Turn a vague request like “catch bad claims” or “make search better” into a specific moment where someone needs help. Identify the user, the decision point, the action that follows, and what would change if the model worked well.

Decide whether the model should predict a value, rank options, generate content, or make a rule-bound decision. Compare common framings so the same request can become the right kind of model task instead of a mismatched solution.

Apply the previous explanations in a guided problem.

State exactly what the model receives and what it must return at the moment it is used. Define the unit of work, the available context, the output format, and whether the result should be a class, score, ordered list, text, or structured object.

Translate words like “good,” “risky,” “relevant,” or “high quality” into a target the model can learn or produce. Pick an observable outcome or proxy, set the time horizon, and name the cases that are in or out of scope.

Apply the previous explanations in a guided problem.

Check your understanding with a short quiz.

Build practical limits into the task before refinement starts. Reason through safety rules, privacy limits, latency needs, human review, abstention, and the difference between a model prediction and the action taken from it.

Use a small set of realistic example requests to expose hidden ambiguity in the task. Write expected outputs for normal, borderline, and out-of-scope cases so stakeholders can agree on what the model is actually being asked to do.

Apply the previous explanations in a guided problem.

Check your understanding with a short quiz.

Review this chapter with practice based on your mistakes.