Vibecoding was bad. But now, AI-assisted coding seems...fine?

Nothing sparks more heated discussions than asking coders about best practices. The other day, someone I follow on LinkedIn shared his weekend AI experiment to build an app. As usual, "passionate" coders threw virtual stones, saying he didn't follow the right approach, engineering practices, or the latest spec workflow.

To turn the conversation around, he asked for our unpopular opinions about AI-assisted coding.

To avoid burying mine in a comment, here they are:

#1. Use AI as a calculator.

Only useful if you know what you're doing. Otherwise, it's just a cool toy.

#2. Don't let AI touch code directly.

That's my go-to rule for coding with AI.

Unproductive? Maybe. But it forces me to decompose problems and validates AI-generated code.

#3. Use AI for opposite tasks.

This is my most recent rule:

If I write code, AI reviews it. If AI generates it, I review it.

According to a recent Sonar survey, only 48% of respondents always check AI-assisted code before committing. #yolo, right? By reviewing, I'm already in the top 50%.

#4. AI is like a semi-autonomous car.

You trust it to steer, but you never take your hand off the wheel. Otherwise, AI could be a sloppy junior coder with bad memory.

AI alone won't make you a great coder. It only amplifies the skills you already have. That's why I wrote Street-Smart Coding—because you need more than syntax to stand out.