スキル一覧に戻る

teach

pjt222
更新日 2 days ago
3 閲覧
17
2
17
GitHubで表示
その他ai

について

`teach`スキルは、学習者のレベルを評価し、既知の概念から未知の概念へと段階的に説明を構築することで、体系的な知識伝達を提供します。ソクラテス的問答法を用いて理解を確認し、ユーザーのフィードバックに基づいて指導を適応させます。このスキルは、ユーザーが「Xはどのように機能しますか?」と尋ね、特に以前の説明が失敗した場合や前提知識にギャップがある概念について、段階的な説明を必要とする際に使用してください。

クイックインストール

Claude Code

推奨
メイン
npx skills add pjt222/agent-almanac -a claude-code
プラグインコマンド代替
/plugin add https://github.com/pjt222/agent-almanac
Git クローン代替
git clone https://github.com/pjt222/agent-almanac.git ~/.claude/skills/teach

このコマンドをClaude Codeにコピー&ペーストしてスキルをインストールします

ドキュメント

Teach

Structured knowledge transfer — assess learner, scaffold known → unknown, calibrated explanation, check via Q, adapt to feedback, reinforce via practice.

Use When

  • User asks "how does X work?" → graduated explanation needed, not data dump
  • Q's reveal gap between current understanding + need
  • Prev explanations didn't land — confused | asking same Q diff way
  • Teaching concept w/ prereqs user may not have
  • After learn built deep mental model needing communication

In

  • Required: Concept, system, skill to teach
  • Required: Learner (implicit — user in conv)
  • Optional: Known context (expertise, background, goals)
  • Optional: Prev failed explanations (what tried)
  • Optional: Time/depth constraint (quick overview vs deep)

Do

Step 1: Assess — Map Learner

Before explaining, determine what learner knows + needs.

Learner Calibration Matrix:
┌──────────────┬────────────────────────────┬──────────────────────────┐
│ Level        │ Explanation Pattern         │ Check Pattern            │
├──────────────┼────────────────────────────┼──────────────────────────┤
│ Novice       │ Analogy-first. Connect to  │ "In your own words, what │
│ (no domain   │ familiar concepts. Avoid   │ does X do?" Accept any   │
│ vocabulary)  │ jargon entirely. Concrete  │ correct paraphrase.      │
│              │ before abstract.           │                          │
├──────────────┼────────────────────────────┼──────────────────────────┤
│ Intermediate │ Build on existing vocab.   │ "What would happen if    │
│ (knows terms,│ Fill gaps with targeted    │ we changed Y?" Tests     │
│ some gaps)   │ explanations. Use code     │ whether they can predict │
│              │ examples that are close    │ from understanding.      │
│              │ to their existing work.    │                          │
├──────────────┼────────────────────────────┼──────────────────────────┤
│ Advanced     │ Skip fundamentals. Focus   │ "How would you compare   │
│ (strong base,│ on nuance, trade-offs,     │ X to Z approach?" Tests  │
│ seeks depth) │ edge cases. Reference      │ integration and judgment. │
│              │ source material directly.  │                          │
├──────────────┼────────────────────────────┼──────────────────────────┤
│ Misaligned   │ Correct gently. Provide    │ "Let me check my under-  │
│ (confident   │ the right model alongside  │ standing — you're saying  │
│ but wrong)   │ why the wrong model feels  │ X?" Mirror back to       │
│              │ right. No shame signals.   │ surface the mismatch.    │
└──────────────┴────────────────────────────┴──────────────────────────┘
  1. Review user's Q's, vocab, goals
  2. Classify likely level for THIS topic (advanced in one, novice in another)
  3. ID Zone of Proximal Dev (ZPD): just beyond reach but achievable w/ support
  4. Note misconceptions to address before correct model lands
  5. ID best entry: what they know connecting to need

Got: Clear picture: what learner knows, needs, what bridge connects. Specific enough to choose explanation strategy.

If err: Level unclear → calibration Q: "Familiar w/ [prereq]?" Not test, gather data. Awkward → default intermediate, adjust on response.

Step 2: Scaffold — Bridge Known → Unknown

Build path from known → new concept.

  1. ID anchor: 1 concept learner definitely understands related to target
  2. State connection explicit: "X (you know) works like Y in this new context because..."
  3. 1 new idea at a time — never 2 in same sentence
  4. Concrete examples before abstract principles
  5. Layered complexity: simple first, then nuance
  6. Prereqs missing → teach prereq first (mini-scaffold) before main

Got: Scaffolded path, each step builds on prev. Learner never lost — each new idea connects to existing.

If err: Gap too large for single scaffold → break into smaller steps. No familiar anchor (entirely novel) → analogy to diff domain known. Imperfect → acknowledge limits: "Like X, except for..."

Step 3: Explain — Calibrate Depth + Style

Right level, right mode.

  1. Open w/ core idea in 1 sentence — headline before article
  2. Expand w/ scaffolded explanation (Step 2)
  3. Learner's vocab, not domain jargon (unless advanced)
  4. Code: minimal working example, not comprehensive
  5. Abstract: concrete instance first, then generalize
  6. Processes: walk specific case step-by-step before general rules
  7. Monitor confusion signs: next Q doesn't build on explanation → didn't land

Got: Explanation neither too shallow (leaves Q's) nor too deep (overwhelms). Uses their language, connects to context.

If err: Too long → core idea buried, restate 1-sentence headline. More confused after → entry wrong, try diff anchor/analogy. Genuinely complex → acknowledge vs hide: "3 parts, they interact. Start w/ first."

