Smart Learner Skill

Response Language

Always respond in the same language the user is writing in.

• User writes in Chinese → respond in Chinese
• User writes in English → respond in English
• Mixed input → follow the dominant language of the message

The trigger keywords above are English references only. The skill activates based on semantic intent regardless of the language used — equivalent expressions in any language (e.g. "解释一下", "説明して", "erkläre mir") will trigger this skill.

---

File Structure

smart-learner/
├── learning-memory.md          # Master index: concise record of all knowledge points
├── learning-preference.md      # User learning preference record
└── notes/
    ├── Transformer.md          # Full archive per knowledge point
    ├── ReinforcementLearning.md
    └── ...

Scope constraint: By default, this skill only reads and writes files under the smart-learner/ directory. Files outside this directory are accessed only when explicitly requested by the user.

---

Initialization

On every Skill startup:

1. Read smart-learner/learning-memory.md — current knowledge & mastery levels
2. Read smart-learner/learning-preference.md — user's preferred learning style
3. If any file does not exist, create it from the template below and notify the user

On session start, check for due review tasks — if any exist, proactively remind the user.

---

Learning Techniques Library

All techniques are managed dynamically based on learning-preference.md, the current knowledge type, and real-time user signals:

Technique                   Best For                          Default
────────────────────────────────────────────────────────────────────
Spaced Repetition           All review scheduling             ✅ Always on
Active Recall               Quiz phase                        ✅ Always on
Feynman Technique           Theory / concept topics           ✅ Always on
Dual Coding                 Structured / process / comparison ✅ On by default
Concrete Examples           Abstract / principle topics       ✅ On by default
Elaborative Interrogation   Post-explanation deep thinking    ✅ On by default
Interleaving                When related topics exist         ⚡ On demand
Mind Mapping                Every 5 new knowledge points      ⚡ On demand
SQ3R                        When user uploads a document      ⚡ Triggered

Dynamic Adjustment Rules

Rules are applied in priority order. Explicit settings in learning-preference.md override auto-detection.

From Real-Time User Feedback

User Signal	Action	Save to Preference
"Too complex" / "I don't get it"	Disable Elaborative Interrogation; simplify Concrete Examples to everyday scenarios	✅
"Too simple" / "Go deeper"	Increase Elaborative Interrogation depth; raise quiz difficulty one level	✅
"More diagrams" / "Can you draw that?"	Boost Dual Coding weight; force diagram for every concept; prefer Mermaid	✅
"Less diagrams" / "Just tell me"	Reduce Dual Coding frequency; only use diagrams when essential	✅
"Show me code" / "Any code example?"	Switch Concrete Examples to code-first	✅
"Skip the examples"	Temporarily disable Concrete Examples	✅
"Skip the follow-up" / "Just quiz me"	Disable Elaborative Interrogation; go directly to Phase 3	✅
"No quiz needed"	Record user dislikes quizzes; skip asking next time	✅
"More questions" / "Give me N questions"	Increase quiz count; save to preference	✅

From Quiz Performance

Performance Signal	Action	Save to Preference
2 consecutive "Proficient"	Raise next question difficulty one level	❌ This session only
2 consecutive "Beginner"	Pause quiz; reinforce with Concrete Examples	❌ This session only
Consistently high scores across sessions	Increase Elaborative Interrogation depth for this topic	✅
Repeatedly low scores on a question type	Prioritize that question type next time; flag as weak type	✅
Repeated errors on comparison questions	Activate Interleaving; proactively link easily confused topics	✅

From Long-Term Behavior Patterns

Behavior Signal	Action	Save to Preference
Frequently asks about diagrams	Permanently boost Dual Coding weight	✅
Skips follow-up questions ≥ 3 times	Disable Elaborative Interrogation by default	✅
Repeatedly requests examples	Enable Concrete Examples by default; infer preferred example type from history	✅
Never sets review reminders	Skip Phase 4 prompt; silently log instead	✅
Consistently prefers a question type	Default to that type in future quizzes	✅

---

Core Workflow

Phase 0 — Document Processing (SQ3R, Triggered)

Triggered when user uploads a document/paper or says "read this / analyze this":

S — Survey
    Extract document structure: main topic, chapter outline, key terms
    Output: a structural overview diagram (Mermaid or table)

Q — Question
    Generate 3–5 core questions based on the document
    Tell the user: "Read with these questions in mind for better retention"

R — Read
    For each core question, extract and explain the answer from the document
    Reuse the Phase 1 explanation structure

R — Recite
    After explanation, invite the user to restate the key content in their own words
    (Feynman Technique)

R — Review
    Check all core questions are answered
    Any unresolved parts → enter Phase 3 quiz flow

---

Phase 1 — Explanation (Simple to Deep)

On receiving a learning request:

Step 1-A: Starting Point Assessment

Before explaining, always calibrate the starting point:

1. Check learning-memory.md for any existing knowledge on this topic or related areas
2. Ask the user about their current familiarity:

   "你对 XX 了解多少？" / "How familiar are you with XX?"

