{"id":13093,"date":"2025-08-14T23:19:36","date_gmt":"2025-08-14T15:19:36","guid":{"rendered":"https:\/\/ihower.tw\/blog\/?p=13093"},"modified":"2025-08-21T11:48:11","modified_gmt":"2025-08-21T03:48:11","slug":"agent-prompting-design","status":"publish","type":"post","link":"https:\/\/ihower.tw\/blog\/13093-agent-prompting-design","title":{"rendered":"\u5f9e Prompting \u57fa\u672c\u7d50\u69cb\u5230 Agent Prompting \u8a2d\u8a08\u539f\u5247"},"content":{"rendered":"\n

Anthropic \u6700\u8fd1\u624d\u91cb\u51fa\u4e86\u4ed6\u5011\u5728 2025\/5\/22 \u958b\u767c\u8005\u5927\u6703\u7684\u5b8c\u6574\u5f71\u7247<\/a>\uff0c\u7576\u6642\u7684\u91cd\u982d\u6232\u662f Claude 4 \u6a21\u578b\u767c\u5e03\u3002\u5176\u4e2d\u6709\u5169\u5834\u95dc\u65bc prompting \u6559\u5b78\u7684\u6f14\u8b1b\u5167\u5bb9\u5f88\u4e0d\u932f\uff0c\u9019\u5169\u5834\u6f14\u8b1b\u5f9e\u57fa\u790e Prompt \u5230\u91dd\u5c0d\u7684 Agent \u7684 prompting \uff0c\u5c55\u73fe\u4e86 prompt engineering \u7684\u4e0d\u540c\u5c64\u6b21\uff0c\u63a8\u85a6\u5927\u5bb6\u4e00\u770b\u3002\u4ee5\u4e0b\u662f\u6211\u7684\u89e3\u8b80\u6574\u7406\u3002<\/p>\n\n\n\n

\u7b2c\u4e00\u5834: Prompting 101<\/a><\/h2>\n\n\n\n
\"\"<\/a><\/figure>\n\n\n\n\n\n\n\n

\u9019\u4e00\u5834\u7528\u4e86\u4e00\u500b\u5be6\u969b\u7684\u6848\u4f8b\u4f86\u8b1b\u89e3\uff1a\u6c7d\u8eca\u4fdd\u96aa\u7406\u8ce0\u7684\u81ea\u52d5\u5316\u8655\u7406\u3002\u60f3\u50cf\u4e00\u4e0b\uff0c\u4fdd\u96aa\u516c\u53f8\u6bcf\u5929\u8981\u8655\u7406\u5927\u91cf\u7684\u8eca\u798d\u7406\u8ce0\uff0c\u6bcf\u500b\u6848\u4ef6\u90fd\u6709\u73fe\u5834\u624b\u5de5\u5716\u548c\u5236\u5f0f\u8eca\u798d\u5831\u544a\u8868\u55ae\uff0cAI \u9700\u8981\u5224\u65b7\u300c\u4e8b\u6545\u767c\u751f\u4e86\u4ec0\u9ebc\u300d\u4ee5\u53ca\u300c\u8ab0\u662f\u904e\u932f\u65b9\u300d\u3002<\/p>\n\n\n\n

\u8b1b\u8005\u76f4\u63a5\u5728 Anthropic \u5f8c\u53f0 Console \u793a\u7bc4\u5982\u4f55\u8a2d\u8a08 prompt\uff0c\u4e26\u8b1b\u89e3\u4e86\u6a19\u6e96\u7684 prompt \u7d50\u69cb\u6700\u4f73\u5be6\u8e10\u3002<\/p>\n\n\n\n

Prompt Structure \u57fa\u790e 5 \u500b\u7d50\u69cb\u5143\u7d20<\/h3>\n\n\n\n
\"\"<\/a><\/figure>\n\n\n\n
    \n
  1. 1-2 \u53e5\u8a71\u5efa\u7acb\u89d2\u8272\u548c\u9ad8\u5c64\u7d1a\u4efb\u52d9\u63cf\u8ff0<\/strong> – \u7c21\u6f54\u5730\u544a\u8a34 AI \u5b83\u7684\u8eab\u4efd\u548c\u4e3b\u8981\u76ee\u6a19<\/li>\n\n\n\n
  2. \u52d5\u614b\/\u6aa2\u7d22\u5167\u5bb9<\/strong> – \u653e\u5165\u9700\u8981\u8655\u7406\u7684\u52d5\u614b\u8cc7\u6599\uff0c\u5982\u4f7f\u7528\u8005\u504f\u597d\u3001\u5730\u9ede\u8cc7\u8a0a\u7b49<\/li>\n\n\n\n
  3. \u8a73\u7d30\u4efb\u52d9\u6307\u4ee4<\/strong> – \u5177\u9ad4\u8aaa\u660e\u8981\u5982\u4f55\u57f7\u884c\u4efb\u52d9\u7684\u6b65\u9a5f\u548c\u898f\u5247<\/li>\n\n\n\n
  4. \u7bc4\u4f8b\/n-shot (optional)<\/strong> – \u63d0\u4f9b\u8f38\u5165\u8f38\u51fa\u7684\u7bc4\u4f8b\u4f86\u5f15\u5c0e\u6a21\u578b<\/li>\n\n\n\n
  5. \u91cd\u8907\u95dc\u9375\u6307\u4ee4<\/strong> – \u5c0d\u65bc\u7279\u5225\u9577\u7684 prompt\uff0c\u5728\u6700\u5f8c\u91cd\u7533\u6700\u91cd\u8981\u7684\u6307\u4ee4<\/li>\n<\/ol>\n\n\n\n

