-2026：像課綱的學生實驗 Repo","2026-03-27T01:21:51.467798+00:00",{"id":139,"slug":140,"title":141,"created_at":142},"80cabc3e-09fc-4ff5-8f07-b8d68f5ae545","ai-trending-github-repos-and-research-feeds-zh","AI Trending：把 AI 資源收成一張表","2026-03-27T01:31:35.262183+00:00"]