Prompting Playbook

Rubrics, Socratic Questions, and Feedback Loops for Better AI Tutor Sessions

Learn a repeatable prompting pattern for guiding an AI tutor to evaluate work with rubrics, ask Socratic questions, and run tight feedback loops that improve the learner’s next attempt.

1) Start with a rubric that the tutor can actually use

Most “rubric” prompts fail because they list criteria without specifying what to do when evidence is missing. For an AI tutor, your job is to make the rubric operational: define what counts as evidence, and decide what the tutor should output for each criterion.

  • Criterion: A single, observable skill (not a vague value judgment).
  • Evidence: Where the learner should have shown it (a sentence, a step, a quote, a computation).
  • Rating behavior: What the tutor does at each band (what to reward, what to request).
  • Next action: The single best improvement request tied to the rubric.

When you write prompts, treat the rubric as a contract. If the model can’t explain how a learner earned or missed a point, the learning loop stalls.

2) Convert feedback into a Socratic question, not just a correction

Feedback works best when it triggers thought. Instead of “You should do X,” ask the learner to surface reasoning. A strong Socratic question has a specific target and uses the rubric as context.

Socratic prompt template

Given rubric criterion “(name it), I see (evidence / missing evidence). What would change in your answer if your goal were to demonstrate (what “good” looks like)?”

  • Anchor the question to a criterion so the learner knows what to improve.
  • Ask for a counterfactual: “What would you revise if…?”
  • Prefer “show your reasoning” over “tell me why,” because it invites concrete work.

3) Build a feedback loop that rehearses the next draft

A feedback loop is more than critique. It is critique plus revision plus verification. The most efficient loops follow a consistent rhythm: evaluate → question → revise → re-check.

  1. Evaluate: Score each rubric criterion and quote the relevant evidence.
  2. Diagnose: Identify one root cause (not five).
  3. Question: Use a Socratic prompt that forces the learner to change the targeted part.
  4. Revise: Ask the learner to produce a revised attempt.
  5. Verify: Re-score only the criteria affected by the revision.

This keeps the session from becoming a back-and-forth commentary loop. Learners leave with an updated artifact, not only notes.

4) Use “calibration prompts” to reduce tutor drift

Sometimes an AI tutor’s rubric scoring drifts after several turns, especially if the learner’s writing style changes. Calibration prompts lock the tutor to your rubric before it grades later work.

Before scoring

“Restate the rubric criteria and define what counts as evidence for each one. Then ask me one question about my goal before you score.”

During scoring

“If evidence is missing, write ‘evidence not present’ and ask for exactly one piece of additional information needed to score confidently.”

After revision

“Re-score only the criterion you targeted. Summarize what changed and whether the new evidence meets the rubric.”

5) Prompt for clarity: role, constraints, and output format

To get consistent tutoring output, specify role and constraints, then demand a structured response. A simple format reduces ambiguity for both the learner and the AI tutor.

  • Role: “You are an AI tutor with rubric-based scoring.”
  • Constraints: “Use rubric evidence only. If missing, ask one question.”
  • Output: “For each criterion: score band, evidence quote, explanation, and one next revision request.”

When you keep the format stable, your feedback becomes comparable across iterations, which makes learning measurable.

6) A complete sample prompt you can reuse

Copy this pattern, then swap the rubric and the learner’s work.

Sample prompt

You are an AI tutor. Use the following rubric to score my work.

Rubric criteria:
1) (Criterion name) — Evidence needed: (describe where it appears). Scoring: 3=fully demonstrated, 2=partly, 1=missing.
2) (Criterion name) — Evidence needed: ...
3) (Criterion name) — Evidence needed: ...

My work:
(paste learner attempt)

Task:
A) Score each criterion with a clear band.
B) Quote the evidence you used (or write ‘evidence not present’).
C) For the lowest-scoring criterion, ask one Socratic question that targets a specific revision.
D) Ask me to produce a revised attempt, then re-score only the targeted criterion.

Notice what’s missing: vague encouragement. The prompt forces evidence, requires a question, and demands an updated attempt.

Closing: what to track across attempts

To make progress visible, track changes per rubric criterion. The goal isn’t to “win” the score. It’s to reduce recurring errors and increase evidence density, one revision at a time.

  • Which criterion improved after the Socratic question?
  • Did the tutor’s feedback cite evidence or drift into generic advice?
  • Was the revised attempt re-checked against the rubric?

If you run the loop consistently, the tutor becomes a coaching system rather than a commentary engine.