marcgg, a developer and startup coach, has a rule: at least half his production code is written by hand. He's worried about what happens when developers stop understanding their own code. He points to aviation, where automated systems eroded pilot skills so badly that when automation failed, pilots couldn't fly their own planes. It happened with ORMs and SQL knowledge too. Now it's happening with AI coding assistants. For quick prototypes and personal tools, marcgg "vibe codes" freely. No code review, no standards, just ship it. But production code gets full scrutiny. He drafts architectural plans before asking Claude for input, then compares the two approaches. When he does use AI for production features, he picks isolated components, what Sandi Metz calls "Omega Messes," things he can test independently without risking the whole codebase. He recently used this approach to add voice features to a boxing app, vibe coding the audio generation tools while carefully integrating the feature into a 5-year-old codebase by hand. He's not alone in worrying about this. Early research suggests developers using AI assistants complete tasks faster but lose debugging ability after a few months, especially juniors. The more you offload, the less you understand what you're shipping. But marcgg isn't anti-AI. He estimates he's 5x faster now. He uses the time saved to think more and ship less. Walk around the neighborhood. Ponder ideas instead of rushing to implement the first one that comes to mind. He'd rather spend a few hours thinking about what to build than build something fast that nobody needs. It's a different kind of productivity, one that values judgment over speed.