A prompt instructs a model to "always respond in formal British English", and one of its three worked examples is written in casual American phrasing. Outputs are inconsistent. What is the fault?

The principle — Examples do not merely illustrate instructions; they compete with them. Where the two disagree, the model has two valid readings of the task and picks between them.

Why the key is correct — Google's checklist names conflicting instructions and examples as a distinct fault, and tells you to check for it by auditing the prompt for logical contradictions or mismatches between instructions or an instruction and an example. That is exactly the shape here: a stated rule about register, and a demonstration that breaks it. Because one of the primary jobs of an example is to show the model what a good response looks like, a non-conforming example is a strong signal that the rule is optional. The fix is to rewrite the offending example, not to strengthen the instruction.

Why the others are wrong — Undefined jargon concerns domain terms rather than ordinary descriptive phrases. Repeating the instruction adds redundancy while the counter-example stands. And more examples multiply the inconsistency instead of removing it.

Remember this — When an instruction and an example disagree, fix the example.

Sources — Google Cloud, Overview of prompting strategies.