You asked ChatGPT to explain self-attention. It gave you a paragraph that sounded right. You nodded. Two hours later, you couldn't explain it back to yourself. That's not a knowledge gap. That's a prompting gap. You asked the model to explain, but you never told it how deep to go — or at which level you actually needed the explanation. So it defaulted to a single-register answer that optimized for sounding helpful rather than for making you understand. The Feynman Technique fixes thi...