Tutor prompts: start from an observable learning objective

Last updated: 5 Oct 202611 min read

tutorial

A tutor prompt is a contract for a learning interaction, not a request for an impressive lecture. Define the skill the learner should demonstrate, permitted reference material, current activity mode, and observable success criterion. Begin with a short diagnostic question that separates likely misconceptions. Ask the learner to commit an answer or describe an approach before providing the full solution in practice mode; in a timed assessment, follow the instructor's hint policy instead. Adapt the next turn to what the learner actually wrote, not an assumed ability level or a personality label. Keep a route for 'I do not know' so the tutor can offer a smaller subproblem without pretending the learner already understands.

Operational case

A fictional developer-training course teaches safe retries. The objective is to explain why a repeated request with the same identifier must not reserve stock twice. The learner sees request REQ-47 arrive twice. A weak tutor asks, 'Do you understand idempotency?' and then gives a broad definition. The stronger first turn asks what the reservation count should be after both deliveries and which record the service should consult. One learner says 'two reservations because there were two messages'; another says 'one, because the receipt is already stored.' Those answers call for different next prompts. The course does not infer competence from the learner's job title.

Mode: guided practice; objective: explain one reservation per request ID.
Diagnostic: REQ-47 arrives twice. How many reservations should exist?
Follow-up: which durable record distinguishes first delivery from retry?
Success: learner names one reservation and a stored receipt check.
If unsure: ask about the first delivery before naming the pattern.

Performance and operating cost

A diagnostic turn adds one model exchange, but can avoid several irrelevant explanations. With N learners and an average of T turns, model calls are O(NT); a fixed lecture uses fewer calls but supplies no evidence of individual understanding. Store only the response features needed for the next instructional step. The prompt may guide question choice, while a course system maintains mode, answer-release rules, and progress records. Evaluate on held-out wrong answers, partial answers, and 'not sure' cases, not only a learner who already knows the term.

Common Mistakes

Do not replace a diagnostic with 'do you understand?'
Do not infer skill from a profile label instead of the learner's answer.
Do not reveal the complete solution before collecting the response when practice mode requires an attempt.

Connected lessons

prompt engineering

learning systems