30 years coding, 8 with AI: The real skill isn’t getting AI to write code – it’s knowing when to ignore its confident answers. AI hallucinates with authority.
The story
I’ve been coding for over 30 years, working with AI for the last eight, and here’s something nobody talks about enough: the real skill isn’t getting AI to write code. It’s knowing when to ignore what it just confidently told you.
AI coding assistants have this tone – polished, certain, like they’ve done this a thousand times. And maybe the pattern has appeared in training data a thousand times. But that doesn’t mean the answer fits your problem.
I call it “vibe coding” now. Something in the response just feels off. Maybe it’s too smooth. Maybe it skips a business rule my whole team knows by heart. Maybe the architecture doesn’t fit. That’s when I stop and dig in.
Here’s what I’ve learned: AI is great at syntax and boilerplate. It’s terrible at understanding your project’s actual constraints. It doesn’t know your architecture, your stakeholders, or why you made that weird decision six months ago that saved everything. When it doesn’t have context, it guesses – and it guesses with confidence.
I’ve seen it invent APIs that don’t exist. Suggest old libraries. Ignore edge cases because they’re rare in training data. The problem isn’t that AI gets things wrong. It’s that it gets them wrong while sounding completely right.
So I treat every suggestion as a draft. I break problems into smaller pieces so I can isolate where the hallucination starts. I test everything. I ask the same question different ways and watch where the answers diverge. Sometimes I’ll run the same prompt through multiple models just to see where they disagree.
Recent Comments