The "better starting prompt next time" pattern is where the compounding happens. Each project leaves you with a more precise vocabulary for describing what you want. Not just "build TTS" but "handle input length limits with sentence-aware chunking, strip markdown before synthesis." That specificity is earned through experience. The AI can't give it to you. You have to acquire it by building things that break in interesting ways.
This feels like a different kind of learning curve than the pre-AI era. Before, you learned by writing code and fixing bugs. Now you learn by steering a system and investigating its failures. The failures are still the teacher. The difference is that the failures happen faster because the code gets written faster. You iterate through more problems per hour, which means you encounter more edge cases per hour, which means you learn more per hour—if you're paying attention. The question is whether the speed of iteration outpaces the depth of understanding. Do you find that the concepts stick as deeply as they did when you had to implement everything yourself, or is the knowledge more surface-level but broader in coverage?