What happened?
“Prompts aren’t Real” reached 114 points on Hacker News, linking to evaluation.club.
Why people care
Prompt engineering is treated as a discipline with stable objects. If prompts are not stable, then benchmarks and evaluation methods built on fixed prompt sets inherit that instability.
The argument in context
The essay challenges the assumption that a prompt can be treated as a durable, standalone object while the model, system instructions, tools and surrounding context keep changing. That is a measurement problem before it is a writing problem.
Why it matters
If the object being evaluated is unstable, a benchmark score can hide changes in the surrounding system. Teams may need to version the full interaction and the evaluator rather than storing only the user-visible prompt.
What to read for
Read the original argument before accepting the headline. The Hacker News discussion shows how people interpreted it, but the source article contains the assumptions that should be tested against a real evaluation workflow.