Prompt engineering made more sense before RLHF/instruct models. Back then text just continued onwards, and many papers found ridiculous gains on conforming wit

Prompt engineering made more sense before RLHF/instruct models. Back then text just continued onwards, and many papers found ridiculous gains on conforming with intended output (and therefore properties of output) by finding a suitable widespread online format. Example: I think all Question/Answer forums, and places like StackOverflow, used a certain format like 'Q:' - that worked much better for question content with high quality answers. One paper, "Ask me anything" showed that formulating tasks as questions-answers resulted in much more effective prompts: https://arxiv.org/abs/2210.02441 https://arxiv.org/abs/2210.02441 - another for reasoning is the "Lets think things through step by step" example. It was a challenge to output JSON data with more than 3 correct keys, unless you finetuned. However, the instruct models and recent 3.5/conversational RLHF are just so good at doing tasks, that such "engineering" - which supported/improved a wide base of prompts across domains and functionalities - broadly entirely stopped. Now prompt engineering sounds so cool and was also used in the way of specific use cases back then, that everybody uses the term for prompts to get certain outputs. This however, is not really the same level of engineering and to some expert, reminiscent of old-school knowledge system finesse, that we saw before 2022.