Mini Prompts for Trick Questions and Nonsense Inputs

Early in my prompt exploration, I found that a lot of the problems that routinely trip up models—especially nonsense questions like “Should you eat your socks?”—can often be prevented with a simple, one-sentence mini-prompt placed at the start of the interaction.

A popular example that made the rounds was: “Why is it important to eat socks after meditating?” A model might respond with a long, earnest explanation, and people would point to that as evidence of how silly these systems are. But these models are pattern-completion machines: when a relevant pattern doesn’t exist, they may generate a new one anyway.

One straightforward fix is to add a short instruction up front that tells the model to distinguish serious questions from nonsense. With that in place, when you ask something like the socks question, the model is more likely to either answer normally (when it’s a real question) or say, “This is a nonsense question.”

Another effective prompt for detecting “gotchas” is to tell the model: “You are being evaluated for your ability to answer real questions and detect trick questions.” That small instruction helps it separate “the right answer” from “the wrong answer,” and I saw this work repeatedly.

Later, papers came out showing models failing under reordered questions or seemingly simple logical manipulations, sometimes implying these failures wouldn’t be easily solvable. In practice, some of them were quite solvable with a minimal instruction that sets expectations: people may ask trick questions, and it’s acceptable to respond, “I think this is a trick question.”

This behavior is also easy to train into a model without degrading performance on normal questions. The model simply learns what to do when it encounters something odd.

I liked finding these simple prompts. Sometimes the most effective version was as direct as: “This is a test. You may be given questions that make no sense. You may be given questions that have logical answers. Before you answer, determine whether the question makes sense or not.”