[{"data":1,"prerenderedAt":490},["ShallowReactive",2],{"article:vibe-coding-discipline":3},{"meta":4,"raw":17,"references":18,"prev":50,"next":60,"body":70,"toc":476},{"slug":5,"title":6,"date":7,"updated":7,"tags":8,"summary":13,"draft":14,"lang":15,"minutes":16},"vibe-coding-discipline","Vibe Coding 的工程纪律：任务、审查与验证怎么分工","2026-10-11",[9,10,11,12],"AI 编程","Claude Code","工程效率","Vibe Coding","对照 OpenAI harness engineering、Claude Code 与生产力研究，解释任务契约、仓库知识、机械约束和独立验收；把审查、返工与实际交付一起计入优化。",false,"zh-CN",6,"AI 能很快生成一段看起来合理的代码，但上线需要另一种能力：确认它在真实环境、边界条件和失败路径中仍然正确。本文讨论工程中的 AI 辅助编程，用“Vibe Coding”这个流行说法作入口，重点是**任务分工、证据和责任**。\n\n本文综合公开工程报告、工具文档和原始生产力研究；建议流程属于工程推导，不把单个团队的报告写成所有项目都能获得的提效倍数。\n\n## 按风险与可验证性分工\n\n| 任务 | 合理分工 | 验证重点 |\n| --- | --- | --- |\n| 文案、样板代码、小范围 UI | AI 起草，快速人工检查 | 实际效果、语义、改动范围 |\n| 业务规则、数据迁移 | 明确约束后起草，领域人员审查 | 边界、旧数据、失败与回滚 |\n| 鉴权、支付、密钥、核心算法 | 人确定不变量和设计，AI 可辅助 | 独立验证、权限、并发、副作用 |\n| 架构和部署决策 | AI 列方案，人依据证据决策 | 运维成本、迁移、可恢复性 |\n\n“核心代码永远自己敲”并非唯一正确分工；更重要的是能否可靠验证，以及谁负责最终结果。数据迁移也不是天然低风险：一个短脚本可能影响全库，不能与文案修改用同样的审查力度。\n\n![AI 编程闭环：任务约束、有限改动、风险审查、真实环境验收](\u002Fdiagrams\u002Fvibe-coding.svg)\n\n*图 1 · 自动化程度取决于风险、范围和验证能力，不能只按代码是否容易生成判断。*\n\n## 给 AI 一份可验收的任务书\n\n任务包含背景、具体失败、范围、约束和验收。例如：\n\n```text\n现象：文章目录在生产静态构建中不能 sticky。\n范围：检查文章页及祖先容器的样式，保留现有视觉布局。\n约束：使用仓库约定的构建命令，不引入新依赖。\n验收：生产产物滚动后目录仍可见，移动端无横向溢出。\n输出：改动原因、相关 diff 和验证结果。\n```\n\n写清事实与猜测：“可能是 overflow”是待验证假设，不是要求模型证明的结论。复杂任务先定位再修改；明确不变量往往比提供大量无关历史更有效。\n\n项目文件如 AGENTS.md 或 CLAUDE.md 保存构建、目录和工作约定，定期删掉过期内容。它们帮助模型工作，不能代替权限和机器检查。\n\n## 审查：优先看最容易隐蔽出错的地方\n\n先看 diff 是否只处理目标问题，再看输入、授权、密钥、并发状态、错误与取消路径。新依赖是否必要、旧数据是否兼容、失败是否能恢复，也应随风险检查。\n\n测试按影响选择：文案与小样式修改通常靠渲染检查；状态、协议和计算逻辑需要有意义的回归。测试先行对可明确描述的 bug 很有效，但不能为了每个低影响修改写一套只镜像实现的测试。\n\n另一个 AI 会话可提供独立审查视角，仍可能同源犯错；不能把“两个模型都说没问题”当作正确证明。已知模式用静态分析，正确性用任务相应的测试和人工判断。\n\n## OpenAI 与 Anthropic：让验收独立于生成\n\n[OpenAI Harness engineering](https:\u002F\u002Fopenai.com\u002Findex\u002Fharness-engineering\u002F)讨论仓库知识、隔离运行环境与机械执行的架构约束。可借鉴的是：给 Agent 可导航的项目知识，把重要不变量写成会执行的检查，让它能在实际环境里验证改动。\n\n[Anthropic 的长任务开发报告](https:\u002F\u002Fwww.anthropic.com\u002Fengineering\u002Fharness-design-long-running-apps)使用 planner、generator 与 evaluator，借助结构化契约推动增量实现与浏览器验收。生成与评审分工有助发现自我评价的盲点，但仍需清晰标准与可靠检查。\n\n报告中更复杂 harness 同时扩大了任务实现范围和执行预算，不能只把展示效果视为同等成本下的严格比较。工程推导：团队实验应控制需求范围、工具版本和验收标准，记录完整时间与费用，再判断规划、独立评审或浏览器验证的净收益。\n\n### 团队怎样把经验写成可执行规范\n\n任务书描述目标、范围和验收；AGENTS.md \u002F CLAUDE.md 提供可导航的项目规则；类型、lint、契约测试和 CI 执行可机械验证的不变量。审查意见与故障复盘回流文档，反复出现且能确定判断的规则再固化为检查。