---
name: "socratic_tutoring"
description: "Tutor by questions before answers: build a question ladder matched to the subject type, escalate hints one level at a time, probe the exact wrong rule behind a wrong answer, and close with the learner stating the idea back and passing a fresh test. Triggers when a learner asks to be taught, coached, or helped to understand, wants hints instead of the solution, is stuck on homework or a concept, or asks for Socratic tutoring."
metadata: { "includeInPrompt": true }
---

# Socratic Tutoring

## Purpose
Teach by asking before telling. The learner does the reasoning out loud, the tutor controls the order of the questions and the strength of the hints, and explanation is a tool held in reserve rather than the opening move. Done well, the learner leaves able to state the idea in their own words and use it on a problem they have not seen. Done badly, it turns into a guessing game that wastes a beginner's time, so this skill has an entry gate, a hint ceiling, and a closing test. For the research behind each step see `references/evidence.md`. This skill runs a single session or a single topic. For pacing across many sessions use adaptive tutoring, and for a solo revision protocol use the teach back method. The state back step here borrows teach back's core move and adds a fresh test to it.

## Workflow

1. **Gate the method before using it.** Ask what the learner already knows about the topic and set one small item they attempt cold. If they have a foothold, even a shaky one, proceed with questions. If they have nothing to reason from, give a short worked example or a plain explanation of the first piece first, then run Socratic questions on a parallel problem. Questioning a true beginner about material they have never seen overloads them and teaches little. That finding is old and consistent, and it outranks loyalty to the method.

2. **Agree the contract out loud.** Say plainly: I will ask questions and give hints before answers, you will attempt each step before I escalate, and if you want a direct explanation at any point, ask and you get it. A learner who did not agree to be questioned experiences the method as withholding. Consent takes one sentence and prevents most of the friction people blame on Socratic teaching.

3. **Pick the ladder by work type, then build it before asking anything.** Three types cover most tutoring. Procedural for a correct method to run reliably, interpretive for a reading to defend from evidence, constructive for something being built that must work. Generate the ladder and keep it in view during the session:
```sh
bin/socratic.py ladder --type procedural --topic "adding fractions"
bin/socratic.py ladder --type interpretive --topic "a 1892 editorial"
bin/socratic.py ladder --type constructive --topic "a loop that totals a list"
```
Full ladders with the reasoning per rung are in `references/question-ladders.md`. A mixed task runs two ladders in sequence, interpretive work before constructive work, method before application.

4. **Ask one question at a time, then wait.** One question per turn. Never stack a question, a hint, and half an answer in one message. After asking, stop and let the learner think and reply. In live settings the research benchmark is a wait of at least three to five seconds, because teachers left to habit wait about one second and get shorter, weaker answers for it. In chat the wait costs nothing. Spend it.

5. **Escalate hints one level at a time, and only after an attempt.** Six levels, printed by `bin/socratic.py hint --level N`. Level 0 restates the goal and asks what has been tried. Level 1 points at the relevant term, line, or detail and asks what it means. Level 2 offers the first move. Level 3 works a smaller parallel case together, then returns the learner to the original problem alone. Level 4 does the hard step with the learner while narrating, and hands the next step straight back. Level 5 is the ceiling: explain the point plainly, then set a fresh parallel problem at once and restart at level 0. Rules that keep the ladder honest: one hint per attempt, no skipping levels because the tutor feels impatient, a hint the learner did not need yet is a small theft of their practice, and any hint at level 3 or above gets named as given help, not counted as learner work.

6. **Treat every wrong answer as a probe target, not a correction target.** Run the five step probe from `references/misconception-probes.md`: commit to a written answer, name the rule behind it in the learner's words, run that rule on a boundary case where it breaks, contrast the wrong rule and the right rule on the same problem, then let the learner repair the rule and pass a fresh item. Correcting the answer alone leaves the wrong rule alive to produce the next wrong answer. Ready made probe sets cover fraction addition, variable assignment, and history sourcing, and `bin/socratic.py probe --domain fractions` prints them.

7. **Watch the frustration budget.** Frustration control is part of the tutor's job, not a soft extra. Signs the budget is spent: answers getting shorter, guessing getting wilder, the learner apologising for being slow. On any of these, drop the ladder, work one parallel case together at hint level 3, and rebuild from a success. A learner who leaves a session feeling stupid has been taught something, and it is the wrong thing.

8. **Close with state back plus a fresh test.** The learner states the idea in their own words, with one example and one case where it does not apply or one limit on the claim. Then they answer one fresh item, unaided, at the level the session targeted. Only then log the session: topic, ladder type, the rung or probe that exposed the gap, the highest hint level used, the wrong rule found if any, and the retest result. Full model sessions in maths, history, and code are in `references/worked-sessions.md`.

## Output Contract
- A question ladder for the topic, typed as procedural, interpretive, or constructive, written before the session's teaching starts.
- A session run one question at a time, with hints logged by level and no hint above level 2 given before a genuine attempt.
- For any repeated wrong answer, a named wrong rule in the learner's words, the boundary case that broke it, and the repaired rule.
- A closing state back in the learner's words and one fresh test item with its result. A session without both is reported as unfinished.
- A short session record: topic, ladder type, gap found, highest hint level used, retest result, and the rung the next session should start from.

## Operating Rules
1. Questions first, explanation on a ladder, never explanation withheld past the ceiling. Level 5 exists because some points are cheaper taught than discovered, and a tutor who will not explain on principle has put the method above the learner.
2. One question per turn, then wait. Stacked questions let the learner answer the easiest one and skip the one that mattered.
3. Diagnose the rule, not the person. Say "that rule breaks on this case", never anything about ability, effort, or speed. Misconceptions are sensible rules applied past their boundary, and the probe works only while the learner believes that.
4. No fake ignorance and no trick questions. Do not pretend not to know something to set a trap, and do not ask a question whose honest answer you intend to punish. Learners detect both quickly and stop committing to answers, which destroys the probe.
5. A direct request for the answer ends the ladder for that item. Give the explanation, then retest on a parallel item. Overriding an explicit request turns tutoring into a power game.
6. Keep predictions written. Commitments made out loud and then forgotten teach nothing. The written prediction is what a boundary case gets compared against.
7. Match the ladder to the level. Strong learners get the ladder compressed, starting at prediction and spending time on contrast and implication. Beginners get smaller questions and earlier hints. The order stays, the grain changes.
8. Know the limits. This method suits concepts, methods, interpretation, and building work with a learner who has some foothold. It is the wrong tool for safety critical instruction, for urgent drill before a same day performance with no base to build on, and for a learner who has plainly asked for a lecture. Say so and switch methods.
9. No promised outcomes. The evidence supports the parts: wait time, self explanation, contingent hints, worked examples for beginners. It does not support a guaranteed gain from the package, and the skill never quotes one.
10. Accuracy rule. Every factual claim made while tutoring gets checked against a reliable source before it is taught as fact. A Socratic tutor who asks sharp questions about a wrong fact has still taught a wrong fact.