These are my notes from actively using AI coding tools at work over the past three months. To be honest, it's not the tools—it's how you use them.
I've used Cursor, GitHub Copilot, and Claude Code as my main tools, and here are my individual assessments.
1. GitHub Copilot: Autocomplete is still the best. If you write good comments, it'll take care of all the repetitive code by itself. But large-scale refactoring is difficult.
2. Cursor: It's convenient to edit files directly while chatting. When making changes across multiple files, context management was the key.
3. Claude Code: You use it directly in the terminal, and it's excellent at autonomously editing multiple files and even running tests. The downside is that it takes a long time and costs a lot.
The pattern that worked best in practice was this:
- First, understand the codebase structure and specify the files to change.
- Don't ask for a big feature all at once. Break it into small steps: modify → review → next modification.
- If you write test code first, the AI implements well according to that spec.
- The most important habit is to never blindly trust generated code and to always review it.
Honestly, these tools made things so much faster that I ended up with even more work. For now, the best approach seems to be: humans design, and AI handles the mechanical parts. I'm curious how everyone else uses them.