\n\n核心算法、性能关键路径、安全与合规设计由负责人确定不变量和风险边界，AI 可辅助分析和起草；合并责任不能转给工具。新成员通过示例任务、审查解释和失败案例理解规则，组件文档写清接口、恢复和使用边界，避免只留一句“已经测试”。\n\n### UI 验证示例：sticky 与滚动祖先\n\n目录无法 sticky 时，检查所有祖先的计算样式与滚动容器，而不是反复增加 top 或 z-index。[MDN position](https:\u002F\u002Fdeveloper.mozilla.org\u002Fen-US\u002Fdocs\u002FWeb\u002FCSS\u002Fposition)与 [overflow-x](https:\u002F\u002Fdeveloper.mozilla.org\u002Fen-US\u002Fdocs\u002FWeb\u002FCSS\u002Foverflow-x)说明相关 CSS 行为；具体问题仍需用 DOM、计算样式与滚动实测定位。\n\n桌面、手机与生产产物验证覆盖长表格、图片放大和跨页导航。开发模式正常不保证构建后的样式顺序与 hydration 正常，这也是 Agent 的验收环境应贴近交付环境的原因。\n\n## 自动检查必须放在真的会执行的位置\n\nHooks 可以在工具执行或提交环节检查，CI 则验证提交后的统一规则。安装依赖脚本、Git hooks 和 AI 本身也可能修改检查逻辑；“本地有 hook”不保证永远绕不过。重要门槛放进受控 CI 与分支策略，必要时审查规则自身的变更。[Claude Code hooks](https:\u002F\u002Fcode.claude.com\u002Fdocs\u002Fen\u002Fhooks)\n\n把权限、审查和发布边界单独配置，详见 [本地 Agent 安全篇](\u002Farticles\u002Fclaude-code-local-agent-security)。能自动执行的低风险动作在明确范围内推进，高影响提交需要可 review 的结果和相应授权。\n\n## 提效：总时间里必须包括审查与返工\n\n```text\n任务总耗时 = 定义任务 + 生成或编码 + 审查\n           + 验证 + 返工 + 上线后的修复\n```\n\n按任务类型记录真实耗时、通过率、缺陷和工具费用，同类任务作比较。不能只计“AI 写了五分钟”，忽略工程师审查两小时；也不能因一个复杂任务负收益就否定所有辅助用途。\n\nMETR 对早期 2025 工具、熟悉大型开源项目的开发者做过随机实验，观察到该样本使用 AI 后完成时间增加约 19%。这有明确的时间、工具和样本边界，既不证明 AI 永远减速，也提醒我们感知提效需要实测。[研究原文](https:\u002F\u002Fmetr.org\u002Fblog\u002F2025-07-10-early-2025-ai-experienced-os-dev-study\u002F)\n\n工具更新很快，过去的倍数不能承诺给今天的项目。最有用的指标是团队在当前工具与流程下的交付质量和总时间。\n\n## 落地检查\n\n1. 按风险与可验证性分工，先写目标、约束和验收。\n2. 事实与假设分开，定位后做范围明确的改动。\n3. 审查优先看权限、并发、错误、副作用和旧数据。\n4. 验证按影响选择，UI 在实际构建产物检查。\n5. 机器门槛有受控执行位置，AI 审查不代替验收。\n6. 测总交付时间、返工和质量，持续调整分工。\n\n## 参考资料\n\n- [Claude Code — Best practices](https:\u002F\u002Fcode.claude.com\u002Fdocs\u002Fen\u002Fbest-practices)\n- [Claude Code — Hooks](https:\u002F\u002Fcode.claude.com\u002Fdocs\u002Fen\u002Fhooks)\n- [METR — Early-2025 developer productivity study](https:\u002F\u002Fmetr.org\u002Fblog\u002F2025-07-10-early-2025-ai-experienced-os-dev-study\u002F)\n- [MDN — position 与 sticky](https:\u002F\u002Fdeveloper.mozilla.org\u002Fen-US\u002Fdocs\u002FWeb\u002FCSS\u002Fposition)\n- [MDN — overflow-x](https:\u002F\u002Fdeveloper.mozilla.org\u002Fen-US\u002Fdocs\u002FWeb\u002FCSS\u002Foverflow-x)\n- [OpenAI — Harness engineering](https:\u002F\u002Fopenai.com\u002Findex\u002Fharness-engineering\u002F)\n- [Anthropic — Harness design for long-running application development](https:\u002F\u002Fwww.anthropic.com\u002Fengineering\u002Fharness-design-long-running-apps)",[19,24,28,33,37,41,46],{"title":20,"url":21,"type":22,"note":23,"checked":7},"Claude