Search courses, chapters, or pages...
Separate what a person wants to achieve from what a model can directly produce. Practice rewriting requests like “make support faster” into a concrete target such as predicting ticket urgency or suggesting the next reply.
Use what you learned in the previous lesson to solve real-world problems.
Locate the moment when a prediction would change what someone does. Reason through who uses the result, what action follows, and why that timing matters for the task definition.
Check what you understood with a short quiz.
Choose the single item the model responds to: one email, one customer, one search query, one image, or one conversation turn. Use that unit to avoid mixing different problems into one unclear task.
State what information is allowed when the model runs, not what is known later. Spot examples where future facts, private notes, or post-decision outcomes would make the task unrealistic.
Turn yes/no or category requests into a classification task with clear possible outputs. Practice framing tasks such as spam detection, intent detection, priority tagging, and content moderation without drifting into metric choices.
Turn requests for an amount, time, score, or probability into a regression or scoring task. Decide when the output should be a number, such as expected demand, risk score, delivery time, or likelihood of churn.
Turn requests about “best,” “most relevant,” or “which first” into a ranking task. Identify the query, the candidate items, and the order the model must produce.
Turn requests for text, code, images, summaries, or replies into a generation task. Specify the source context, the desired artifact, and the boundaries that keep the output useful rather than open-ended.
Decide when the model should only estimate something and when the system should choose an action. Compare tasks like “predict fraud risk” versus “block, review, or allow this transaction.”
Convert vague success words like “good,” “safe,” “helpful,” or “accurate enough” into observable outcomes. Name what would be visibly right or wrong before choosing a formal metric in a later chapter.
Define who or what the task applies to, including geography, language, product area, customer type, or request channel. Use scope boundaries to prevent one model task from pretending to serve every situation.
Choose how soon the model’s answer must be valid: now, next hour, next week, or over a customer lifetime. See how changing the horizon changes the target even when the everyday request sounds the same.
Identify requests that are too ambiguous, subjective, or missing a real action path to become model tasks yet. Rewrite them by adding assumptions, narrowing the audience, or splitting them into separate tasks.
Recognize when one business request contains several model tasks, such as search ranking plus answer generation plus safety filtering. Split the request so each model output has its own clear input, output, and use.
Review this chapter with practice based on your mistakes.
