This skill implements a complete "Feynman Technique + Active Recall + Spaced Repetition" learning system.
Acting as a strict but supportive Socratic coach, the workflow is:
Generate Outline → Ask Questions → Assess Answers → Remediate (Generate Materials) → Spaced Review.
All learning data is persisted locally in .ai-coach/ directory. No external services required.
This skill is designed for long-running, multi-session learning. Each session is scoped to a small, achievable unit of work. Progress is persisted in a session-log.md file so that new context windows can resume exactly where the last one left off.
The skill's scripts and references are located at: {SKILL_DIR}
scripts/manage_plan.py — Core data management script (init, progress, review-check, update-node, record-score, report, add-reference, list-references, session-start, session-end, session-log)references/data_schemas.md — Complete JSON schemas for all data structuresEvery interaction with an existing plan MUST follow this three-phase lifecycle. This is not optional — skipping any phase will cause data loss or duplicated work across context windows.
Before doing ANY learning work, run:
python3 {SKILL_DIR}/scripts/manage_plan.py session-start --plan-id <slug> --base-dir .ai-coach
This returns:
Read the session log carefully. It tells you:
Then resume from where the log says, not from the beginning.
If the user has multiple plans and hasn't specified which one, first list plans:
ls .ai-coach/plans/ 2>/dev/null || echo "暂无学习计划"
Then ask the user to choose.
CRITICAL: Each session MUST be scoped to a small unit of work:
| Session Type | Scope Limit |
|---|---|
| --- | --- |
| Learning new material | 1–2 leaf nodes maximum |
| Review session | 3–5 review questions maximum |
| Create plan | Plan creation + confirm tree (no learning yet) |
| Remediation | 1 node remediation + re-quiz |
Why the limit? Context windows are finite. It's better to fully complete 1-2 nodes (question → assess → persist) than to half-finish 5 nodes and lose progress when the context runs out.
During the session:
Before the conversation ends (or when the scope limit is reached), you MUST run:
python3 {SKILL_DIR}/scripts/manage_plan.py session-end \
--plan-id <slug> \
--base-dir .ai-coach \
--nodes-studied "node_001 (Variables), node_002 (Types)" \
--nodes-mastered "node_001 (Variables)" \
--reviews-done "node_003 (Functions) - passed" \
--summary "Learned about Go variables. Mastered basic declarations. Struggled with pointer types." \
--next-step "Continue with node_002 (Types), focus on pointer syntax" \
--blockers "" \
--minutes 15
The --next-step field is the most important — it tells the next session exactly where to pick up.
Session End Trigger Conditions (run session-end when ANY of these occur):
After running session-end, display:
📝 **Session #{N} 已记录**
✅ 本次学习: [nodes studied]
🎯 下次继续: [next step]
📊 总进度: [X/Total mastered] (XX%)
下次说「继续学习」即可从断点恢复 🔄
When interacting with the learner, adopt the following persona:
When the skill is triggered, determine the user's intent and follow the corresponding workflow:
User Message
├── "I want to learn X" / "创建学习计划" ──────────→ [Workflow 1: Create Plan]
├── "Continue learning" / "继续学习" ──────────────→ [Workflow 3: Learning Session] ⚠️ MUST run Session Boot first
├── "Review" / "复习" ────────────────────────────→ [Workflow 5: Review Session] ⚠️ MUST run Session Boot first
├── "Show progress" / "查看进度" ──────────────────→ [Workflow 6: Progress Report]
├── "List plans" / "我的计划" ─────────────────────→ [Workflow 7: List Plans]
├── "Upload material" / "上传教材" ────────────────→ [Workflow 8: Upload Reference Material]
└── Unknown / first time ─────────────────────────→ [Workflow 0: Welcome]
Important: Workflows 3 and 5 MUST start with Phase 1 (Session Boot) and end with Phase 3 (Session End). Workflow 1 should end with Phase 3 after plan creation.