Code — Best practices","https:\u002F\u002Fcode.claude.com\u002Fdocs\u002Fen\u002Fbest-practices","官方文档","Claude Code 的上下文、任务与验证建议；实际效率需计入审查和返工。",{"title":25,"url":26,"type":22,"note":27,"checked":7},"Claude Code — Hooks","https:\u002F\u002Fcode.claude.com\u002Fdocs\u002Fen\u002Fhooks","hooks 的执行位置与配置；关键发布门槛仍需受控 CI。",{"title":29,"url":30,"type":31,"note":32,"checked":7},"METR — Early-2025 developer productivity study","https:\u002F\u002Fmetr.org\u002Fblog\u002F2025-07-10-early-2025-ai-experienced-os-dev-study\u002F","研究报告","特定开发者、任务与早期 2025 工具的随机生产力研究；不构成当代工具的统一效率结论。",{"title":34,"url":35,"type":22,"note":36,"checked":7},"MDN — position 与 sticky","https:\u002F\u002Fdeveloper.mozilla.org\u002Fen-US\u002Fdocs\u002FWeb\u002FCSS\u002Fposition","sticky 定位与滚动祖先的 CSS 行为；具体页面问题需实测。",{"title":38,"url":39,"type":22,"note":40,"checked":7},"MDN — overflow-x","https:\u002F\u002Fdeveloper.mozilla.org\u002Fen-US\u002Fdocs\u002FWeb\u002FCSS\u002Foverflow-x","overflow 计算值与滚动行为；不能仅看某个节点的声明判断布局。",{"title":42,"url":43,"type":44,"note":45,"checked":7},"OpenAI — Harness engineering","https:\u002F\u002Fopenai.com\u002Findex\u002Fharness-engineering\u002F","工程报告","仓库知识、隔离验收环境与机械约束的工程实践；案例效率不能视为普遍承诺。",{"title":47,"url":48,"type":44,"note":49,"checked":7},"Anthropic — Harness design for long-running application development","https:\u002F\u002Fwww.anthropic.com\u002Fengineering\u002Fharness-design-long-running-apps","规划、生成与独立评审的长任务实验；任务范围与预算差异限制直接比较。",{"slug":51,"title":52,"date":7,"updated":7,"tags":53,"summary":58,"draft":14,"lang":15,"minutes":59},"prompt-context-engineering","Prompt 与上下文工程：OpenAI、Anthropic 的方法，版本发布与优化实验",[54,55,56,57],"Prompt","上下文工程","OpenAI","Anthropic","从官方提示实践与长上下文研究出发，解释指令、证据、历史和工具如何组合；用版本化发布包、受控实验、影子验证与灰度回滚，把 prompt 优化纳入工程流程。",7,{"slug":61,"title":62,"date":63,"updated":7,"tags":64,"summary":69,"draft":14,"lang":15,"minutes":16},"minors-ai-safety-engineering","面向未成年人的 AI 产品：合规要求与安全工程","2026-09-15",[65,66,67,68],"安全合规","未成年人保护","内容安全","LLM","核对法规适用范围与年龄边界，补齐生成内容标识、授权和删除链路；把数据保护、流式审核、风险分流、防依赖与人工处置落实到工程。",{"type":71,"children":72},"root",[73,89,94,100,205,210,219,228,234,239,252,257,262,268,273,278,283,289,302,313,318,324,329,334,340,361,366,371,382,395,401,410,415,426,431,436,471],{"type":74,"tag":75,"props":76,"children":77},"element","p",{},[78,81,87],{"type":79,"value":80},"text","AI 能很快生成一段看起来合理的代码，但上线需要另一种能力：确认它在真实环境、边界条件和失败路径中仍然正确。