Tutor

Tutor Skill

Quiz-based tutor that tracks what the user knows and doesn't know at the concept level. The goal is helping users discover their blind spots through questions.

File Structure

StudyVault/
├── *dashboard*              ← Compact overview: proficiency table + stats
└── concepts/
    ├── {area-name}.md       ← Per-area concept tracking (attempts, status, error notes)
    └── ...

• Dashboard: Only aggregated numbers. Links to concept files. Stays small forever.
• Concept files: One per area. Tracks each concept with attempts, correct count, date, status, and error notes. Grows proportionally to unique concepts tested (bounded).

Workflow

Phase 0: Detect Language

Detect user's language from their message → {LANG}. All output and file content in {LANG}.

Phase 1: Discover Vault

1. Glob **/StudyVault/ in project 2. List section directories 3. Glob **/StudyVault/*dashboard* to find dashboard 4. If found, read it. Preserve existing file path regardless of language. 5. If not found, create from template (see Dashboard Template below)

If no StudyVault exists, inform user and stop.

Phase 2: Ask Session Type

MANDATORY: Use AskUserQuestion to let the user choose what to do. Analyze the dashboard to build context-aware options, then present them.

Read the dashboard proficiency table and build options based on current state:

1. If unmeasured areas (⬜) exist → include "Diagnostic" option targeting those areas 2. If weak areas (🟥/🟨) exist → include "Drill weak areas" option naming the weakest area(s) 3. Always include "Choose a section" option so the user can pick any area 4. If all areas are 🟩/🟦 → include "Hard-mode review" option

Present these as an AskUserQuestion with header "Session" and concise descriptions showing which areas each option targets. The user MUST select before proceeding.

Phase 3: Build Questions

1. Read markdown files in target section(s) 2. If drilling weak area: also read concepts/{area}.md to find 🔴 unresolved concepts — rephrase these in new contexts (don't repeat the same question) 3. Craft exactly 4 questions following references/quiz-rules.md

CRITICAL: Read references/quiz-rules.md before crafting ANY question. Zero hints allowed.

Phase 4: Present Quiz

Use AskUserQuestion:

• 4 questions, 4 options each, single-select
• Header: "Q1. Topic" (max 12 chars)
• Descriptions: neutral, no hints

Phase 5: Grade & Explain

1. Show results table (question / correct answer / user answer / result) 2. Wrong answers: concise explanation 3. Map each question to its area

Phase 6: Update Files

1. Update concept file (concepts/{area}.md)

For each question answered:

• New concept: Add row to table + if wrong, add error note under ### 오답 메모 (or localized equivalent)
• Existing 🔴 concept answered correctly: Increment attempts & correct, change status to 🟢, keep error note (learning history)
• Existing 🟢 concept answered wrong again: Increment attempts, change status back to 🔴, update error note

Table format:

| Concept | Attempts | Correct | Last Tested | Status |
|---------|----------|---------|-------------|--------|
| concept name | 2 | 1 | 2026-02-24 | 🔴 |

Error notes format (only for wrong answers):

### Error Notes

**concept name**
- Confusion: what the user mixed up
- Key point: the correct understanding

2. Update dashboard

• Recalculate per-area stats from concept files (sum attempts/correct across all concepts in that area)
• Update proficiency badges: 🟥 0-39% · 🟨 40-69% · 🟩 70-89% · 🟦 90-100% · ⬜ no data
• Update stats: total questions, cumulative rate, unresolved/resolved counts, weakest/strongest

Dashboard stays compact — no session logs, no per-question details.

Dashboard Template

Create when no dashboard exists. Filename localized to {LANG}. Example in English:

# Learning Dashboard

> Concept-based metacognition tracking. See linked files for details.

---

## Proficiency by Area

| Area | Correct | Wrong | Rate | Level | Details |
|------|---------|-------|------|-------|---------|
(one row per section, last column = [[concepts/{area}]] link)
| **Total** | **0** | **0** | **-** | ⬜ Unmeasured | |

> 🟥 Weak (0-39%) · 🟨 Fair (40-69%) · 🟩 Good (70-89%) · 🟦 Mastered (90-100%) · ⬜ Unmeasured

---

## Stats

- **Total Questions**: 0
- **Cumulative Rate**: -
- **Unresolved Concepts**: 0
- **Resolved Concepts**: 0
- **Weakest Area**: -
- **Strongest Area**: -

Concept File Template

Create per area when first question is asked. Example:

# {Area Name} — Concept Tracker

| Concept | Attempts | Correct | Last Tested | Status |
|---------|----------|---------|-------------|--------|

### Error Notes

(added as concepts are missed)

Important Reminders

• ALWAYS read references/quiz-rules.md before creating questions
• NEVER include hints in option labels or descriptions
• NEVER use "(Recommended)" on any option
• Randomize correct answer position
• After grading, ALWAYS update both concept file AND dashboard
• Communicate in user's language

Quiz Design Rules

Zero-Hint Policy (CRITICAL)

Every question must be answerable ONLY by someone who actually knows the material.

1. Option descriptions: NEVER reveal correctness

• BAD: label: "stderr", description: "Error output stream used by Cloud Run for error classification"
• GOOD: label: "stderr", description: "Standard error stream"

2. No "(Recommended)" tag on any option

3. Randomize correct answer position — never always first or last

4. Question phrasing: Ask about behavior/purpose/output, don't hint at the answer

• BAD: "Which error stream does error() use?"
• GOOD: "Where does error() method output go?"

5. Plausible distractors: Wrong options must be real concepts from the domain, representing common misconceptions

Question Types

1. Factual recall: "What HTTP status code is returned when...?" 2. Conceptual understanding: "Why does the system use X pattern?" 3. Behavioral prediction: "What happens when X fails?" 4. Comparison/distinction: "What is the difference between X and Y?" 5. Debugging scenario: "Given this error, what is the most likely cause?"

Difficulty Balancing

• Diagnostic: easy 40%, medium 40%, hard 20%
• Weak-area drill: medium 30%, hard 70%
• Review: all levels evenly

Drilling Unresolved Concepts

When targeting 🔴 concepts from concept files:

• Do NOT repeat the exact same question — rephrase in a new context
• Test the same underlying knowledge from a different angle
• E.g., if user confused "400 vs 422", ask a scenario question where they must choose the correct status code for a new situation

AskUserQuestion Format

• 4 questions per round, 4 options each, single-select
• Header: max 12 chars, "Q1. Topic"

File Update Protocol

After grading: 1. Update concepts/{area}.md — add/update concept rows + error notes 2. Update dashboard — recalculate area stats from concept files 3. Badges: 🟥 0-39% · 🟨 40-69% · 🟩 70-89% · 🟦 90-100% · ⬜ no data

Language Rule

All file content and output in the user's detected language. Badge emojis are universal.