A regional bank says: "We want to use machine learning to reduce credit-card defaults." A data scientist insists on a written framing before any code.
Which framing is complete enough to start building a model?
Why did you pick that answer? Two or three sentences. The act of articulating it is what builds the judgment — not the click that follows.
A usable framing names the prediction unit, the target with a horizon, when the score is produced, and the decision that changes because of it. Every later choice — features allowed, metric, threshold — follows from those four lines. Option A is the only one that answers all four.
Naming the algorithm and an accuracy target says nothing about what is predicted, for whom, or by when. 95% accuracy on a 3%-default portfolio is worse than predicting "no one defaults".
"Better visibility" is not a decision. Without a decision, there is no way to choose the horizon, the costs of errors, or the threshold.
Building first is how leakage and the wrong horizon creep in. Feature importances cannot tell you which decision the business needs.