Last week I ran a small experiment. I let AI coding assistants handle as much of my daily work as possible — Cursor, Copilot, Claude in the terminal, the works. Not to prove a point. I genuinely wanted to know where the ceiling is right now.
Here's my honest take after five days.
What it nailed
Boilerplate. Setting up Express routes, writing test skeletons, generating TypeScript interfaces from JSON — all that stuff that eats 20 minutes but requires zero creativity. AI handles it faster than I can type.
Explaining unfamiliar code. I inherited a gnarly regex-heavy config parser last month. Pasting it into Claude and asking "what does this do, line by line" saved me an hour of squinting.
Refactoring suggestions. "Extract this into a utility function" type stuff. Not always right, but it got me thinking in the right direction more often than not.
What it got wrong — every single time
Deleting things it shouldn't. I asked Cursor to clean up a utility file. It removed every comment. Not refactored — deleted. The comments were the only documentation that file had. I've seen this pattern repeatedly: AI assistants treat comments as noise.
Making up APIs that don't exist. I lost count of how many times it generated code calling methods that were never part of the library. It writes with such confidence that you don't question it until runtime. And if you're not running tests immediately, that bug hides for days.
Ignoring the surrounding codebase. This is the big one. AI writes code that works in isolation, but doesn't match the patterns already in your project. Different error handling style. Different naming conventions. It's like hiring a contractor who does good work but never reads your team's style guide.
The pattern I actually settled into
By Thursday, I'd stopped asking AI to "write this feature." Instead I'd:
- Write the rough implementation myself (15 min)
- Ask AI to review it for edge cases (2 min)
- Ask AI to write the tests (5 min)
- Actually read the tests and fix the ones that test the wrong thing (10 min)
Total: 32 minutes instead of my usual 45, but with better test coverage. Not revolutionary. Just... better.
The uncomfortable truth
AI coding tools right now are like a very fast junior dev who never sleeps but also never asks clarifying questions. The output looks professional until you look closely.
The people getting the most out of these tools aren't the ones outsourcing their thinking. They're the ones who already know what good code looks like and use AI to get there faster.
If you're still figuring out your approach, start small. Use it for the boring stuff. Keep your hands on the wheel for anything that matters.
I've been building workflows around AI tools for the past year — figuring out what actually sticks vs. what's just demo-worthy. If you're interested in that kind of practical approach, I share more on my profile.
Top comments (0)