AI writes more code, but programmers still need to read and question it

AI coding tools now write a large share of new code. This guide is for developers and engineering managers in Korean companies and international teams. It sets out the author's view: what still makes someone a programmer is the habit of asking questions of the code and finding the answers in the code itself.

How AI changed the job

Tools such as Claude Code and Cursor let people with little coding experience build working applications in a short time. Anthropic describes Claude Code as an agentic coding tool that reads a codebase, edits files and runs commands, and it is available in the terminal, in code editors, in a desktop app and in the browser.

Two data points show the scale of the change:

• In March 2025, Y Combinator managing partner Jared Friedman said that a quarter of the startups in its Winter 2025 batch had codebases that were 95% AI-generated. The figure excluded code written to import libraries, and Friedman stressed that these founders were highly technical (as reported by TechCrunch).
• In a 2022 GitHub experiment, 95 professional developers were asked to build an HTTP server in JavaScript. Those using GitHub Copilot finished about 55% faster: 1 hour 11 minutes on average against 2 hours 41 minutes. 78% of the Copilot group completed the task, against 70% of the others. It was one narrow task, so the result does not cover all development work.

Typing speed and memorized syntax now set people apart less than before.

Old yardsticks and a better one

Debates about who is a "real developer" used to focus on technical depth: algorithms and data structures, mastery of a language and system design. These still matter, but AI now produces much of the code those skills used to produce.

In the author's view, the core of the job was never typing code. It is turning a business problem into software that works. The yardstick that still separates programmers is simple: can you ask questions of a piece of code and answer them from the code?

Using AI versus programming

Take "vibe coding" with an AI tool. You describe what you want in plain language, check whether the result behaves as expected, and ask the AI to fix it when it does not. You never read the code. This is useful work, but it is closer to building a product with an AI tool than to programming.

Programming means understanding how the code works, even when an AI wrote it, and being able to change and improve it yourself. When the software behaves in an unexpected way, a programmer looks first at the code that is actually running, not only at documents or outside sources.

TechCrunch's report on Y Combinator makes the same point: Y Combinator CEO Garry Tan warned that founders still need to understand their products in depth, because early reasoning models are weak at debugging.

Code avoidance at work

The author often sees developers who avoid reading code. On a new project they rely on documents or on colleagues' explanations. They say the documentation is too thin to start work, yet they rarely write documentation themselves. In the author's experience this happens surprisingly often among senior developers with more than 10 years of experience, while juniors with 1~3 years often adapt quickly to new languages and frameworks. The author sees this as a difference in approach, not in ability to learn.

This habit hurts the whole team. When working software already runs the business, the code is the most accurate and current record of how that business works. A developer who does not read it cannot help the team understand the business.

A common excuse is that "the legacy code is messy". Yet the same people often say this about projects they joined from the start. The problem is then the will or ability to understand, not the code. Some developers also judge code by whether they personally understand it, which gives no objective standard. Developers who read a lot of different code learn to compare their own code with other people's fairly, and to see where explanations or documents are needed.

Let AI help you read code

AI tools such as Claude and ChatGPT can explain what a function or module does in plain language. That lowers the barrier to reading code. But the explanation is only a starting point. You still need to compare it with the real code, find the gap between intended and actual behavior, and decide how to fix it. Agentic tools that summarize a whole project's structure help most when you already want to understand the code.

Checklist for developers and teams

1. When software breaks, open the code before anything else.
2. Run code reviews and expect every team member to understand the working code.
3. Treat messy legacy code as something to study, not to avoid.
4. Use AI to learn: ask why the code works, not only what it should be.
5. When hiring, the author suggests valuing people who use AI tools well and can check and verify the code those tools produce. Defining the problem and understanding the business context remain human work.
6. Prompt design and learning new tools are useful skills, but they work best on top of a real understanding of code.

The author's conclusion: AI does not remove the programmer's role. It raises the bar. Developers who can question code and find answers in it will keep creating value.