Prompt Engineering is Complicated and Contingent

Summary

Repeated tests illustrate why a successful answer does not establish a reliable prompting technique.

Using repeated difficult-question trials, the authors show that prompt effects can reverse across questions and that evaluation thresholds alter conclusions about reliability.

An editorial teaching idea is to have learners define reliability before comparing repeated outputs. This turns prompting advice into an inspectable experiment; that classroom application was not tested here.

Model/benchmark-specific evidence. The suggested classroom use is an editorial inference.