Andrej Karpathy moved over to Anthropic just five weeks ago. A teammate let someone peek at the exact LOOPS.md file Karpathy uses for prompt engineering. That document defines how you structure inputs, manage context windows, and handle iterative feedback loops with Claude. Dropping LOOPS.md straight into your own setup transforms Claude’s behavior. Instead of vague, one-size-fits-all replies, the model zeroes in on the specific task you’ve defined. In one test, the very first response shifted from a generic summary to a detailed, step-by-step breakdown—mirroring the way the user thinks through problems. The key idea: you’re not talking directly to the model. You’re building a small system around it—templates, token budgets, chain-of-thought triggers—that guides Claude toward the style and depth you need. The file spells out prompt formats, variable placeholders, and protocols for handling multiple passes over the same prompt to refine results. If you work with Claude or plan to, LOOPS.md is a game changer. Save it now before it disappears from your feed. Then try running it against a few of your existing prompts and watch how your interactions improve.