本文讨论工程中的 AI 辅助编程，用“Vibe Coding”这个流行说法作入口，重点是",{"type":74,"tag":82,"props":83,"children":84},"strong",{},[85],{"type":79,"value":86},"任务分工、证据和责任",{"type":79,"value":88},"。",{"type":74,"tag":75,"props":90,"children":91},{},[92],{"type":79,"value":93},"本文综合公开工程报告、工具文档和原始生产力研究；建议流程属于工程推导，不把单个团队的报告写成所有项目都能获得的提效倍数。",{"type":74,"tag":95,"props":96,"children":98},"h2",{"id":97},"按风险与可验证性分工",[99],{"type":79,"value":97},{"type":74,"tag":101,"props":102,"children":103},"table",{},[104,128],{"type":74,"tag":105,"props":106,"children":107},"thead",{},[108],{"type":74,"tag":109,"props":110,"children":111},"tr",{},[112,118,123],{"type":74,"tag":113,"props":114,"children":115},"th",{},[116],{"type":79,"value":117},"任务",{"type":74,"tag":113,"props":119,"children":120},{},[121],{"type":79,"value":122},"合理分工",{"type":74,"tag":113,"props":124,"children":125},{},[126],{"type":79,"value":127},"验证重点",{"type":74,"tag":129,"props":130,"children":131},"tbody",{},[132,151,169,187],{"type":74,"tag":109,"props":133,"children":134},{},[135,141,146],{"type":74,"tag":136,"props":137,"children":138},"td",{},[139],{"type":79,"value":140},"文案、样板代码、小范围 UI",{"type":74,"tag":136,"props":142,"children":143},{},[144],{"type":79,"value":145},"AI 起草，快速人工检查",{"type":74,"tag":136,"props":147,"children":148},{},[149],{"type":79,"value":150},"实际效果、语义、改动范围",{"type":74,"tag":109,"props":152,"children":153},{},[154,159,164],{"type":74,"tag":136,"props":155,"children":156},{},[157],{"type":79,"value":158},"业务规则、数据迁移",{"type":74,"tag":136,"props":160,"children":161},{},[162],{"type":79,"value":163},"明确约束后起草，领域人员审查",{"type":74,"tag":136,"props":165,"children":166},{},[167],{"type":79,"value":168},"边界、旧数据、失败与回滚",{"type":74,"tag":109,"props":170,"children":171},{},[172,177,182],{"type":74,"tag":136,"props":173,"children":174},{},[175],{"type":79,"value":176},"鉴权、支付、密钥、核心算法",{"type":74,"tag":136,"props":178,"children":179},{},[180],{"type":79,"value":181},"人确定不变量和设计，AI 可辅助",{"type":74,"tag":136,"props":183,"children":184},{},[185],{"type":79,"value":186},"独立验证、权限、并发、副作用",{"type":74,"tag":109,"props":188,"children":189},{},[190,195,200],{"type":74,"tag":136,"props":191,"children":192},{},[193],{"type":79,"value":194},"架构和部署决策",{"type":74,"tag":136,"props":196,"children":197},{},[198],{"type":79,"value":199},"AI 列方案，人依据证据决策",{"type":74,"tag":136,"props":201,"children":202},{},[203],{"type":79,"value":204},"运维成本、迁移、可恢复性",{"type":74,"tag":75,"props":206,"children":207},{},[208],{"type":79,"value":209},"“核心代码永远自己敲”并非唯一正确分工；更重要的是能否可靠验证，以及谁负责最终结果。