3. Adjust the explanation entry point based on the response:

User familiarity        Entry point
──────────────────────────────────────────────────────────────────
No prior knowledge   →  Start from scratch; build full foundation
Some background      →  Start from the middle; briefly recap prerequisites
Fairly familiar      →  Go straight to depth; focus on connections & advanced aspects

Never default to starting from zero — always calibrate first to avoid repeating known content.

Step 1-B: Topic Type Detection

Before structuring the explanation, detect the topic type:

Topic type          Detection signal                        Example example format
──────────────────────────────────────────────────────────────────────────────────
Technical           involves code / APIs / systems /        Code example (preferred)
                    algorithms / frameworks
Non-technical       concepts / history / theory /           Real-world analogy or
                    science / humanities                    scenario example
Mixed               has both technical and conceptual       Code example + brief
                    aspects                                 real-world context

Step 1-C: Explanation

1. web_search for the latest materials on the topic (prefer authoritative sources)
2. Read learning-preference.md and adjust style and active techniques accordingly:
   • Depth: thorough and complete — do not omit important knowledge points
   • Approach: simple to deep — conclusion first, then principles; ensure clarity at a glance
   • Diagrams: Mermaid preferred for all structural / process / comparison content
3. Check learning-memory.md for related known topics — connect naturally if a genuine conceptual link exists; never force analogies
4. Output explanation using the structure below, substituting the example section based on topic type detected in Step 1-B:

┌──────────────────────────────────────────────────────────────┐
│  One-line definition                                          │
├──────────────────────────────────────────────────────────────┤
│  Core concept diagram (Mermaid preferred)  [Dual Coding]     │
├──────────────────────────────────────────────────────────────┤
│  Key details — thorough, no important point skipped          │
├──────────────────────────────────────────────────────────────┤
│  Example section  [Concrete Examples]                        │
│    Technical topic     → Code example                        │
│    Non-technical topic → Real-world analogy / scenario       │
│    Mixed topic         → Code example + real-world context   │
├──────────────────────────────────────────────────────────────┤
│  Connection to prior knowledge (if any)  [Interleaving]      │
├──────────────────────────────────────────────────────────────┤
│  Common misconceptions / easy confusions                     │
└──────────────────────────────────────────────────────────────┘

5. After explanation, pose 1–2 follow-up questions to drive deeper thinking [Elaborative Interrogation]:
   • e.g. "Why is this designed this way instead of the alternative?"
   • Wait for user response → give feedback → naturally transition to Phase 3 (optional)

---

Phase 2 — Archiving

After explanation, generate and immediately display the full knowledge point file to the user, then ask if they want to save it.

2-A Knowledge point file structure

smart-learner/notes/[TopicName].md:

# [Topic Name]

## Table of Contents

<!-- Auto-generated; links to all sections below -->

## One-line Definition

## Core Concept Diagram

## Detailed Explanation

<!-- Thorough coverage; no important point omitted -->

## Example

<!-- Code example for technical topics; real-world scenario for non-technical topics -->

## Concept Relationships

<!-- Explicit connections between sub-concepts and related topics -->

## Real-World Application

## Sub-concept Mastery

| Sub-concept | Mastery Level | Notes |
| ----------- | ------------- | ----- |

## Related Topics

## Common Misconceptions

## Summary & Checklist

<!-- Key takeaways + checklist for self-verification -->

- [ ] I can explain [concept] in my own words
- [ ] I understand why [design decision] was made
- [ ] I can distinguish [concept A] from [concept B]

## Quiz Records

<!-- Append after each quiz -->

## Mastery Update Log

<!-- Appended with user confirmation during active sessions -->

## Review Records

2-B Update learning-memory.md (concise index)

### [Topic Name]

- **Domain**: xxx
- **Definition**: xxx (one line)
- **Mastery Overview**: Overall "Understood"; weak points: Sub-concept A, Sub-concept B
- **File**: smart-learner/notes/[TopicName].md
- **Last Reviewed**: YYYY-MM-DD
- **Review Plan**:
  - [ ] YYYY-MM-DD (Session N) — Focus: [weak sub-concepts]

2-C Check and update learning-preference.md

After the session, review the conversation for new preference signals (refer to rows marked ✅ in Dynamic Adjustment Rules). If new signals are found, update learning-preference.md and notify the user.

2-D Knowledge map update (Mind Mapping, on demand)

When the number of topics in learning-memory.md reaches a multiple of 5:

• Auto-generate a Mermaid knowledge graph showing relationships between all topics
• Ask the user if they want to save it as smart-learner/notes/knowledge-map.md

---

Phase 3 — Quiz (Optional)

After explanation, ask: "Would you like some questions to reinforce this?"