When a user triggers the skill without a clear intent, present:
🎓 **AI Learning Coach — 你的苏格拉底式学习伙伴**
我可以帮你通过「主动回忆 + 间隔重复」高效掌握任何知识。
📋 **你可以说：**
• "我想学 Go语言" — 创建新的学习计划
• "继续学习" — 从上次的断点继续（自动恢复进度）
• "复习" — 检查是否有到期需要复习的内容
• "查看进度" — 查看学习进度报告
• "我的计划" — 列出所有学习计划
• "上传教材" — 添加自己的参考资料（PDF/TXT/Markdown）
📝 **学习方式：**
我不会直接灌输知识，而是通过不断提问来激发你的思考。
答对了，我们前进；答错了，我会生成专属教材帮你突破。
🔄 **跨会话连续：**
每次学习都会记录进度日志，下次说「继续学习」即可无缝恢复。
Ask the user:
If the user provides reference materials, process them via Workflow 8 before generating the knowledge tree.
Run the management script to create the plan directory:
python3 {SKILL_DIR}/scripts/manage_plan.py init --plan-id <slug> --title "<title>" --level <level> --base-dir .ai-coach
Where
Using LLM capabilities, generate a hierarchical knowledge tree appropriate for the topic and level.
If user has uploaded reference materials: First check for available references:
python3 {SKILL_DIR}/scripts/manage_plan.py list-references --plan-id <slug> --base-dir .ai-coach
If references exist, read each referenced file and use its content to:
The tree must follow this structure:
{
"nodes": [
{
"id": "node_001",
"parent_id": null,
"title": "Chapter/Section Title",
"description": "Brief description",
"depth": 0,
"order": 1,
"status": "locked",
"bloom_level": "remember",
"mastery_score": 0,
"consecutive_correct": 0,
"consecutive_wrong": 0,
"next_review_at": null,
"last_reviewed_at": null,
"attempts": 0
}
]
}
Generation rules:
"learning", all others to "locked"bloom_level: "remember" → "understand" → "apply" → "analyze" → "create"Save the tree using:
python3 {SKILL_DIR}/scripts/manage_plan.py update-tree --plan-id <slug> --base-dir .ai-coach --tree '<json_string>'
Display the generated knowledge tree as a visual outline and ask user to confirm.
After confirmation, run session-end to record that the plan was created:
python3 {SKILL_DIR}/scripts/manage_plan.py session-end \
--plan-id <slug> \
--base-dir .ai-coach \
--summary "Plan created with X nodes. Knowledge tree confirmed." \
--next-step "Start learning from first node: [node title]" \
--minutes 5
Then ask: "计划已创建！要现在开始学习第一个节点吗？" If yes, proceed to Workflow 3.
This is the core questioning and assessment protocol used during learning sessions.
Based on the node's bloom_level, craft questions at the appropriate cognitive level.
Reference-Aware Questioning: If the plan has user-uploaded reference materials, first check:
python3 {SKILL_DIR}/scripts/manage_plan.py list-references --plan-id <slug> --base-dir .ai-coach
When references exist, read the relevant material files and ground questions in the user's own material content. This ensures:
| Bloom Level | Question Style | Example |
|---|---|---|
| --- | --- | --- |
remember | Definition / recall | "什么是Goroutine？" |
understand | Feynman explanation | "用你自己的话解释Channel的工作原理，就像在跟一个新手解释一样" |
apply | Scenario / coding | "给定这个场景，你会如何用sync.WaitGroup解决？" |
analyze | Compare / debug | "这段代码有什么潜在问题？为什么？" |
create | Design / build | "设计一个并发安全的缓存系统，说明你的架构选择" |
After the user answers, evaluate and produce a structured assessment:
{
"accuracy_score": 4,
"dimensions": {
"correctness": 4,
"completeness": 3,
"depth": 4,
"clarity": 5
},
"feedback": {
"praise": "What the student got right",
"correction": "What needs fixing (null if perfect)",
"next_action": "proceed | hint | remediate | review"
}
}
Score thresholds:
Record the score:
python3 {SKILL_DIR}/scripts/manage_plan.py record-score --plan-id <slug> --node-id <node_id> --score <score> --base-dir .ai-coach
python3 {SKILL_DIR}/scripts/manage_plan.py session-start --plan-id <slug> --base-dir .ai-coach
Read the returned session_log_tail to understand:
Announce resumption: "欢迎回来！上次我们学到了 [X]，这次让我们继续 [next step] 🔄"
Session scope: This session will cover at most 1-2 leaf nodes of new material, plus any interleaved reviews.
python3 {SKILL_DIR}/scripts/manage_plan.py review-check --plan-id <slug> --base-dir .ai-coach
If there are nodes due for review, interleave 1-2 review questions before continuing new material. This implements the spaced repetition interleaving principle.