    \u64f4\u5145\u5230\u66f4\u7d30\u7bc0\u7684 10 \u9ede\u7d50\u69cb<\/h3>\n\n\n\n
    \"\"<\/a><\/figure>\n\n\n\n

    \u7576\u9700\u8981\u8655\u7406\u66f4\u8907\u96dc\u7684\u4efb\u52d9\u6642\uff0c\u53ef\u4ee5\u5c07 prompt \u7d50\u69cb\u64f4\u5145\u5230 10 \u500b\u5143\u7d20\uff1a<\/p>\n\n\n\n

      \n
    1. \u4efb\u52d9\uff08Task context\uff09<\/strong> – \u66f4\u8a73\u7d30\u7684\u89d2\u8272\u8a2d\u5b9a\u548c\u80cc\u666f\u8aaa\u660e<\/li>\n\n\n\n
    2. \u8a9e\u8abf\uff08Tone context\uff09<\/strong> – \u660e\u78ba\u8a2d\u5b9a\u56de\u61c9\u7684\u8a9e\u6c23\u3001\u98a8\u683c\u548c\u500b\u6027<\/li>\n\n\n\n
    3. \u80cc\u666f\u8cc7\u6599\u3001\u6587\u4ef6\u548c\u5716\u7247<\/strong> – \u6240\u6709\u9700\u8981\u53c3\u8003\u7684\u975c\u614b\u8cc7\u6599<\/li>\n\n\n\n
    4. \u8a73\u7d30\u4efb\u52d9\u63cf\u8ff0\u8207\u898f\u5247<\/strong> – \u5b8c\u6574\u7684\u57f7\u884c\u6b65\u9a5f\u3001\u5224\u65b7\u6a19\u6e96\u548c\u8655\u7406\u898f\u5247<\/li>\n\n\n\n
    5. \u7bc4\u4f8b\uff08Examples\uff09<\/strong> – \u591a\u500b\u4e0d\u540c\u60c5\u5883\u7684\u8f38\u5165\u8f38\u51fa\u7bc4\u4f8b<\/li>\n\n\n\n
    6. \u5c0d\u8a71\u6b77\u53f2\uff08Conversation history\uff09<\/strong> – \u5982\u679c\u6709\u5148\u524d\u7684\u5c0d\u8a71\u8a18\u9304\u9700\u8981\u53c3\u8003<\/li>\n\n\n\n
    7. \u7acb\u5373\u4efb\u52d9\u63cf\u8ff0\u6216\u8acb\u6c42<\/strong> – \u7576\u524d\u9019\u4e00\u8f2a\u5177\u9ad4\u8981\u8655\u7406\u7684\u4efb\u52d9<\/li>\n\n\n\n
    8. \u9010\u6b65\u601d\u8003\u63d0\u793a<\/strong> – \u52a0\u5165\u300cThink step by step\u300d\u6216\u300cTake a deep breath\u300d\u7b49\u5f15\u5c0e\u8a9e<\/li>\n\n\n\n
    9. \u8f38\u51fa\u683c\u5f0f\u8981\u6c42\uff08Output formatting\uff09<\/strong> – \u8a73\u7d30\u6307\u5b9a\u56de\u61c9\u7684\u7d50\u69cb\u3001\u683c\u5f0f\u548c\u6a23\u5f0f<\/li>\n\n\n\n
    10. \u9810\u586b\u56de\u61c9\uff08Prefilled response\uff09<\/strong> – \u9810\u5148\u586b\u5165\u90e8\u5206\u56de\u61c9\u5167\u5bb9\u4f86\u5f15\u5c0e\u6a21\u578b (\u7de8\u6309: \u9019\u662f Claude API \u624d\u6709\u7684\u529f\u80fd\uff0c\u53ef\u4ee5\u9810\u586b Assistant \u56de\u61c9\u3002OpenAI API \u6c92\u6709\u9019\u529f\u80fd)<\/li>\n<\/ol>\n\n\n\n

      \u9019\u500b\u64f4\u5145\u7248\u672c\u7279\u5225\u9069\u5408\u8655\u7406\u9700\u8981\u591a\u91cd\u8003\u91cf\u7684\u8907\u96dc\u4efb\u52d9\uff0c\u6bcf\u500b\u5143\u7d20\u90fd\u6709\u5176\u7279\u5b9a\u7528\u9014\uff0c\u53ef\u4ee5\u6839\u64da\u5be6\u969b\u9700\u6c42\u9078\u64c7\u4f7f\u7528\u3002<\/p>\n\n\n\n

