AI Can Write the Code. But Can You Debug It? 😅

Fixing a collapsible sidebar menu.

This was before AI agents, when I was building a web app by copying and pasting code from ChatGPT. Asking for large chunks of code often broke the AI’s context, so I started asking for the smallest components individually.

Even then, it took a lot of time. Sometimes the menu overlapped with the menu button; other times, the button disappeared completely.

This taught me a lesson about maintainability: use AI to build small, atomic components, but handle the integration manually.

The sidebar example is a good reminder that some UI bugs aren't really component bugs at all. I've seen layouts where increasing z-index does nothing because an ancestor's transform has created a separate stacking context. A position: fixed menu can also end up positioned relative to that ancestor instead of the viewport.

That's the kind of issue where blindly asking AI to tweak CSS just creates a pile of conflicting overrides. Inspecting the computed styles and DOM hierarchy usually gets you closer to the actual cause. For UI components, I'd also test the open/closed states at different viewport widths, especially during resize, rather than relying on one screenshot at one screen size. The hard part isn't generating the CSS; it's understanding which browser layout rule is actually responsible for the behavior.

One thing worth distinguishing is fixing a bug from learning why it happened. At IT Path Solutions, we see the value of treating AI-assisted debugging as a hypothesis-testing process rather than repeatedly asking the model to rewrite code. A useful workflow is to reproduce the failure, identify the violated assumption, make the smallest targeted change, and add a regression test that would have caught the original issue. Otherwise, the code may start working again without anyone understanding the underlying cause, leaving the same failure mode ready to return in a different form. AI becomes much more useful when it helps developers build a better mental model of the system, not just produce another patch.

Đúng trọng điểm. Viết code nhanh là một chuyện, nhưng gánh đựng technical debt ngầm từ những đoạn code "chạy được nhưng không hiểu tại sao" lại là chuyện hoàn toàn khác.

Tôi thấy hai thách thức lớn nhất khi debug code do AI sinh ra:

1. Mất context logic nghiệp vụ: AI thường viết "correct by syntax" chứ không phải "correct by domain". Bug thường nằm ở chỗ logic nghiệp vụ bị hiểu sai nhẹ (off-by-one trong rule tính phí, race condition trong luồng thanh toán...), thứ mà unit test do chính AI viết thường không bao phủ được vì nó test theo implementation chứ không test theo spec.
2. Khó reproduce để fix: Khi tự viết, ta có mental model để đoán bug ở đâu. Với code AI, stack trace trỏ vào một hàm mà ta chưa từng đọc kỹ. Việc "đọc hiểu code người khác viết" (code review) giờ trở thành kỹ năng cốt lõi hàng ngày thay vì chỉ làm lúc merge PR.

Thói quen tôi bắt buộc team tuân thủ: Không commit bất kỳ dòng code AI nào nếu không thể giải thích được tại sao nó đúng cho test case xấu nhất mà bạn tự nghĩ ra. Buộc AI sinh test case edge-case trước, mới sinh implementation sau — giúp chuyển từ "vibe PS: the tool I meant is on labagent .tech