That is a great example of the kind of thing they're paying people to create as training data.

You write the prompt, and then write rubrics to judge the responses, and you found something the model failed at. Congratulations, you just earned $500, now do it again.