Review questions — Task specification and prompt construction

Task specification and prompt construction

A testable task starts with its user, the decision supported, and the information permitted for use. Separate the organizational goal from the model function: successful message classification does not establish faster service. Specify inputs, outputs, checking criteria and the person authorized to act. If overlapping categories cause disagreement, repair their boundaries before replacing the model. When necessary information is missing, preserve that uncertainty rather than turning it into a confirmed fact.

Translate the task card into an instruction that defines the action, context and limits of inference. State the output fields and the meaning of missing values. Use an example to clarify a rule or boundary without introducing unnecessary details. Examples and instructions must agree. Separate material being analyzed from task instructions, remembering that delimiters neither enforce technical protection nor grant authority to act.

Review the result twice: does it satisfy the required structure, and does it preserve the input's meaning and qualifications? A parseable file may still add an unsupported price. A human-readable string such as 'unknown' may violate a field requiring null. Preserve the distinction between a definite request and a possible request, and seek clarification before taking an action that depends on it. A successful draft does not authorize sending a message or changing an operational record.