{"id":2378,"date":"2026-04-20T15:50:06","date_gmt":"2026-04-20T15:50:06","guid":{"rendered":"https:\/\/deepinsightai.io\/?p=2378"},"modified":"2026-04-20T15:50:07","modified_gmt":"2026-04-20T15:50:07","slug":"how-to-properly-do-vibe-coding","status":"publish","type":"post","link":"https:\/\/deepinsightai.io\/ko\/how-to-properly-do-vibe-coding\/","title":{"rendered":"How to properly do Vibe Coding? This is a masterclass from the head of programming agents at Anthropic"},"content":{"rendered":"

If you break your hand, wear a cast for two months, but work cannot stop, what should a programmer do? Erik Schluntz, a researcher at Anthropic and co-author of Building Effective Agents<\/em>, gives an answer: hand everything over to Claude<\/a>.<\/p>\n\n\n\n

Today, as AI is forcefully reshaping the rules of the software industry, Vibe Coding has become an unavoidable question for companies that want to multiply productivity.<\/p>\n\n\n\n

A few months ago, Schluntz stepped forward with his unusual experience of being forced into \u201cfully automated work,\u201d and discussed a somewhat controversial topic: how to responsibly practice Vibe Coding in production environments.<\/p>\n\n\n\n

This talk is full of practical insights, and in recent days it has gone viral again on X. A user named Movez even praised it as better than 100 paid courses.<\/p>\n\n\n\n

In the spirit of Vibe Coding, we also used AI to organize Schluntz\u2019s talk.<\/p>\n\n\n\n

If you would like to watch this excellent talk, you can click here to access it<\/a>.<\/p>\n\n\n\n

\"comment<\/figure>\n\n\n\n

What Is Vibe Coding<\/h2>\n\n\n\n
\"What<\/figure>\n\n\n\n

Many people equate heavy use of AI tools like Cursor or Copilot for generating code with Vibe Coding. That is not accurate. As long as the developer maintains a tight feedback loop with the model\u2014reviewing and modifying code line by line\u2014it cannot be called true \u201cvibe.\u201d<\/p>\n\n\n\n

Andre Karpathy provided a more precise definition: \u201cfully immerse yourself in the vibe, embrace the exponential growth of technology, and completely forget that the code even exists.\u201d<\/p>\n\n\n\n

This working style significantly lowers the barrier to development, allowing people without an engineering background to independently build complete applications. However, in the past, successful cases of this approach were mostly limited to personal games or low-risk projects. Once non-professionals bring this approach into real production environments, it often leads to out-of-control situations such as exhausting API quotas, bypassing subscription validation, or arbitrarily modifying databases.<\/p>\n\n\n\n

Why Vibe Coding Matters: The Logic of Exponential Growth<\/h2>\n\n\n\n

If high-risk business environments contain uncontrollable factors, why should we still push this technology forward? The core driving force lies in the \u201cexponential growth\u201d of AI capabilities.<\/p>\n\n\n\n

Currently, the length of tasks that AI can independently handle doubles approximately every 7 months. Today, AI can reliably complete coding tasks that take about 1 hour, and developers still have the energy to review them line by line. But by next year or the year after, when AI can generate code equivalent to 1 day or even 1 week of human work in one go, if we still insist on traditional synchronous review and modification, human engineers will inevitably become the bottleneck in the explosion of computing power.<\/p>\n\n\n\n

\"The<\/figure>\n\n\n\n

We can refer to the history of compilers. Early developers did not trust compilers and would still check the underlying assembly code. As system scale grew, developers had to learn to trust higher-level abstractions. Looking ahead, the entire software engineering field needs to consider in advance: how to safely and responsibly accept systems directly generated by large models in production environments.<\/p>\n\n\n\n

How to Apply Vibe Coding in Production<\/h2>\n\n\n\n

The core principle of practicing Vibe Coding in production is: forget the existence of code, but always focus on the existence of the product.<\/p>\n\n\n\n

In modern enterprise management, CTOs rely on acceptance tests to manage technical experts, product managers validate features by experiencing the product, and CEOs audit financial models through key data slices. None of them dive into the lowest-level execution details. Software engineers also need to establish similar abstraction layers that can be verified without reading underlying code.<\/p>\n\n\n\n

The key is to find abstraction layers that you can verify.<\/p>\n\n\n\n

However, current AI coding has a tricky technical obstacle: technical debt. At present, apart from reading the entire source code, it is extremely difficult to measure or verify technical debt through systematic methods.<\/p>\n\n\n\n

Based on this, Erik Schluntz proposes focusing on \u201cleaf nodes\u201d in the codebase. These nodes refer to terminal functions or auxiliary components that are not depended on by any other modules. In these areas, even if some technical debt is introduced, it is acceptable, because they change infrequently and do not block future module construction.<\/p>\n\n\n\n

In contrast, for the core trunk and underlying architecture of the system, engineers still need to deeply understand and strictly protect scalability.<\/p>\n\n\n\n

It is worth noting that as model capabilities improve, the level of code we can trust AI to handle is extending downward. In recent internal tests at Anthropic, the success rate of AI generating high-quality architectures is increasing, and this boundary is dynamically changing.<\/p>\n\n\n\n

The Core Skill of Vibe Coding: Think Like a Product Manager<\/h2>\n\n\n\n

To get AI to produce high-quality engineering code, developers need to shift their mindset and treat themselves as the product manager of Claude<\/a>. Do not ask what Claude can do for you; ask what you can do for Claude.<\/p>\n\n\n\n

When facing complex development tasks, developers need to guide AI as if onboarding a new employee on their first day. Simply throwing out an instruction like \u201cimplement this feature\u201d is bound to fail. Developers need to provide detailed codebase navigation and clearly define requirements and constraints.<\/p>\n\n\n\n

Before letting Claude actually write code, Schluntz usually spends 15 to 20 minutes interacting with it. This includes letting AI explore the codebase, find relevant files, and jointly create a clear execution plan. Then all this well-organized context and specifications are merged into a single prompt before execution.<\/p>\n\n\n\n

Under this process, the model\u2019s task success rate increases exponentially.<\/p>\n\n\n\n

Real Case: 22,000 Lines of Code Merged in One Day<\/h2>\n\n\n\n

Schluntz revealed an extreme real-world case from inside Anthropic. Their team successfully merged a code change of up to 22,000 lines into a production reinforcement learning codebase, most of which was written by Claude<\/a>.<\/p>\n\n\n\n

To responsibly complete this merge, the team adopted four strategies:<\/p>\n\n\n\n