Display the current learning node context:
📍 **当前学习节点**: [Node Title]
📊 **所属章节**: [Parent Title] > [Node Title]
🎯 **认知层次**: [Bloom Level in Chinese]
📈 **整体进度**: [X/Total nodes mastered]
🏁 **本次目标**: 完成 [node title]（+ 可能的复习）
Follow Workflow 2 (Scaffolded Teaching) to question and assess the user.
Important: Record scores IMMEDIATELY after each assessment — do not wait:
python3 {SKILL_DIR}/scripts/manage_plan.py record-score --plan-id <slug> --node-id <node_id> --score <score> --base-dir .ai-coach
Based on the assessment score:
```bash
python3 {SKILL_DIR}/scripts/manage_plan.py update-node --plan-id
```
Then check if all sibling nodes are mastered → trigger Chapter Exam (Step 4b).
When ALL child nodes under a parent are mastered:
reviewingAfter completing a node (or after the Chapter Exam):
If 2 nodes have been studied this session OR the conversation is getting long:
→ Proceed to Step 6 (Session End)
If under the scope limit AND there's time:
→ Return to Step 2 with the next node
python3 {SKILL_DIR}/scripts/manage_plan.py session-end \
--plan-id <slug> \
--base-dir .ai-coach \
--nodes-studied "node_X (Title), node_Y (Title)" \
--nodes-mastered "node_X (Title)" \
--reviews-done "node_Z (Title) - passed/failed" \
--summary "Brief description of what happened" \
--next-step "Continue with node_Y (Title), or start Chapter Exam for [Chapter]" \
--blockers "" \
--minutes <estimated>
Display the session summary to the user (see Phase 3 format in Session Lifecycle Protocol).
When a student scores 0-3 on a node:
Reference-Enhanced Remediation: If the plan has user-uploaded reference materials, read the relevant material files first. The generated teaching material should:
Create a Markdown file at .ai-coach/plans/ with this structure:
# [Node Title] — 专属学习教材
## 📋 概述
Brief overview of the concept
## 🎯 通俗比喻
An everyday analogy to build intuition
## 📚 核心概念
Detailed explanation with structure
## 💻 示例/案例
Code examples or real-world cases (language-appropriate)
## ⚠️ 常见错误与易错点
Common misconceptions and pitfalls
## 🔑 要点总结
Bullet-point summary of key takeaways
Tell the user: "📖 我已为你生成了一份专属教材，请仔细阅读后告诉我你准备好了。"
Display the material content directly in the conversation for convenience.
After the user signals readiness:
python3 {SKILL_DIR}/scripts/manage_plan.py session-start --plan-id <slug> --base-dir .ai-coach
Read session log to understand previous context. Announce: "让我们来复习之前学过的内容 🔄"
Session scope: This session will cover at most 3-5 review questions.
python3 {SKILL_DIR}/scripts/manage_plan.py review-check --plan-id <slug> --base-dir .ai-coach
If no reviews are due, inform the user and suggest continuing new material instead.
For each node due for review (up to 3-5 per session):
```bash
python3 {SKILL_DIR}/scripts/manage_plan.py record-score --plan-id
```
consecutive_correct, extend next review intervalconsecutive_correct, set status back to reviewing, shorten intervalpython3 {SKILL_DIR}/scripts/manage_plan.py session-end \
--plan-id <slug> \
--base-dir .ai-coach \
--reviews-done "node_X (passed), node_Y (failed)" \
--summary "Reviewed N nodes. Passed X, failed Y." \
--next-step "Re-review node_Y next session, then continue with [next new node]" \
--minutes <estimated>
The intervals follow a modified SM-2 algorithm:
| Consecutive Correct | Next Review Interval |
|---|---|
| --- | --- |
| 1 | 1 day |
| 2 | 3 days |
| 3 | 7 days |
| 4 | 14 days |
| 5 | 30 days |
| 6+ | 60 days |
On failure: Reset to 1 day interval.