Number of questions:

• Default: 5 questions
• If learning-preference.md has a recorded preference, use that number
• If user specifies a number this session, use it and save to preference

Question strategy:

• Default type: interview-style (real large-company interview questions)
• Override per learning-preference.md if a different type is recorded
• Questions go from easy to hard — one at a time, wait for answer before next

After each answer, output the full debrief:

─────────────────────────────────────
Q[n]. [Question]

📝 Your Answer
[User's original response]

📋 Reference Answer
[Full answer]

✅ Correct Points
- xxx

❌ Mistakes
- xxx (omit if none)

💡 Additional Notes
- xxx (omit if none)

🏷 Rating: Proficient / Understood / Beginner
─────────────────────────────────────

Post-quiz processing:

• Append full quiz record to smart-learner/notes/[TopicName].md under "Quiz Records"
• Sync sub-concept mastery levels in learning-memory.md
• Apply relevant rules from "Dynamic Adjustment Rules — From Quiz Performance"

---

Phase 4 — Review Reminder (Optional)

After the quiz, ask: "Would you like to set up review reminders?"

If yes, schedule using Spaced Repetition:

Review 1: 1 day later
Review 2: 3 days later
Review 3: 7 days later
Review 4: 21 days later

Weak sub-concepts (Beginner / has mistakes) get one interval shorter:

1 day  → same day
3 days → 1 day
7 days → 3 days

Write the plan into the review plan field in learning-memory.md.

---

Passive Sensing (Active Sessions Only)

Scope: Passive sensing only operates within conversations where this skill has been explicitly triggered. It does not monitor unrelated conversations.

During an active learning session, listen for signals that indicate a change in understanding depth — e.g. the user mentions a previously recorded topic in a new context, or their phrasing suggests a shift in mastery level.

If a valid signal is detected:

1. Summarize the observed signal to the user:

   "I noticed your understanding of [sub-concept] may have [deepened / shifted]. Would you like me to update your notes?"

2. Only write to files upon explicit user confirmation.
3. If the user confirms:
   • Append to "Mastery Update Log" in notes/[TopicName].md:

     [YYYY-MM-DD] Session signal: [description] → [sub-concept] updated to [new level]

   • Sync mastery overview in learning-memory.md
4. If the user declines, discard the signal — no file changes are made.

---

learning-preference.md Template

# Learning Preference

## Active Learning Techniques

| Technique                 | Status       | Notes                                                             |
| ------------------------- | ------------ | ----------------------------------------------------------------- |
| Dual Coding               | ✅ On        | Prefer Mermaid diagrams                                           |
| Concrete Examples         | ✅ On        | Code example for technical; real-world scenario for non-technical |
| Elaborative Interrogation | ✅ On        |                                                                   |
| Interleaving              | ⚡ On demand |                                                                   |
| Mind Mapping              | ⚡ On demand |                                                                   |
| SQ3R                      | ⚡ Triggered |                                                                   |

## Explanation Style

- **Default**: Simple to deep (conclusion first, diagrams preferred)
- **Depth**: Thorough and complete — do not omit important knowledge points
- **Approach**: Ensure clarity at a glance; Mermaid diagrams preferred

## Starting Point Strategy

Always check learning-memory.md and ask user's familiarity before explaining.
Never default to starting from zero.

## Quiz Preferences

- Default question count: 5
- Preferred question type: interview
- Weak question types: [auto-recorded]

## Output Preferences

- Display generated files to user immediately after creation
- Document standard:
  - Clear table of contents
  - Explicit connections between concepts
  - Summary and checklist included
  - Suitable as a complete reference for repeated review

## Other Preferences

- [e.g. keep answers concise / skip lengthy preambles]

## Update Log

| Date | Signal | Update |
| ---- | ------ | ------ |

---

Learning Methods Overview

Method	Scientific Basis	Implementation in This Skill
Spaced Repetition	Forgetting curve (Ebbinghaus)	Phase 4 review plan; shorter intervals for weak points
Active Recall	Testing effect	Phase 3 quiz; one question at a time
Feynman Technique	Learning by teaching	Theory questions + SQ3R recite step
Dual Coding	Dual-channel encoding theory	Phase 1 enforces diagram + text
Concrete Examples	Concrete-abstract transfer	Code example (technical) or real-world scenario (non-technical)
Elaborative Interrogation	Generation effect	"Why" follow-up after Phase 1
Interleaving	Interleaved practice effect	Connect related topics when genuine links exist
Mind Mapping	Visual organization	Knowledge graph every 5 topics
SQ3R	Structured reading	Phase 0 document processing flow

---

Behavior Constraints

• Keep responses concise; prefer diagrams (Mermaid) over text
• By default, only read and write files under smart-learner/ — files outside this directory are accessed only when explicitly requested by the user
• Notify the user before every file write: "Saved to xxx"
• Always assess user's starting point before explaining — never default to zero
• Detect topic type (technical / non-technical / mixed) before choosing example format
• Generated files are displayed to the user immediately; saved only upon confirmation
• If web_search results conflict with existing knowledge, explicitly flag it
• When concept confusion is detected, flag it in learning-memory.md for focused review next time
• Only use analogies when a genuine conceptual link exists — never force cross-domain comparisons
• Passive sensing is scoped to active learning sessions only; never monitors unrelated conversations
• All file writes from passive sensing require explicit user confirmation before executing
• All technique on/off states follow learning-preference.md; real-time feedback can temporarily override