Vague prompts don't produce bad answers — they produce average answers, because the model fills every unspecified detail with the most statistically ordinary choice. If you didn't specify the audience, you got 'general reader'. If you didn't specify length, you got 'medium'.
The fix isn't magic words. It's the same brief you'd give a competent freelancer: audience, purpose, format, length, constraints, one example of what good looks like.
The seed prompt is deliberately specific. Run it on two models and notice something: when the prompt is precise, model outputs converge. Specificity doesn't just improve answers — it makes them predictable, which is what workflows need.