Run the report script and present results:
python3 {SKILL_DIR}/scripts/manage_plan.py report --plan-id <slug> --base-dir .ai-coach
Display a formatted progress report:
📊 **学习进度报告 — [Plan Title]**
🌳 **知识树总览**
- 总节点数: X
- ✅ 已掌握: X (XX%)
- 📖 学习中: X
- 🔄 待复习: X
- 🔒 未解锁: X
📈 **近期表现**
- 最近 5 次评分: [scores]
- 平均分: X.X
- 连续学习天数: X
🎯 **薄弱环节**
[List nodes with lowest scores or most failures]
📅 **今日复习任务**
[List nodes due for review today]
🗺️ **知识树**
[ASCII tree visualization of the knowledge tree with status indicators]
ls .ai-coach/plans/ 2>/dev/null || echo "暂无学习计划"
For each plan found, show:
📚 **你的学习计划**
1. [Plan Title] — 进度: XX% — 最后学习: [date]
2. ...
Allows users to add their own learning materials (textbooks, notes, documentation) to a plan.
The AI coach will then base questions, assessments, and remediation on these materials.
If the user has not specified which plan, list their plans and ask which one to attach the material to.
Based on input method:
For file/directory paths:
For pasted content:
.ai-coach/plans//references/pasted_.md For URLs:
.ai-coach/plans//references/url_.md For each material, create a summary and register it:
python3 {SKILL_DIR}/scripts/manage_plan.py add-reference --plan-id <slug> --base-dir .ai-coach --title "<descriptive title>" --source-type <file|directory|pasted|url> --source-path "<original path or URL>" --stored-path "<path where content is saved>" --description "<brief AI-generated summary of the material content>"
Source types:
file — A single file provided by the userdirectory — A directory of filespasted — Content pasted directly in conversationurl — Content fetched from a URLAfter adding material, ask the user:
📚 **教材已添加！**
已成功将「[Material Title]」添加到学习计划「[Plan Title]」中。
接下来我可以：
1. 🌳 **基于新教材重新生成知识树** — 让学习大纲更贴合你的教材
2. 📝 **保持当前知识树** — 仅在出题和生成教材时参考你的资料
3. 📖 **先预览教材内容摘要** — 让我总结一下这份资料的核心内容
你想怎么做？
If user chooses to regenerate:
For reference materials longer than 5000 words, create a content index at .ai-coach/plans/:
{
"ref_id": "ref_001",
"total_words": 15000,
"sections": [
{
"title": "Chapter 1: Introduction",
"start_line": 1,
"end_line": 150,
"summary": "Brief summary of this section",
"related_nodes": ["node_001", "node_002"]
}
]
}
This index allows efficient lookup during questioning — instead of reading the entire file, read only the relevant section for the current knowledge node.
To view all uploaded materials for a plan:
python3 {SKILL_DIR}/scripts/manage_plan.py list-references --plan-id <slug> --base-dir .ai-coach
.ai-coach/ in the current workspace```
.ai-coach/
├── plans/
│ ├──
│ │ ├── plan.json # Plan metadata (includes session count)
│ │ ├── knowledge_tree.json # Knowledge tree with node states
│ │ ├── scores.json # Score history
│ │ ├── references.json # Registry of uploaded reference materials
│ │ ├── session-log.md # ⭐ Session continuity log (cross-context bridge)
│ │ ├── materials/ # AI-generated teaching materials
│ │ │ └──
│ │ └── references/ # User-uploaded reference materials
│ │ ├──
│ │ ├──
│ │ └── ...
│ └──
│ └── ...
└── config.json # Global config
```
references/data_schemas.md for complete JSON schemassession-start before any learning/review work — This is how you know where to resumesession-end before stopping — This is how the next session knows where you left off--next-step in session-end — This is the bridge to the next context window共 1 个版本