      Extended Thinking \u7528\u9014<\/h3>\n\n\n\n
      \"\"<\/a><\/figure>\n\n\n\n

      \u8b1b\u8005\u6700\u5f8c\u63d0\u5230 Extended Thinking \u7684\u7528\u9014: \u4f55\u6642\u7528? \u6709\u6c92\u6709\u7f3a\u9ede?<\/p>\n\n\n\n

      \n

      \u7de8\u6309: Extended Thinking<\/a> \u5c31\u662f Claude \u6a21\u578b\u662f\u5426\u8981\u555f\u7528\u589e\u5f37\u63a8\u7406\u80fd\u529b\uff0c\u6703\u7522\u751f\u984d\u5916\u7684 reasoning tokens\uff0c\u9019\u500b\u9810\u8a2d\u662f\u95dc\u7684\u3002<\/p>\n<\/blockquote>\n\n\n\n

      Extended Thinking \u53ef\u4ee5\u7528\u4f86\u8f14\u52a9\u4f60\u7684 system prompt \u958b\u767c\u3002\u5728\u958b\u767c\u521d\u671f\u4f7f\u7528\u9019\u500b\u529f\u80fd\uff0c\u89c0\u5bdf\u6a21\u578b\u7684\u63a8\u7406\u904e\u7a0b\u662f\u600e\u9ebc\u60f3\u7684\uff0c\u4e86\u89e3\u6a21\u578b\u7684\u601d\u8003\u65b9\u5f0f\u6709\u52a9\u65bc\u4f60\u512a\u5316 prompt\u3002<\/p>\n\n\n\n

      Extended Thinking \u7684\u7f3a\u9ede: \u56e0\u70ba\u6bcf\u6b21\u90fd\u91cd\u65b0\u767c\u660e\u8f2a\u5b50\uff0c\u56e0\u6b64\u82b1\u8cbb tokens \u6210\u672c\u8f03\u9ad8\uff1b\u53e6\u5916\u56e0\u70ba\u6eab\u5ea6\u56fa\u5b9a\u662f 1 \u4e0d\u80fd\u8abf\u6574\uff0c\u6240\u4ee5\u6bd4\u8f03\u4e0d\u5bb9\u6613\u91cd\u73fe\u76f8\u540c\u7684\u7d50\u679c\u3002<\/p>\n\n\n\n

      \u7b2c\u4e8c\u5834\uff1aPrompting for Agents<\/a><\/h2>\n\n\n\n
      \"\"<\/a><\/figure>\n\n\n\n

      \u7b2c\u4e8c\u5834\u6f14\u8b1b\u5247\u9032\u5165\u4e86\u4e0d\u540c\u7684\u5834\u666f: \u91dd\u5c0d Agent \u7cfb\u7d71\u7684 prompting \u8a2d\u8a08\uff0c\u4e5f\u5c31\u662f Agent \u7684 system prompt \u8981\u600e\u9ebc\u5beb\u3002\u8b1b\u8005\u9996\u5148\u63d0\u4f9b\u4e86\u4e00\u500b\u5be6\u7528\u7684\u6aa2\u67e5\u6e05\u55ae\uff0c\u5e6b\u52a9\u5224\u65b7\u662f\u5426\u771f\u7684\u9700\u8981\u5efa\u69cb Agent\uff0c\u6216\u662f\u7528 workflow \u5c31\u597d\u3002<\/p>\n\n\n\n

      \n

      \u9019\u88e1\u6211\u5c31\u4e0d\u5c55\u958b\u4ec0\u9ebc\u662f LLM-based Agent \u4e86\uff0c\u6c92\u6709\u9019\u500b\u77e5\u8b58\u5132\u5099\u7684\u8a71\uff0c\u8acb\u770b\u6211\u4e4b\u524d\u7684 Agent \u5206\u4eab<\/a>\u3002<\/p>\n<\/blockquote>\n\n\n\n

      \"\"<\/a><\/figure>\n\n\n\n
      \"\"<\/a><\/figure>\n\n\n\n
      \"\"<\/a><\/figure>\n\n\n\n

      \u63a5\u8457\u624d\u662f\u91cd\u982d\u6232\uff0c\u5982\u4f55\u8a2d\u8a08\u7d66 Agent \u7684 System prompt\u3002\u8acb\u6ce8\u610f\u9019\u500b\u5834\u666f\u8acb\u4e0d\u8981\u5957\u7528\u4e0a\u4e00\u56de\u7d50\u69cb\u5316\u4e94\u500b\u5143\u7d20\u4e86\u3002<\/p>\n\n\n\n

      Agent Prompting \u7684 7 \u500b\u6838\u5fc3\u539f\u5247\uff1a<\/h3>\n\n\n\n
      \"\"<\/a><\/figure>\n\n\n\n
        \n
      1. \u50cf\u4f60\u7684 Agent \u4e00\u6a23\u601d\u8003<\/strong>\n