A brief guide to Vibe Coding

Filed under Software Engineering, AI, LLM on March 25, 2025. Last updated on March 25, 2025.

What is Vibe Coding?

Vibe Coding is a concept proposed by Andrej Karpathy in February 2025.

It is undeniable that using AI tools can make our programming work easier. But is it really that simple?

My View on Vibe Coding

When I first heard the term "Vibe Coding", I thought it sounded too easy and simple. This might make some people outside the industry think that programming is an easy job. And it might also deepen the anxiety of young engineers in the industry: Will programmers be replaced by AI? Can everyone easily complete a project?

My view is:

Programming is not an easy job. To write good code, you need clear logic, solid programming skills, and lots of programming experience. So, the bar will be higher, but the demand for talent in this industry will also increase. You can refer to Uncle Bob's new book We, Programmers: A Chronicle of Coders from ADA to AI for more details.
You can use AI to accelerate your learning process. If you only need to understand a field at a basic level, AI tools can help you get started quickly.
Using AI tools to reduce tedious coding work, so you have more time to think and design. Most of the time, I only need to press the Tab key to use Cursor.

How to use AI tools to improve work efficiency in daily work is a question I have been practicing and trying to answer.

Remember this existential conclusion: No one can help you make decisions, AI can't either. You need to be responsible for your decisions, and you need to be responsible for your code. Use AI, try to make mistakes quickly, and continue to learn.

Vibe Coding vs. AI-assisted Programming

Compared to "Vibe Coding", I prefer the term "AI-assisted Programming", to emphasize the role of AI tools in the programming process.

Comparison Item	Vibe Coding	AI-assisted Programming
Audience	Everyone	Professionals
Project Size	Small, Experimental	Medium, Large, Official Project
Skill Requirements	Programming Basics	Architecture, Abstraction, Design
Maintainability	Low	High

From Vibe Coding to AI-assisted Programming

Here is my personal flow of using AI tools to assist me in programming. You can see that the AI's participation is limited. All the parts that involve decision-making are still under the control of the programmer.

I will use 🤖 and 👨‍💻 to represent the parts that are dominated by AI and the programmer, respectively.

Preparation 👨‍💻

Currently, I use Cursor and ChatGPT the most. Cursor is used to handle any code-related tasks (editing, modifying, debugging), and ChatGPT is used for analysis and learning.

Generate Ideas 👨‍💻

First, you need to think, ask questions, and doubt. Without these creative and original ideas, AI is not able to help you.

Secondly, you need to be brave to ask questions and try. No one can be perfect, so don't be afraid to make mistakes. Follow your own pace and keep learning.

AI-assisted Programming and Requirements Analysis 🤖

In the current environment, requirements analysis is very simple. Using tools like ChatGPT, we only need to describe the requirements clearly, and we can get a preliminary solution.

For example, I can directly ask ChatGPT "I want to use RL to train a model to play Hollow Knight, what should I do?"。 Of course, I can also directly input my requirements in the ChatGPT window.

Generate Code 🤖

Based on my experience, there are some things to note when using AI to generate code:

Highlight the Determinate Parts 👨‍💻

Whether it's requirements, function signatures, type design, any parts you are sure of, you need to clearly point them out in the prompt, to ensure that AI can generate the code correctly. Or, you can also write this part of the code yourself, and let AI complete it.

Learn and Understand the Generated Code 👨‍💻

You may encounter some parts that you don't understand in the code generated by AI. At this point, you need to stop and understand this part of the code, then, based on your understanding, let AI modify this part of the code.

Iterate and Test the Code 🤖

Usually, after a few iterations, a stage of work is completed. At this point, you need to stop and reflect on whether the work of this stage has met your expectations. If it has, then continue to the next stage; if it hasn't, then continue to iterate.

Refactor, Refactor, and Refactor 👨‍💻

If you are not sure whether you should refactor, you can use tools like SonarQube to help you, to find the Bad Smells in the code. My experience is that, after a few iterations, the code generated by AI will become more and more complex, and it will become more and more difficult to maintain. At this point, refactoring becomes particularly important.

Testing (TDD) 👨‍💻

AI will not test the code it generates by itself, so, please test it. Of course, you can continue to use AI to generate test code, but you need to lead this process.

Generate Test Code 🤖

For developers, investing time in function signatures and type design can make it easier to test the code. Most languages, even dynamic languages, at least provide some degree of annotation ability, such as type annotations in Python.

These type annotations can help AI generate more accurate test code.

Run Test Code 🤖

Currently, Cursor has already been able to generate test commands very well. Especially when you have accumulated more and more test cases, using Cursor to generate test commands in the command line is very convenient.

In the latest version, Cursor even can generate commands in the chat window, you can click to run them directly in the chat window.

Fix and Improve the Code 👨‍💻

Whether the existing implementation has problems, or the test cases themselves have problems, you need to evaluate and fix these problems. Maybe AI can help you, but the final decision is still in your hands.

Release and Deploy 👨‍💻

For small projects, the code generated by AI can basically be released and deployed directly. However, for large projects, there are more issues to consider, such as separation of front and back ends, containerization, observability, multi-cloud deployment, etc. All these issues that need to be discussed further, AI is not able to give you answers directly, but you need to make decisions.

Summary

AI as an auxiliary tool can make your programming work more comfortable. But, the individual is the key to the success of the project. If you have any ideas, please try them. Even if they fail, you will gain valuable experience.