{"id":710,"date":"2026-05-24T02:04:03","date_gmt":"2026-05-24T02:04:03","guid":{"rendered":"https:\/\/ketoha.co.jp\/blog\/?p=710"},"modified":"2026-05-24T02:15:43","modified_gmt":"2026-05-24T02:15:43","slug":"cursorsdk-%e3%81%a7-llm-as-a-judge-%e3%81%aa%e3%82%b3%e3%83%bc%e3%83%89%e3%83%ac%e3%83%93%e3%83%a5%e3%83%bc%e3%82%92%e3%81%95%e3%81%9b%e3%82%8b","status":"publish","type":"post","link":"https:\/\/ketoha.co.jp\/blog\/?p=710","title":{"rendered":"CursorSDK \u3067 LLM as a judge \u306a\u30b3\u30fc\u30c9\u30ec\u30d3\u30e5\u30fc\u3092\u3055\u305b\u308b"},"content":{"rendered":"\n
\"\"<\/a><\/figure>\n\n\n\n

<\/p>\n\n\n\n

AI\u30de\u30f3\u30bb\u30fc\u306a\u79c1\u3068\u30ec\u30d3\u30e5\u30fc\u306b\u75b2\u308c\u305f\u5148\u8f29<\/h2>\n\n\n\n

\u30d0\u30a4\u30d6\u30b3\u30fc\u30c7\u30a3\u30f3\u30b0\u306e\u7d50\u679c\u306f\u30b9\u30ad\u30eb\u30b9\u30bf\u30c3\u30af\u306a\u5148\u8f29\u30a8\u30f3\u30b8\u30cb\u30a2\u304c\u30ec\u30d3\u30e5\u30fc\u3057\u3066\u307e\u3059\u304b\uff1f
AI\u3092\u77e5\u3089\u306a\u3044\uff08\u4fe1\u7528\u3057\u306a\u3044\uff09\u5148\u8f29\u306f\u3001\u4eba\u529b\u30ec\u30d3\u30e5\u30fc\u306e\u5ea6\u306bAI\u306e\u30d0\u30a4\u30a2\u30b9\u3084\u30cf\u30eb\u30b7\u30cd\u30fc\u30b7\u30e7\u30f3\u306b\u3046\u3093\u3056\u308a\u3057\u3066\u304d\u3066\u3044\u307e\u3059\u3002
\u305d\u3046\u3057\u3066AI\u99c6\u52d5\u958b\u767a\u306e\u662f\u975e\u306f\u30a8\u30f3\u30b8\u30cb\u30a2\u306e\u4e2d\u3067\u3082\u4e8c\u6975\u5316\u304c\u9032\u3093\u3067\u3044\u3063\u3066\u3057\u307e\u3046\u306e\u3067\u3057\u3087\u3046\u3002<\/p>\n\n\n\n

\u5148\u8f29\u3092\u697d\u3055\u305b\u3088\u3046\uff08\u30a2\u30f3\u30c1\u5148\u8f29\u30d1\u30bf\u30fc\u30f3\uff09<\/h2>\n\n\n\n

\u3053\u306e\u307e\u307e\u3060\u3068\u305d\u306e\u3046\u3061\u5148\u8f29\u306b\u300cAI\u958b\u767a\u7981\u6b62\uff01\u300d\u3068\u8a00\u308f\u308c\u3066\u3057\u307e\u3044\u305d\u3046\u3002
\u672a\u6765\u3057\u304b\u306a\u3044\u79c1\u306eAI\u30c9\u30ea\u30d6\u30f3\u30e9\u30a4\u30d5\u3092\u305d\u3093\u306a\u5c5e\u4eba\u7684\u306a\u8cb0\u3044\u4e8b\u6545\u307f\u305f\u3044\u306a\u3082\u306e\u3067\u596a\u308f\u308c\u3066\u306a\u3093\u3066\u306a\u308b\u3082\u306e\u304b\uff01
\u3044\u3048\u3001\u3059\u307f\u307e\u305b\u3093\u3001\u5148\u8f29\u304c\u75b2\u308c\u3063\u3071\u306a\u3057\u306a\u306e\u306fAI\u304c\u30ec\u30d3\u30e5\u30fc\u5de5\u7a0b\u3092\u5148\u8f29\u306b\u6295\u3052\u3063\u3071\u306a\u3057\u3060\u304b\u3089\u304b\u3082\u3057\u308c\u307e\u305b\u3093\u3002<\/p>\n\n\n\n

