Recently, tech self-media has been hyping up Vibe Coding, which sounds cool. The general idea is that you don't need to understand rigid programming syntax; as long as you have a feel and an idea, AI can write the code for you.
Many people say that the barrier to programming has completely disappeared, and everyone can be a programmer. But from my personal experience, this has been overhyped.
In fact, in recent years, programming tools have always been at the forefront of AI applications. Before general-purpose agents like Manus entered the public eye, tools like Cursor, Windsurf, and Copilot were already widely used by engineers.
Every time a new tool appears, a bunch of self-media outlets claim that programming is dead and that ordinary people will soon be able to develop software as easily as ordering takeout. But is that really the case?
As long as the code is still there, the barrier remains
If you open any AI code editor, such as Cursor, Antigravity, Trea, or Tongyi Lingma, you'll see dense lines of code still displayed on the screen.
This is actually a very important signal. The reason code is still shown is that what AI produces is not yet 100% reliable. In the professional software development field, code is the only verification standard. It serves functions like validation, debugging, accountability, and maintenance. Once a system becomes complex, there must be an intermediate form that can be checked and corrected. At the current stage, code is that intermediate form. If AI makes a mistake or the functionality doesn't meet expectations, you still have to go back and look at where the code went wrong.
Any user-facing tool product won't expose its core business logic. If AI editors were truly fully reliable, there would be no need for you to see the code, just like how Meitu Xiuxiu doesn't show you the algorithm instructions for image processing when you edit a photo.
So, current Vibe Coding actually serves those with basic engineering judgment. AI saves developers the effort of memorizing syntax and writing boilerplate files, but it can't think through the entire system's logic for professional developers.
Therefore, a more accurate positioning of Vibe Coding is to reduce the operational burden on professional developers, not to allow non-professional users to bypass engineering complexity. It reduces repetitive labor and shortens implementation time, but still requires users to have the ability to judge right from wrong.
The steering wheel in autonomous driving, you still have to hold it
To understand this, we can draw an analogy with a familiar scenario—autonomous driving.
In closed, simple, and standardized road conditions, like highways, autonomous driving performs excellently. AI can steadily keep the car in its lane. But once it enters complex urban road conditions, it's prone to errors.
So, even with the most advanced onboard systems running the most sophisticated autonomous driving technology, cars still retain the design of a steering wheel, accelerator pedal, and brake pedal. This is so that when AI encounters situations it can't handle, the driver can quickly take over to prevent accidents.
Similarly, with Vibe Coding, when you ask AI to write a small tool or a single web page feature, it performs very well. But when developing a complex software system with multiple integrated modules, issues like AI hallucinations and attention lapses emerge. The complexity here isn't about the amount of code, but the continuous expansion of states, the increase in implicit constraints, and the amplification of early decisions later on. Current models struggle to maintain long-term global consistency and lack reliable risk perception and self-degradation capabilities. Errors generated early on can accumulate over time, ultimately leading to program failure.
AI editors expose the code because humans are still seen as the regular fallback, requiring developers to take over or verify at any time. Just like cars with autonomous driving still have a steering wheel and pedals, requiring the driver to be ready to take control.
Should ordinary people learn Vibe Coding?
Nevertheless, this doesn't mean Vibe Coding is overrated to the point of being useless.
I believe Coding Agents are likely to become key middleware for various AI agents in the future. Code is a universal, composable, and verifiable way to express actions. Translating high-level intentions into executable structures is an unavoidable part of intelligent systems.
However, for a considerable period, it will be more suitable as a constrained execution module rather than an independent entity handling complex goals.
From a product experience perspective, ordinary people don't need to learn how to use Coding Agents. Any truly mature product aimed at the general public should aim for zero learning cost. If a product requires users to understand context management, failure rollback, or handling bugs caused by model biases, it means the product hasn't been fully packaged.
If you're only learning these complex tool operations out of FOMO, I think it's completely unnecessary. Tools are constantly changing, and they are inevitably becoming simpler. If a tool still requires you to deliberately learn how to use it, it means it's not good enough yet.
However, if you want to use these tools to understand how programs work, that's very valuable.
The future is a world driven by AI and data. While you may not need to write code yourself, understanding how programs break down problems and what logical constraints are is like learning to drive in the age of autonomous driving. It's not about becoming a professional driver; it's about understanding the rules of how the world operates, so you can make better judgments when problems arise.
From the ebb of Prompt Engineering to the rise of Vibe Coding
Looking back at prompt engineering, which was hot in 2023 but has since cooled down significantly, we can see a similar pattern.
So-called techniques are essentially temporary patches when a system is not yet mature. Once model capabilities improve, these techniques are quickly absorbed. What ultimately remains is not the wording of prompts, but more fundamental abilities: the ability to clearly express needs and the ability to verify results.
Whether it's prompt engineering or Vibe Coding, what truly determines the outcome is never the technique itself, but whether you know what you want and what results are acceptable.
Vibe Coding has indeed made development faster, but it hasn't made programming itself disappear. It's more like giving a professional chef a more modern set of cooking tools, but these tools can't turn someone who has never stepped into a kitchen into a Michelin-star chef overnight.
AI hasn't created a completely new threshold of ability; it amplifies the differences that already exist between people. Those who express themselves clearly will be more efficient. Those with poor judgment will only get unusable results faster.
AI is just a tool. In the end, the person responsible for setting goals, making judgments, and verifying results is always you.