最近看到好多问题，我自己的经历，朋友问过来的问题，都在指向同一个问题。

在我们的指使下，AI 可以生成非常多的代码，但随着代码增加，灾难就出现了。

一个常见的做法是，给 Prompt，让 AI 重新做。抽卡，总能解决问题。

小项目、小产品是可以的，再复杂的呢？等待 AI 更强么？

朋友找到我的时候，我 review 了整个项目，表面的问题好解决，但过程中积累的问题必须要清理、重构了。

这些问题，要趁早解决，开始可能就是一行代码的问题，但积累起来，可能只能重写了。

A new service category is quietly emerging in tech: Vibe Coding cleanup. What started as LinkedIn jokes about "fixing AI messes" has become a real business opportunity. The harsh reality nobody wants to admit: most AI-generated code is production-unready, and companies are desperately hiring specialists to fix it before their technical debt spirals out of control.

科技行业正悄然兴起一个新的服务类别：Vibe Coding清理服务。最初在LinkedIn上关于"修复AI烂摊子"的调侃，如今已经成为真正的商业机会。没人愿意承认的残酷现实是：大多数AI生成的代码都无法直接投入生产，公司正拼命雇佣专家来修复这些代码，以免技术债务失控。

Vibe Coding爆发

When Andrej Karpathy coined "vibe coding" in early 2025, he perfectly captured how developers now work: chatting with AI to generate entire functions instead of writing them. The approach promises 10x productivity gains through natural language programming[1]. GitHub reports that 92% of developers now use AI coding tools[2], with Copilot alone generating billions of lines of code monthly.

当Andrej Karpathy在2025年初创造"vibe coding"这个词时，他完美地概括了开发者现在的工作方式：与AI对话来生成整个函数，而不是自己编写。这种方法承诺通过自然语言编程实现10倍的生产力提升[1]。GitHub报告称92%的开发者现在使用AI编程工具[2]，仅Copilot每月就生成数十亿行代码。

But there's a problem nobody talks about at conferences. GitClear's analysis of 150 million lines of code reveals AI assistance correlates with 41% more code churn[3] - code that gets reverted or rewritten within two weeks. Stanford researchers found that developers using AI assistants produce significantly less secure code while believing it's more secure[4]. The tools amplify bad practices: no input validation, outdated dependencies, and architectural decisions that make senior engineers weep.

但有一个问题在会议上无人提及。GitClear对1.5亿行代码的分析显示，AI辅助与41%的代码流失率相关[3]——这些代码在两周内被回滚或重写。斯坦福研究人员发现，使用AI助手的开发者产生的代码安全性明显较低，但他们却认为更安全[4]。这些工具放大了不良实践：没有输入验证、过时的依赖项，以及让资深工程师痛哭的架构决策。

清理经济正在发生

404 Media's investigation reveals developers are building entire careers around fixing AI-generated code[5]. Hamid Siddiqi manages 15-20 cleanup projects simultaneously, charging premium rates to untangle what he calls "AI spaghetti" - inconsistent interfaces, redundant functions, and business logic that makes no sense. Software consultancy Ulam Labs now advertises "Vibe Coding cleanup" as a core service[6].

404 Media的调查显示，开发者正在围绕修复AI生成的代码构建整个职业生涯[5]。Hamid Siddiqi同时管理着15-20个清理项目，收取高额费用来解开他称之为"AI意大利面条"的代码——不一致的接口、冗余的函数，以及毫无意义的业务逻辑。软件咨询公司Ulam Labs现在将"Vibe Coding清理"作为核心服务进行宣传[6]。

The demand is so high that VibeCodeFixers.com launched as a dedicated marketplace. Within weeks, 300 specialists signed up and dozens of projects were matched. Founder Swatantra Sohni describes a typical client: "They burned through $5,000 in OpenAI credits, have a half-working prototype they're emotionally attached to, and need it production-ready yesterday." TechCrunch reports that 25% of Y Combinator's current startup cohort has codebases that are 95% AI-generated[7], highlighting the massive scale of this trend across Silicon Valley.

需求如此之高，以至于VibeCodeFixers.com作为专门的市场平台推出。几周内，300名专家注册，数十个项目得到匹配。创始人Swatantra Sohni描述了典型客户："他们烧掉了5000美元的OpenAI积分，有一个半成品原型，对此有感情依恋，需要昨天就让它准备好投产。"TechCrunch报告称，Y Combinator当前创业公司队列中25%的代码库是95% AI生成的[7]，突显了这一趋势在硅谷的巨大规模。

为什么AI代码在规模化的时候会失败

The fundamental issue isn't that AI writes bad code - it's that it writes locally optimized code without understanding system context. Stack Overflow's analysis shows AI excels at small, isolated tasks but fails at architectural decisions[8]. Every prompt creates technical debt: inconsistent patterns, duplicated logic, and security holes that automated scanners miss.

根本问题不是AI写出糟糕的代码——而是它在不理解系统上下文的情况下编写局部优化的代码。Stack Overflow的分析显示，AI在小型、独立任务上表现出色，但在架构决策上失败[8]。每个提示都会产生技术债务：不一致的模式、重复的逻辑，以及自动扫描器遗漏的安全漏洞。

Georgetown University research shows that at least 48% of AI-generated code contains security vulnerabilities[9]. The tools leak secrets into code, suggest deprecated libraries, and create race conditions that only appear under load. Worse, developers often don't understand the generated code well enough to spot these issues. Thoughtworks warns this creates "competency debt" - teams lose the ability to maintain their own systems[10] as they become dependent on AI-generated code they don't fully understand.

