The saying 'prompt words will be eliminated' becomes popular every few months. The reason is usually: as models become smarter and agents become more autonomous, who still needs to be careful with their wording?
This judgment is partially correct, but overlooks a key difference: what was eliminated was the "chat style prompt word technique", leaving behind a complete set of system engineering methods.
Early prompt word engineering was indeed like a craft job. Everyone is competing for 'magic words': adding' please think step by step 'will have an immediate effect; Put it another way, the output quality is vastly different. These experiences are valuable, but difficult to accumulate and evaluate, more like personal skills than engineering abilities.
And when the large model enters the production system, the nature of the prompt words changes. It is no longer a text written for models to read, but a component of a system: system prompt words define the role and boundaries of the product, example samples are implicit rule libraries, thought chains are used to guide complex reasoning, and the description of external tool calls determines whether the intelligent body can use the interface correctly. These components require version management, comparative testing, and regression evaluation - this is no longer 'writing prompts', but' designing prompt systems'.
An important trend is the closed-loop evaluation. In the past, judging the quality of prompt words relied on manually drawing a few outputs; Now the team will build a review set and change the prompt words into quantifiable experiments: whether the accuracy has increased or decreased, how much delay has increased, and how much cost has changed. The role of prompt word engineers is increasingly resembling the intersection of product managers and algorithm engineers: they understand both the boundaries of model capabilities and user needs, and can speak with data.
So, the prompt words will not disappear, only 'mysticism' will disappear. As model capabilities continue to improve, perhaps future humans only need to say 'help me do this', but the ability in the middle layer to' translate fuzzy requirements into precise instructions that the model can execute 'will become even more important - although it may be renamed as system design, orchestration, or intelligent agent engineering.
Instead of worrying about whether the prompt words will become outdated, it's better to use them as a key to understanding the big model: every adjustment in wording is a small experiment on the model's behavior. This experimental ability will never depreciate in any era.