\u6700\u8fd1Cursor SDK\u304c\u30ea\u30ea\u30fc\u30b9\u3055\u308c\u3001\u4eca\u306f\uff085\/26\u307e\u3067\uff0990%\u30aa\u30d5\u3067\u5229\u7528\u3067\u304d\u308b\u3053\u3068\u3092\u77e5\u3063\u305f\u79c1\u3002
\u4f55\u304b\u3067\u304d\u308b\u3053\u3068\u306f\u306a\u3044\u304b\u306a\u3068\u8003\u3048\u3001”LLM as a judge”\u306e\u4ed5\u7d44\u307f\u3092\u53d6\u308a\u5165\u308c\u305fAI\u30b3\u30fc\u30c9\u30ec\u30d3\u30e5\u30fc\u3092\u7528\u610f\u3057\u3066\u3001\u5148\u8f29\u306b\u3057\u3063\u304b\u308a\u7652\u3048\u3066\u3082\u3089\u3046\u3053\u3068\u306b\u3057\u307e\u3057\u305f\u3002<\/p>\n\n\n\n

AI\u30b3\u30fc\u30c9\u30ec\u30d3\u30e5\u30fc<\/h2>\n\n\n\n

“LLM as a judge”\u306f\u3053\u3053\u3067\u306f\u8a73\u3057\u304f\u66f8\u304d\u307e\u305b\u3093\u3002
\u7c21\u5358\u306b\u3044\u3046\u3068\u3001\u540c\u3058\u30ec\u30d3\u30e5\u30fc\u4f9d\u983c\u306e\u5185\u5bb9\u3067\u3082\u6c7a\u5b9a\u8ad6\u3092\u6301\u305f\u305a\u306b\u30d9\u30af\u30c8\u30eb\u3067\u63fa\u3089\u3050AI\u306b\u5bfe\u3057\u3066\u3001\u8907\u6570\u30e2\u30c7\u30eb\u3067\u30ec\u30d3\u30e5\u30fc\u3055\u305b\u3066\u3001\u305d\u308c\u3092\u307e\u305f\u5225\u306e\u30e2\u30c7\u30eb\u3067A\/B\u30c6\u30b9\u30c8\u7684\u306b\u8a55\u4fa1\u3055\u305b\u3066\u30ec\u30d3\u30e5\u30fc\u7d50\u679c\u306e\u7cbe\u5ea6\u3092\u3042\u3052\u3088\u3046\u4f5c\u6226\u3001\u3067\u3059\u3002<\/p>\n\n\n\n

\u30ec\u30c4\u30c8\u30e9\u30a4\uff01
(\u5b9f\u969b\u306e\u30b3\u30fc\u30c9\u3068\u306f\u90e8\u5206\u7684\u306b\u5909\u3048\u3066\u3044\u308b\u305f\u3081\u3001\u53c2\u8003\u7528\u3068\u3057\u3066\u304f\u3060\u3055\u3044)<\/p>\n\n\n\n