乔治城大学的研究显示，至少48%的AI生成代码包含安全漏洞[9]。这些工具会将机密信息泄露到代码中，建议使用已弃用的库，并创建只在负载下才出现的竞态条件。更糟糕的是，开发者往往不够了解生成的代码，无法发现这些问题。Thoughtworks警告这会产生"能力债务"——团队失去维护自己系统的能力[10]，因为他们依赖于自己并不完全理解的AI生成代码。

市场机会

The Vibe Coding cleanup market is growing rapidly, though exact numbers are hard to pin down. What we know: Gartner predicts 75% of enterprise software engineers will use AI code assistants by 2028[11]. If even a fraction of those projects need cleanup - and current data suggests most will - we're looking at a massive emerging market.

Vibe Coding清理市场正在快速增长，尽管确切数字难以确定。我们知道的是：Gartner预测到2028年75%的企业软件工程师将使用AI代码助手[11]。如果这些项目中哪怕只有一小部分需要清理——而当前数据表明大多数都需要——我们面对的将是一个巨大的新兴市场。

The economics are compelling. Startups save weeks getting to MVP with Vibe Coding, then spend comparable time and budget on cleanup. But that's still faster than traditional development. The specialists who can efficiently refactor AI messes command $200-400/hour rates. Some are building productized services: fixed-price cleanup packages, AI code audits, and "vibe-to-production" pipelines.

经济效益很有吸引力。初创公司通过Vibe Coding节省数周时间达到MVP，然后花费相当的时间和预算进行清理。但这仍然比传统开发更快。能够高效重构AI烂摊子的专家收费200-400美元/小时。一些人正在构建产品化服务：固定价格的清理包、AI代码审计，以及"从vibe到生产"的流水线。

Thoughtworks reports that refactoring activity has declined while code churn increases with AI assistance[12], with most AI-assisted projects requiring significant cleanup before production. Multiple consultancies are now hiring specifically for "AI code remediation" roles. The market is real, growing, and largely untapped.

Thoughtworks报告称，在AI辅助下，重构活动减少而代码流失增加[12]，大多数AI辅助项目在投产前需要大量清理。多家咨询公司现在专门招聘"AI代码修复"角色。这个市场是真实的、增长的，而且基本未被开发。

这对工程实践的影响

We're witnessing a fundamental shift in how software gets built. AI handles the initial implementation, humans handle architecture, testing, and cleanup. It's not the future we expected, but it's the one we're getting.

我们正在见证软件构建方式的根本性转变。AI处理初始实现，人类处理架构、测试和清理。这不是我们期望的未来，但这是我们正在得到的未来。

Gergely Orosz argues AI tools are like "very eager junior developers" - they write code quickly but need constant supervision[13]. The difference is that AI juniors never become seniors. They'll always need cleanup specialists.

Gergely Orosz认为AI工具就像"非常积极的初级开发者"——他们写代码很快但需要持续监督[13]。不同之处在于AI初级开发者永远不会成为高级开发者。他们永远需要清理专家。

This creates interesting career paths. Junior developers who master Vibe Coding cleanup can command senior salaries within two years. Senior engineers who understand both AI capabilities and limitations become invaluable. Companies that build robust cleanup processes gain competitive advantage.

这创造了有趣的职业道路。掌握Vibe Coding清理的初级开发者可以在两年内获得高级薪资。既了解AI能力又了解其局限性的高级工程师变得无价。建立强大清理流程的公司获得竞争优势。

我们的态度

At Donado Labs, we've cleaned up enough vibe-coded disasters to recognize the pattern. AI acceleration works, but only with professional cleanup built into the process. We use AI for prototyping and routine tasks, but architecture and critical logic remain human-written. Our "Vibe to Production" service takes AI prototypes and makes them enterprise-ready: proper testing, security hardening, and documentation that won't make your successor cry.

在Donado Labs，我们清理了足够多的vibe-coded灾难，已经能够识别模式。AI加速是有效的，但只有在流程中内置专业清理时才有效。我们使用AI进行原型设计和常规任务，但架构和关键逻辑仍然由人类编写。我们的"从Vibe到生产"服务将AI原型转化为企业级：适当的测试、安全加固，以及不会让你的继任者哭泣的文档。

The companies succeeding with AI coding aren't the ones using it most - they're the ones using it smartly. They prototype with AI, then invest in cleanup before technical debt compounds. They treat Vibe Coding like any other tool: powerful but dangerous without expertise.

在AI编程方面成功的公司不是使用最多的公司——而是明智使用的公司。他们用AI做原型，然后在技术债务复合之前投资清理。他们将Vibe Coding视为任何其他工具：强大但没有专业知识就很危险。

Next time someone claims AI will replace programmers, ask them who's going to clean up the code. That's where the real opportunity lies.

下次有人声称AI将取代程序员时，问问他们谁来清理代码。这就是真正的机会所在。