A colleague asks a model for "a short, professional summary" of a meeting transcript. Each time they run it they get a different length and a different level of formality, and they are becoming convinced the tool is unreliable. What is the actual fault?

The principle — A word that two readers could interpret differently will be interpreted differently by a model on different runs. Non-determinism turns an ambiguous specification into visible inconsistency.

Why the key is correct — Google's prompt-health checklist lists ambiguity as a distinct fault and says to avoid subjective or relative qualifiers that lack a concrete, measurable definition, providing objective constraints instead. Its own example is replacing "write a brief summary" with "write a summary of 3 sentences or less". Both "short" and "professional" are exactly this kind of qualifier: neither can be checked by counting or by looking. Rewriting them as "six bullet points or fewer, one sentence each, no contractions" removes the variation, because the model and the requester are now working from the same definition.

Why the others are wrong — More sentences of equally subjective instruction reproduce the problem at greater length. Calling the variation stylistic freedom misdiagnoses a resolution of undefined terms as a deliberate choice. And varying input length cannot explain different outputs from the same transcript.

Remember this — If you cannot verify compliance by counting or looking, the model cannot either.

Sources — Google Cloud, Overview of prompting strategies.