数据迁移也不是天然低风险：一个短脚本可能影响全库，不能与文案修改用同样的审查力度。",{"type":74,"tag":75,"props":211,"children":212},{},[213],{"type":74,"tag":214,"props":215,"children":218},"img",{"alt":216,"src":217},"AI 编程闭环：任务约束、有限改动、风险审查、真实环境验收","\u002Fdiagrams\u002Fvibe-coding.svg",[],{"type":74,"tag":75,"props":220,"children":221},{},[222],{"type":74,"tag":223,"props":224,"children":225},"em",{},[226],{"type":79,"value":227},"图 1 · 自动化程度取决于风险、范围和验证能力，不能只按代码是否容易生成判断。",{"type":74,"tag":95,"props":229,"children":231},{"id":230},"给-ai-一份可验收的任务书",[232],{"type":79,"value":233},"给 AI 一份可验收的任务书",{"type":74,"tag":75,"props":235,"children":236},{},[237],{"type":79,"value":238},"任务包含背景、具体失败、范围、约束和验收。例如：",{"type":74,"tag":240,"props":241,"children":246},"pre",{"className":242,"code":244,"language":79,"meta":245},[243],"language-text","现象：文章目录在生产静态构建中不能 sticky。\n范围：检查文章页及祖先容器的样式，保留现有视觉布局。\n约束：使用仓库约定的构建命令，不引入新依赖。\n验收：生产产物滚动后目录仍可见，移动端无横向溢出。\n输出：改动原因、相关 diff 和验证结果。\n","",[247],{"type":74,"tag":248,"props":249,"children":250},"code",{"__ignoreMap":245},[251],{"type":79,"value":244},{"type":74,"tag":75,"props":253,"children":254},{},[255],{"type":79,"value":256},"写清事实与猜测：“可能是 overflow”是待验证假设，不是要求模型证明的结论。复杂任务先定位再修改；明确不变量往往比提供大量无关历史更有效。",{"type":74,"tag":75,"props":258,"children":259},{},[260],{"type":79,"value":261},"项目文件如 AGENTS.md 或 CLAUDE.md 保存构建、目录和工作约定，定期删掉过期内容。它们帮助模型工作，不能代替权限和机器检查。",{"type":74,"tag":95,"props":263,"children":265},{"id":264},"审查优先看最容易隐蔽出错的地方",[266],{"type":79,"value":267},"审查：优先看最容易隐蔽出错的地方",{"type":74,"tag":75,"props":269,"children":270},{},[271],{"type":79,"value":272},"先看 diff 是否只处理目标问题，再看输入、授权、密钥、并发状态、错误与取消路径。新依赖是否必要、旧数据是否兼容、失败是否能恢复，也应随风险检查。",{"type":74,"tag":75,"props":274,"children":275},{},[276],{"type":79,"value":277},"测试按影响选择：文案与小样式修改通常靠渲染检查；状态、协议和计算逻辑需要有意义的回归。测试先行对可明确描述的 bug 很有效，但不能为了每个低影响修改写一套只镜像实现的测试。",{"type":74,"tag":75,"props":279,"children":280},{},[281],{"type":79,"value":282},"另一个 AI 会话可提供独立审查视角，仍可能同源犯错；不能把“两个模型都说没问题”当作正确证明。已知模式用静态分析，正确性用任务相应的测试和人工判断。",{"type":74,"tag":95,"props":284,"children":286},{"id":285},"openai-与-anthropic让验收独立于生成",[287],{"type":79,"value":288},"OpenAI 与 Anthropic：让验收独立于生成",{"type":74,"tag":75,"props":290,"children":291},{},[292,300],{"type":74,"tag":293,"props":294,"children":297},"a",{"href":43,"rel":295},[296],"nofollow",[298],{"type":79,"value":299},"OpenAI Harness engineering",{"type":79,"value":301},"讨论仓库知识、隔离运行环境与机械执行的架构约束。