A summary of “the science behind prompt engineering” has been going around, built on five research findings. It is a good list of ideas. But the numbers attached to them are the kind that get copied from post to post without anyone opening the paper, so I opened the papers. Two of the five claims check out. Three do not — at least not in the form they are being repeated. In one case the real research points in the opposite direction from the advice. That is not a reason to throw the list away. The underlying advice is mostly sound. But if you are going to change how you work based on a study, it is worth knowing which study, and what it actually measured. Here is each claim, what I could verify, and what I would actually do about it. Quick summary Claim Verdict Leading questions push models toward the answer you implied ✅ Confirmed ”Think step by step” now loses to a short role prompt (IBM, 430,000 evaluations) ❌ Study not found — and persona research...