A magnificent tunnel through the wrong yard

There is no perfect "vibe coder." I've spent a year and serious hard cash learning why.

We've all seen the magic: ask AI for a feature, and it gives you a magnificent, elegant tunnel - fast, correct, code-perfect. But the deeper we go into the AI-assisted era, the more I think we risk losing the exact collaborative friction that keeps projects safe.

The signal that stops arriving

For twenty-five years my genuine concern has been understanding what bothers my teams, offering out-of-the-box thinking to help them, and guiding all of us to success. In a traditional setup, leadership relies on human empathy to do that.

You learn to read a room - not to catch people out, but to hear the subtle hesitation or stress in a confident voice that tells you: we might be hitting a wall here, let's pause and look at this together. It's an intuition built entirely on supporting each other through human interaction.

But AI doesn't leak its stress. It doesn't hesitate. It provides the same fluent, certain answer whether it's right or wrong. When I've asked an assistant "are you sure? this doesn't look right", it doubles down with the same corporate confidence every time.

It's a mirror. The human indicators we use to protect our teams and catch design flaws early suddenly return nothing.

Why is that wall load-bearing at all?

This is where the trap opens up, for developers and leaders alike.

I've spent years walking the floor asking one big-picture question: why is that wall we're tunnelling around load-bearing at all? It's the three-day obstacle that dissolves the moment we empower an engineer to challenge the premise.

The danger with AI isn't that it writes bad code. It's that it will happily build a magnificent tunnel through the wrong yard, never once asking if the wall needs to exist. It inflames our natural bias to just keep executing - infinite agreeable tunnelling, without the out-of-the-box friction that saves a project.

Software was always a team sport

The complete skillset - verifying the mechanical answers at the keyboard and interrogating the high-level questions - rarely lives in one seat. AI didn't change that. It just made us more prone to tunnelling deep into isolation, amplified by a very confident accomplice.

The scarce skill of the next decade isn't prompting. It's the leadership and engineering mindset that stops to ask about the wall before ten thousand flawless lines get written around it.

My rule now, for myself and the teams I care about: the model never gets the last word. Doubt ends in evidence - the log, the test, the running system. Never in the dialogue.

← All writing