可借鉴的是：给 Agent 可导航的项目知识，把重要不变量写成会执行的检查，让它能在实际环境里验证改动。",{"type":74,"tag":75,"props":303,"children":304},{},[305,311],{"type":74,"tag":293,"props":306,"children":308},{"href":48,"rel":307},[296],[309],{"type":79,"value":310},"Anthropic 的长任务开发报告",{"type":79,"value":312},"使用 planner、generator 与 evaluator，借助结构化契约推动增量实现与浏览器验收。生成与评审分工有助发现自我评价的盲点，但仍需清晰标准与可靠检查。",{"type":74,"tag":75,"props":314,"children":315},{},[316],{"type":79,"value":317},"报告中更复杂 harness 同时扩大了任务实现范围和执行预算，不能只把展示效果视为同等成本下的严格比较。工程推导：团队实验应控制需求范围、工具版本和验收标准，记录完整时间与费用，再判断规划、独立评审或浏览器验证的净收益。",{"type":74,"tag":319,"props":320,"children":322},"h3",{"id":321},"团队怎样把经验写成可执行规范",[323],{"type":79,"value":321},{"type":74,"tag":75,"props":325,"children":326},{},[327],{"type":79,"value":328},"任务书描述目标、范围和验收；AGENTS.md \u002F CLAUDE.md 提供可导航的项目规则；类型、lint、契约测试和 CI 执行可机械验证的不变量。审查意见与故障复盘回流文档，反复出现且能确定判断的规则再固化为检查。",{"type":74,"tag":75,"props":330,"children":331},{},[332],{"type":79,"value":333},"核心算法、性能关键路径、安全与合规设计由负责人确定不变量和风险边界，AI 可辅助分析和起草；合并责任不能转给工具。新成员通过示例任务、审查解释和失败案例理解规则，组件文档写清接口、恢复和使用边界，避免只留一句“已经测试”。",{"type":74,"tag":319,"props":335,"children":337},{"id":336},"ui-验证示例sticky-与滚动祖先",[338],{"type":79,"value":339},"UI 验证示例：sticky 与滚动祖先",{"type":74,"tag":75,"props":341,"children":342},{},[343,345,351,353,359],{"type":79,"value":344},"目录无法 sticky 时，检查所有祖先的计算样式与滚动容器，而不是反复增加 top 或 z-index。",{"type":74,"tag":293,"props":346,"children":348},{"href":35,"rel":347},[296],[349],{"type":79,"value":350},"MDN position",{"type":79,"value":352},"与 ",{"type":74,"tag":293,"props":354,"children":356},{"href":39,"rel":355},[296],[357],{"type":79,"value":358},"overflow-x",{"type":79,"value":360},"说明相关 CSS 行为；具体问题仍需用 DOM、计算样式与滚动实测定位。",{"type":74,"tag":75,"props":362,"children":363},{},[364],{"type":79,"value":365},"桌面、手机与生产产物验证覆盖长表格、图片放大和跨页导航。开发模式正常不保证构建后的样式顺序与 hydration 正常，这也是 Agent 的验收环境应贴近交付环境的原因。",{"type":74,"tag":95,"props":367,"children":369},{"id":368},"自动检查必须放在真的会执行的位置",[370],{"type":79,"value":368},{"type":74,"tag":75,"props":372,"children":373},{},[374,376],{"type":79,"value":375},"Hooks 可以在工具执行或提交环节检查，CI 则验证提交后的统一规则。安装依赖脚本、Git hooks 和 AI 本身也可能修改检查逻辑；“本地有 hook”不保证永远绕不过。重要门槛放进受控 CI 与分支策略，必要时审查规则自身的变更。",{"type":74,"tag":293,"props":377,"children":379},{"href":26,"rel":378},[296],[380],{"type":79,"value":381},"Claude Code hooks",{"type":74,"tag":75,"props":383,"children":384},{},[385,387,393],{"type":79,"value":386},"把权限、审查和发布边界单独配置，详见 ",{"type":74,"tag":293,"props":388,"children":390},{"href":389},"\u002Farticles\u002Fclaude-code-local-agent-security",[391],{"type":79,"value":392},"本地 Agent 安全篇",{"type":79,"value":394},"。