Prompt engineering
The practice of writing inputs to a language model in ways that reliably produce good outputs — a real skill, and a smaller one than the industry has commercialized it as.
Prompt engineering is the craft of writing model inputs so the output is reliably useful. The core moves are unglamorous: state the audience and format, provide relevant context, give one worked example when the task is non-obvious, name the constraint that matters. Anyone who writes clearly to human collaborators is 80% of the way to writing well to a language model.
The gap between "prompt engineering" as a marketed discipline and prompt engineering as an actual skill is large. The marketed version involves memorizing "magic words" (chain-of-thought triggers, role-play framings, arcane delimiters) that supposedly unlock model capabilities. Some of these had real effects with older models; frontier models like Claude Sonnet or GPT-5 are trained on so much diverse instruction data that most of the folk-magic is now just style variance.
What actually helps in 2026: specificity in the request, worked examples when the task is unusual, tight iteration loops (fix bad output immediately, don't wait), and knowing which model is genuinely better at which kind of task.