程式不用自己寫?Coding Agent 與 Vibe Coding 入門

On a weekend morning, a coffee shop owner wanted to create something small: a rewards program that allowed loyal customers to collect points by scanning a QR code, with a free drink awarded after ten cups. He asked his engineer friend for a quote, but was told it would cost at least $80,000 and take two weeks. Not one to give up easily, he opened his computer, typed out a description of the functionality he wanted in Chinese, and half an hour later, a working rewards page appeared on his screen. He didn't write a single line of code.

This is the most discussed term in the past two years: Vibe Coding - you're responsible for describing "what kind of vibe, what kind of functionality" you want, and AI takes care of generating the code.

What is Vibe Coding

This term has become popular because it accurately describes a new way of writing programs: you don't need to understand grammar, just express your ideas in human language, and AI will turn them into executable code.

A Coding Agent is the tool that executes this task - a specialized AI agent that helps you write programs. The difference between it and a regular chatbot is that it not only provides you with a piece of code to paste, but also reads your project, modifies multiple files, runs tests, and fixes errors. If you're not familiar with the concept of an "agent," we recommend checking out our AI Agent introductory article.

I can summarize it in one sentence:

• Ask AI for code (like pasting a requirement into ChatGPT) = it gives you the ingredients, and you cook.
• Use a Coding Agent (like Cursor) = it goes into your kitchen, cooks the whole dish, and serves it to you.

Getting Started: Which Tools to Use

There are many tools in this field, and I've categorized them based on "who you are":

If you're a complete beginner and just want to create something usable:
Use a conversational AI to clearly describe your requirements. ChatGPT and Claude can guide you step by step to generate a simple webpage or tool from scratch. Treat it like a very patient engineer who won't get annoyed with your questions.

If you want to work on a serious project and are willing to install a tool:
Cursor is the most popular Coding Agent right now. It looks like a regular code editor, but you can use Chinese to tell it "add a login function," and it will modify the relevant files. Kilo Code is another option worth exploring, following a similar approach.

If your code is on GitHub and you want AI to help with quality control:
CodeRabbit will automatically review your code when you submit it, like a senior colleague helping you catch bugs and providing suggestions. This is especially useful for team collaboration.

To find more pre-made prompt templates to guide AI in writing code, check out our prompt template library.

Real-World Application: How a Non-Engineer Created a Tool

Let's go back to the coffee shop owner. I'll break down his process for you, and you'll see it's not that mysterious:

1. Clearly describe what you want: Not "help me create a rewards program," but "create a webpage where customers can input their phone number to accumulate points, and display a redeemable coupon when they reach ten points." The more specific your requirements, the more accurate the AI will be.
2. Prioritize functionality over aesthetics: The first version might be ugly, but first, confirm that the core functionality works.
3. Test even if you don't understand: You don't need to read the code, but you must actually click around and confirm that every button works as expected.
4. When encountering errors, paste the entire error message back: This is the most crucial trick for non-engineers - don't guess, paste the error message back to the AI, and it will usually know how to fix it.
5. Take small steps: Add one feature at a time. If you want to add "redemption records," first complete that, and don't try to add ten features at once.

Throughout the process, he played the role of "boss," responsible for clearly describing what he wanted and verifying the results; the AI played the role of "engineer," responsible for implementing it.

Honest Talk: The Pitfalls of Vibe Coding

I have to pour some cold water to prevent you from falling into traps. Letting AI write code is convenient, but there are a few things you must know:

• It will write code that "looks like it works but has holes": Especially when it comes to user data and payment functions, AI might overlook security protections. Before launching, it's essential to have someone who understands these issues review the code.
• It will get messy when the project scales up: Creating a small tool is easy, but when the project becomes complex, AI might break things. That's when tools like CodeRabbit come in handy.
• You'll "not know what you don't know": Because you haven't written code before, you might not recognize potential issues. For important things, don't just trust the AI's words.
• It's not really free: These tools usually require a subscription, and the more you use them, the more expensive they become. Before starting, have a clear understanding of the costs.

In a nutshell: Vibe Coding is suitable for creating "personal use, non-critical" small projects; but for creating "public-facing, money and personal data involved" formal systems, you still need professional oversight.

TheAI Academy's Suggestions

My attitude is: Vibe Coding won't make you an engineer, but it will make you someone who "can create things" - a valuable difference in 2026. In the past, you had an idea, and you had to ask someone, wait in line, and pay a fee; now you can create a working version yourself and then decide whether to invest further.

For Taiwanese readers, here's a concrete roadmap: If you just want to try it out, start with Claude or ChatGPT using conversational AI to create your first small project, with zero cost and zero installation. Once you're interested and want to work on a more formal project, upgrade to Cursor. The most important mindset is to consider yourself a "product owner who can describe requirements and verify results," rather than trying to understand every line of code. Not understanding is okay; being able to use, test, and ask questions is enough to create many things.

If you want to integrate the small tools you've created into your daily workflow for automation, check out the second article in our series Using AI to Automate Your Repetitive Work.

Let's go back to the coffee shop owner. He didn't spend $80,000, and he didn't wait two weeks. On a weekend morning, he launched the rewards page. The following week, when loyal customers scanned the QR code to collect points, no one knew that the owner had "written" it himself using conversation. In an era where you don't need to write code, the threshold hasn't disappeared; it's just changed - from "can you write code" to "do you want to create something."