能自动执行的低风险动作在明确范围内推进，高影响提交需要可 review 的结果和相应授权。",{"type":74,"tag":95,"props":396,"children":398},{"id":397},"提效总时间里必须包括审查与返工",[399],{"type":79,"value":400},"提效：总时间里必须包括审查与返工",{"type":74,"tag":240,"props":402,"children":405},{"className":403,"code":404,"language":79,"meta":245},[243],"任务总耗时 = 定义任务 + 生成或编码 + 审查\n           + 验证 + 返工 + 上线后的修复\n",[406],{"type":74,"tag":248,"props":407,"children":408},{"__ignoreMap":245},[409],{"type":79,"value":404},{"type":74,"tag":75,"props":411,"children":412},{},[413],{"type":79,"value":414},"按任务类型记录真实耗时、通过率、缺陷和工具费用，同类任务作比较。不能只计“AI 写了五分钟”，忽略工程师审查两小时；也不能因一个复杂任务负收益就否定所有辅助用途。",{"type":74,"tag":75,"props":416,"children":417},{},[418,420],{"type":79,"value":419},"METR 对早期 2025 工具、熟悉大型开源项目的开发者做过随机实验，观察到该样本使用 AI 后完成时间增加约 19%。这有明确的时间、工具和样本边界，既不证明 AI 永远减速，也提醒我们感知提效需要实测。",{"type":74,"tag":293,"props":421,"children":423},{"href":30,"rel":422},[296],[424],{"type":79,"value":425},"研究原文",{"type":74,"tag":75,"props":427,"children":428},{},[429],{"type":79,"value":430},"工具更新很快，过去的倍数不能承诺给今天的项目。最有用的指标是团队在当前工具与流程下的交付质量和总时间。",{"type":74,"tag":95,"props":432,"children":434},{"id":433},"落地检查",[435],{"type":79,"value":433},{"type":74,"tag":437,"props":438,"children":439},"ol",{},[440,446,451,456,461,466],{"type":74,"tag":441,"props":442,"children":443},"li",{},[444],{"type":79,"value":445},"按风险与可验证性分工，先写目标、约束和验收。",{"type":74,"tag":441,"props":447,"children":448},{},[449],{"type":79,"value":450},"事实与假设分开，定位后做范围明确的改动。",{"type":74,"tag":441,"props":452,"children":453},{},[454],{"type":79,"value":455},"审查优先看权限、并发、错误、副作用和旧数据。",{"type":74,"tag":441,"props":457,"children":458},{},[459],{"type":79,"value":460},"验证按影响选择，UI 在实际构建产物检查。",{"type":74,"tag":441,"props":462,"children":463},{},[464],{"type":79,"value":465},"机器门槛有受控执行位置，AI 审查不代替验收。",{"type":74,"tag":441,"props":467,"children":468},{},[469],{"type":79,"value":470},"测总交付时间、返工和质量，持续调整分工。",{"type":74,"tag":95,"props":472,"children":474},{"id":473},"参考资料",[475],{"type":79,"value":473},[477,479,480,481,486,487,488,489],{"id":97,"depth":478,"text":97},2,{"id":230,"depth":478,"text":233},{"id":264,"depth":478,"text":267},{"id":285,"depth":478,"text":288,"children":482},[483,485],{"id":321,"depth":484,"text":321},3,{"id":336,"depth":484,"text":339},{"id":368,"depth":478,"text":368},{"id":397,"depth":478,"text":400},{"id":433,"depth":478,"text":433},{"id":473,"depth":478,"text":473},1792400816298]