A prompt asking a model to categorise a book returns the right category but as a full explanatory sentence, when you needed one of four fixed labels. Which iteration strategy does Google demonstrate for this?

The principle — When a model will not follow an instruction, a different framing of the same task is often more effective than a firmer version of the instruction that already failed.

Why the key is correct — This is Google's own worked example. A prompt asking which category The Odyssey belongs to returns a correct but discursive answer that does not stay within the supplied options, and Google's stated remedy is to switch to an analogous task: rephrase the instructions as a multiple choice question and ask the model to choose an option. The revised prompt returns the bare label. The generalisation Google draws is that if you cannot get a model to follow your instructions for a task, giving it instructions for an analogous task that achieves the same result is a legitimate strategy.

Why the others are wrong — Rebuilding discards a prompt that is already getting the categorisation right. Repeating the instruction is redundancy rather than emphasis. And stripping the sentence afterwards leaves you maintaining fragile text handling that a prompt change removes.

Remember this — When the instruction is not working, change the shape of the question rather than its volume.

Sources — Google, Prompt design strategies.