In this video I break down my “anti‑vibecoder” way of building with AI after basically living inside Claude Code and Cursor for the last year. The big idea is simple: LLMs are guessing, and your job as the human is to make those guesses better. If you treat AI like a magic button and just smash “continue” like a pigeon that’s had brain surgery, you’ll end up in error hell, restarting projects and burning tokens. I walk through the specific habits that changed everything for me: using plan mode first so the model maps the project, finds edge cases, and saves you time/money; dictating prompts because voice prompts tend to be longer, more specific, and more human; and using MCP servers (I like Context7) as a “USB port” that lets AI pull in live docs and real context. I also explain why I treat AI like a junior dev with ADHD—fast and often brilliant, but it needs tight context and small tasks. If you’re worried about learning, I use it like pair programming: ask why, slow down when it matters, and don’t sacrifice your growth for a company’s productivity. Finally, I show a design hack: screenshot a UI you like (Mobbin), then constrain AI to only use the color scheme and font so it doesn’t redesign your whole app.