Do you ever work with these AI agents and find yourself repeating the same things over and over — only to see your virtual coding partner churn out yet another overly complex function with no tests? When it’s a clean slate, the AI can generate working solutions with a single prompt. But once it has to deal with the technical debt it created, it slows to a crawl — and becomes worse than useless.

Recently, I discovered a trick that made my clumsy AI coding agent so effective, it migrated a giant, horribly written legacy app from MongoDB to SQL in just a few hours — with almost no input from me.

It all comes down to one realization: the real gap between AI agents and humans isn’t creativity or logic — it’s the inability to form habits. So I built a system that lets me transfer my own habits into Claude Code, and the result is an agent that performs like a mid-level engineer who knows what good looks like.

In this talk, I’ll share the exact setup — including the prompt structure, quality checks, post-commit hooks, and a weirdly effective emoji trick. You’ll see how I apply this system live, and walk away with a blueprint you can use to train your AI coding agent.