Prompt Engineering Is Not Dead

It’s become a cliche to say prompt engineering is dead. The obituaries are confident that the skill of optimizing instructions to guide AI models like ChatGPT or Gemini is not needed anymore. In June 2025 the AI researcher Andrej Karpathy popularized a tidier replacement term, calling context engineering the delicate art of filling a model’s context window with the right information at each step. A month later, Gartner was telling clients that "context engineering is in, and prompt engineering is out." If you’re celebrating the funeral of prompt engineering, then you never understood what it was.Someone coined the pretentious term ‘prompt engineering’, but it was only ever one thing. The refinement of language to express an idea clearly enough that it can be acted upon. Strip the term away and what's left is a skill as old as teaching, as old as managing people, as old as raising children. That is, making your intent legible to someone, or something, else. That isn't going anywhere. If anything, the current debate shows how thoroughly we've confused the label for the thing it points at.What The Obituaries Get RightThe AI models have changed substantially since 2022. Early systems rewarded a kind of incantation. Magic phrases. Role-play tricks. The odd threat or bribe typed into a chatbot to squeeze out something better. That barely matters now. What moves the needle is how you structure the whole task.Tech entrepreneur Andrew Ng made this plain at Sequoia Capital’s AI Ascent back in 2024, when he shared results on the HumanEval coding benchmark. Handed a single prompt, GPT-3.5 solved roughly 48% of the problems and the stronger GPT-4 around 67%. Then his team wrapped the weaker model in an iterative loop, letting it draft, check its own work and revise across several passes. Its accuracy climbed to about 95%, beating GPT-4's single attempt outright. The lesson is to stop polishing one perfect sentence and start designing the system the model works...