index.js<\/p>\n\n\n\n

import { Agent } from \"@cursor\/sdk\";\nimport { execSync } from 'child_process';\nimport path from 'path';\nimport fs from 'fs';\n\nconst reviewerPromptPath = path.resolve(process.cwd(), '.reviewer-prompt.txt');\nconst reviewerPrompt = fs.readFileSync(reviewerPromptPath, 'utf-8') + \"\\n\";\nconst judgerPromptPath = path.resolve(process.cwd(), '.judger-prompt.txt');\nconst judgerPrompt = fs.readFileSync(judgerPromptPath, 'utf-8') + \"\\n\";\n\nconst userPrompt = process.argv[2];\nif (!userPrompt) {\n  console.error(\"\u30e6\u30fc\u30b6\u30fc\u30d7\u30ed\u30f3\u30d7\u30c8(\u30ec\u30d3\u30e5\u30fc\u4f9d\u983c\u306e\u5185\u5bb9)\u3092\u5f15\u6570\u306b\u6307\u5b9a\u3057\u3066\u304f\u3060\u3055\u3044\");\n  process.exit(1);\n}\n\n\/\/console.log(judgerPrompt);\n\/\/console.log(reviewerPrompt);\n\nconst reviewer1_model = process.env.REVIEWER1_MODEL ?? 'gemini-3-flash'; \nconst reviewer1_think = process.env.REVIEWER1_THINK ?? 'low'; \nconst reviewer2_model = process.env.REVIEWER1_MODEL ?? 'composer-2'; \nconst reviewer2_think = process.env.REVIEWER1_THINK ?? 'low'; \nconst judger_model = process.env.JUDGER_MODEL ?? 'claude-3-7-sonnet'; \nconst judger_think = process.env.JUDGER_THINK ?? 'high'; \n\n\nconst reviewer1 = await Agent.create({\n  apiKey: process.env.CURSOR_API_KEY,\n  model: { id: reviewer1_model, value: reviewer1_think },\n  local: { cwd: process.cwd() },\n});\n\nconst reviewer2 = await Agent.create({\n  apiKey: process.env.CURSOR_API_KEY,\n  model: { id: reviewer2_model, value: reviewer2_think },\n  local: { cwd: process.cwd() },\n});\n\nconst judge = await Agent.create({\n  apiKey: process.env.CURSOR_API_KEY,\n  model: { id: judger_model, value: judger_think },\n  local: { cwd: process.cwd() },\n});\n\n\nconst [review1, review2] = await Promise.all([\n  (async () => {\n    const run1 = await reviewer1.send(`[\u30b7\u30b9\u30c6\u30e0\u304b\u3089\u306e\u6307\u793a]\\n${reviewerPrompt}\\n[\u30e6\u30fc\u30b6\u30fc\u304b\u3089\u306e\u4f9d\u983c]\\n${userPrompt}`);\n    let result = '';\n    for await (const event1 of run1.stream()) {\n      if (event1.type === 'assistant' && event1.message?.content) {\n        for (const block1 of event1.message.content) {\n          if (block1.type === 'text') {\n            result += block1.text.replace(\/[\\r\\n]+\/g, '');\n          }\n        }\n      }\n    }\n    console.log(reviewer1_model, result);\n    return result;\n  })(),\n\n  (async () => {\n    const run2 = await reviewer2.send(`[\u30b7\u30b9\u30c6\u30e0\u304b\u3089\u306e\u6307\u793a]\\n${reviewerPrompt}\\n[\u30e6\u30fc\u30b6\u30fc\u304b\u3089\u306e\u4f9d\u983c]\\n${userPrompt}`);\n    let result = '';\n    for await (const event2 of run2.stream()) {\n      if (event2.type === 'assistant' && event2.message?.content) {\n        for (const block2 of event2.message.content) {\n          if (block2.type === 'text') {\n            result += block2.text.replace(\/[\\r\\n]+\/g, '');\n          }\n        }\n      }\n    }\n    console.log(reviewer2_model, result);\n    return result;\n  })()\n]);\n\n\nconst run = await judge.send(`[\u30b7\u30b9\u30c6\u30e0\u304b\u3089\u306e\u6307\u793a]\\n${judgerPrompt}\\n[\u30ec\u30d3\u30e5\u30a2\u30fc\u306e\u56de\u7b54]\\n{\"reviewer1\": ${review1}, \"reviewer2\": ${review2}}`);\n\nlet result = '';\nfor await (const event of run.stream()) {\n  if (event.type === 'assistant' && event.message?.content) {\n    for (const block of event.message.content) {\n      if (block.type === 'text') {\n        result += block.text.replace(\/[\\r\\n]+\/g, '');\n      }\n    }\n  }\n}\n\ntry {\n  execSync('glow -', { \n    input: result, \n    stdio: ['pipe', process.stdout, process.stderr],\n    encoding: 'utf-8' \n  });\n} catch (error) {\n  console.log(result);\n}<\/code><\/pre>\n\n\n\n
