How to Keep Your Prompts from Getting Boring

Sometimes a model’s first answer is competent but familiar. The second and third answers can be more of the same: polite, safe, and shaped like every other answer to the prompt.

A 2025 preprint, Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity, studies one possible reason. The authors argue that preference data can favor typical responses, reducing output diversity after post-training. They introduce Verbalized Sampling, an inference-time approach that asks a model to produce several responses and verbalize a distribution over them. The paper reports improved diversity in its evaluated tasks; it remains a preprint, so its results should not be treated as a universal rule for every model or prompt. (paper)

The practical lesson does not require treating a chatbot’s stated probabilities as calibrated confidence. For ordinary creative work, ask for distinct options and describe what must differ.

A useful prompt pattern

Give me five genuinely different approaches to the following task. Each must satisfy the same factual constraints, but use a distinct angle or structure. Label the difference in one sentence before each answer. Do not rank the options by truth or confidence.

For a short story, the difference might be point of view, emotional tone, setting, or conflict. For an article opening, it might be a firsthand observation, a concrete scene, a counterintuitive question, a historical comparison, or a plain definition.

The labels matter. “Five options” alone often produces five near-duplicates. Naming the dimension of difference gives the model—and the person selecting from the list—a real constraint to work with.

Use variety on purpose

More options are not always better. For a factual or consequential task, a single carefully sourced answer may be safer than a menu of clever alternatives. Variety is useful when you are exploring form, language, examples, or possible directions—not when you are choosing what is true.

After the model offers alternatives, the human job remains the same: reject options that distort the facts, choose the one that suits the audience, and revise it until it sounds like the person responsible for publishing it.

That is how a prompt stays interesting without confusing novelty for quality.