.reviewer-prompt.txt<\/p>\n\n\n\n
\u3042\u306a\u305f\u306f\u30b3\u30fc\u30c9\u30ec\u30d3\u30e5\u30fc\u306e\u5c02\u9580\u5bb6\u30a8\u30f3\u30b8\u30cb\u30a2\u3067\u3059\u3002\n\u56de\u7b54\u306f\u5fc5\u305a{\"name\":\"\u3042\u306a\u305f\u306e\u30e2\u30c7\u30eb\u540d-\u30d0\u30fc\u30b8\u30e7\u30f3\u540d\", \"status\":\"pass\/fail\/etc\", \"answer\":\"\u3042\u306a\u305f\u306e\u30ec\u30d3\u30e5\u30fc\u7d50\u679c\"}\u306ejson\u5f62\u5f0f\u3067\u8fd4\u3059\u3053\u3068\n\u4f9d\u983c\u5185\u5bb9\u304c\u66d6\u6627\u306a\u5834\u5408\u306f\u30ec\u30d3\u30e5\u30fc\u305b\u305a\u306b\u305d\u306e\u65e8\u3092\u7b54\u3048\u308b\u3053\u3068\n\u5206\u304b\u3089\u306a\u3044\u3053\u3068\u306f\u7d20\u76f4\u306b\u5206\u304b\u3089\u306a\u3044\u3068\u8a00\u3046\u3053\u3068\nstatus\u306f\u4ee5\u4e0b\u306e\u5b9a\u7fa9\u306b\u6cbf\u3063\u3066\u9078\u629e\u3059\u308b\u3053\u3068\n* pass=\u6307\u793a\u3055\u308c\u305f\u5185\u5bb9\u306b\u57fa\u3065\u304f\u30ec\u30d3\u30e5\u30fc\u7d50\u679c\u306b\u554f\u984c\u304c\u306a\u3044\u5834\u5408\n* fail=\u6307\u793a\u3055\u308c\u305f\u5185\u5bb9\u306b\u57fa\u3065\u304f\u30ec\u30d3\u30e5\u30fc\u7d50\u679c\u306b\u554f\u984c\u304c\u3042\u308b\u5834\u5408\n* etc=\u305d\u308c\u4ee5\u5916\u306e\u5834\u5408<\/code><\/pre>\n\n\n\n
.judger-prompt.txt<\/p>\n\n\n\n
\u3042\u306a\u305f\u306fLLM-as-a-Judge\u3092\u7406\u89e3\u3057\u305f\u30b3\u30fc\u30c9\u30ec\u30d3\u30e5\u30fc\u306e\u5c02\u9580\u5bb6\u3067\u3059\u3002\n\u4ee5\u4e0b\u306e\u8907\u6570\u306e\u30ec\u30d3\u30e5\u30a2\u30fc\u306e\u56de\u7b54\u3092\u5143\u306b\u30b9\u30b3\u30a2\u8a55\u4fa1\uff081=\u4f4e\u30015=\u9ad8\uff09\u3092\u884c\u3044\u3001\u305d\u306e\u8a55\u4fa1\u3092\u304a\u3053\u306a\u3063\u305f\u6839\u62e0\u3092\u8a73\u3057\u304f\u8aac\u660e\u3059\u308b\u3053\u3068\n\u306a\u304a\u3001\u30ec\u30d3\u30e5\u30a2\u30fc\u306fjson\u5f62\u5f0f\u3067\u30ec\u30d3\u30e5\u30fc\u7d50\u679c\u3092\u56de\u7b54\u3057\u3066\u3044\u308b\u304c\u3001\u305d\u306e\u4e2d\u306estatus\u306f\u4ee5\u4e0b\u306e\u5b9a\u7fa9\u306b\u6cbf\u3063\u3066\u63a1\u7528\u3055\u308c\u3066\u3044\u308b\n* pass=\u6307\u793a\u3055\u308c\u305f\u5185\u5bb9\u306b\u57fa\u3065\u304f\u30ec\u30d3\u30e5\u30fc\u7d50\u679c\u306b\u554f\u984c\u304c\u306a\u3044\u5834\u5408\n* fail=\u6307\u793a\u3055\u308c\u305f\u5185\u5bb9\u306b\u57fa\u3065\u304f\u30ec\u30d3\u30e5\u30fc\u7d50\u679c\u306b\u554f\u984c\u304c\u3042\u308b\u5834\u5408\n* etc=\u305d\u308c\u4ee5\u5916\u306e\u5834\u5408\n\n\u305d\u306e\u305f\u3081etc\u540c\u58eb\u306e\u6bd4\u8f03\u7b49\u3067\u30b9\u30b3\u30a2\u8a55\u4fa1\u304c\u96e3\u3057\u3044\u5834\u5408\u306f\u3001\u7406\u7531\u3068\u5171\u306b\u305d\u306e\u65e8\u3092\u56de\u7b54\u3059\u308b\u3053\u3068\u3067\u30b9\u30b3\u30a2\u8a55\u4fa1\u3092\u898b\u9001\u308b\u3053\u3068\u304c\u3067\u304d\u308b\n\u307e\u305f\u3001\u51fa\u529b\u306fmd\u5f62\u5f0f\u3067\u304a\u3053\u306a\u3046\u3053\u3068<\/code><\/pre>\n\n\n\n
node index.js \"{\u5177\u4f53\u7684\u306a\u30ec\u30d3\u30e5\u30fc\u4f9d\u983c\u306e\u5185\u5bb9}\" > works\/review.txt<\/code><\/pre>\n\n\n\n
AI\u30de\u30f3\u30bb\u30fc\u306b\u75b2\u308c\u306a\u3044\u5148\u8f29\uff08\u8131\u30fb\u30a2\u30f3\u30c1\u5148\u8f29\u30d1\u30bf\u30fc\u30f3\uff09<\/h2>\n\n\n\n
\u79c1\u3060\u3063\u3066\u3084\u308b\u306e\u3060\u3001\u3068\u3044\u3046\u66f8\u304d\u51fa\u3057<\/p>\n\n\n\n