Step 4: Check — Verify Understanding

Don't assume worked. Test via Q's revealing mental model.

  1. Ask Q requiring application not recall: "Given X, what would you expect?"
  2. Ask paraphrase: "Explain back in your own words?"
  3. Present variation: "What if we changed this one thing?"
  4. Look for specific understanding: predict, not just repeat?
  5. Answer reveals misconception → note specific err for Step 5
  6. Correct → push slightly further: can generalize?

Got: Reveals working mental model vs parroting. Working model handles variations; memorized cannot.

If err: Can't answer → explanation didn't build right model. Not their failure — feedback on teaching. Note what didn't land → Step 5.

Step 5: Adapt — Respond to Feedback

Adjust based on check.

  1. Solid → reinforce (Step 6) | advance to next concept
  2. Specific misconception → address direct w/ evidence, not repetition
  3. General confusion → completely diff explanation approach
  4. Ahead of assessment → accelerate, skip scaffolding → nuance
  5. Behind → slow down, teach missing prereq
Adaptation Responses:
┌──────────────────┬─────────────────────────────────────────────────┐
│ Signal           │ Adaptation                                       │
├──────────────────┼─────────────────────────────────────────────────┤
│ "I think I get   │ Push gently: "Great — so what would happen      │
│ it"              │ if...?" Verify before moving on.                 │
├──────────────────┼─────────────────────────────────────────────────┤
│ "I'm confused"   │ Change modality: if verbal, show code. If code, │
│                  │ use analogy. If analogy, draw a diagram.         │
├──────────────────┼─────────────────────────────────────────────────┤
│ "But what about  │ Good sign — they are testing the model. Address  │
│ [edge case]?"    │ the edge case, which deepens understanding.      │
├──────────────────┼─────────────────────────────────────────────────┤
│ "That doesn't    │ They have a competing model. Explore it: "What   │
│ seem right"      │ do you think happens instead?" Reconcile the two.│
├──────────────────┼─────────────────────────────────────────────────┤
│ Silence or       │ They may be processing, or lost. Ask: "What      │
│ topic change     │ part feels least clear?" Lower the bar gently.   │
└──────────────────┴─────────────────────────────────────────────────┘

Got: Teaching adapts in real time. No identical repetition — each retry diff approach. Responsive not mechanical.

If err: Multiple adaptations fail → may be missing prereq so fundamental neither party ID'd. Ask explicit: "What part feels biggest jump?" Often reveals hidden gap.

Step 6: Reinforce — Practice

Solidify via application, not repetition.

  1. Practice problem requiring concept (not trick)
  2. Coding context → small modification to existing code using concept
  3. Conceptual → present scenario, apply model
  4. Connect forward: "Now you understand X, this connects to Y, can explore next"
  5. Reference material for independent: docs, related files, further reading
  6. Close loop: "To summarize..." — 1 sentence for core concept

Got: Learner applied concept ≥1×, has resources for continued learning. Summary anchors for future recall.

If err: Too hard → teaching jumped too far, simplify. Can do but can't explain why → procedural w/o conceptual, return Step 3 focused on "why" not "how".

Check

  • Level assessed before explanation
  • Scaffolded known → unknown, not data dump
  • ≥1 check Q asked to verify (not assumed)
  • Teaching adapted on feedback, not repeated identical
  • Learner can apply, not just recall
  • Honest gaps acknowledged, not glossed

Traps

  • Curse of knowledge: Forget learner doesn't share teacher context. Jargon, assumed prereqs, implicit reasoning = primary culprits.
  • Explain to impress vs teach: Comprehensive, precise explanations demonstrating knowledge but leaving learner behind.
  • Repeat louder: Doesn't land → repeat w/ more emphasis vs diff approach.
  • Test vs teach: Check Q's as gotchas vs diagnostic. Goal = reveal understanding, not catch failure.
  • Silence ≠ understanding: Absence of Q's ≠ explanation worked. Often means learner doesn't know what to ask.
  • One-size-fits-all depth: Novice gets advanced explanation "should see full picture" → overwhelms; expert gets beginner "better safe" → wastes time.

  • teach-guidance — human-guidance variant coaching person to be better teacher
  • learn — systematic knowledge acquisition building understanding to teach from
  • listen — deep receptive attention reveals actual needs beyond stated Q
  • meditate — clear assumptions between teaching episodes, fresh approach each learner

GitHub リポジトリ

pjt222/agent-almanac
パス: i18n/caveman-ultra/skills/teach
0
agentsagentskillsai-assisted-developmentclaude-codeskillsteams

関連スキル

llamaguard

その他

LlamaGuardは、暴力やヘイトスピーチなど6つの安全性カテゴリーにおいて、LLMの入力と出力をモデレートするMetaの70-80億パラメータモデルです。94〜95%の精度を提供し、vLLM、Hugging Face、Amazon SageMakerを使用してデプロイ可能です。このスキルを使用して、AIアプリケーションにコンテンツフィルタリングと安全策を簡単に統合できます。

スキルを見る

cost-optimization

その他

このClaudeスキルは、リソースの適正サイジング、タグ付け戦略、支出分析を通じて、開発者がクラウドコストを最適化することを支援します。AWS、Azure、GCPにわたるクラウド支出の削減とコストガバナンスの実施のためのフレームワークを提供します。インフラコストの分析、リソースの適正サイジング、または予算制約への対応が必要な際にご利用ください。

スキルを見る

quantizing-models-bitsandbytes

その他

このスキルは、bitsandbytesを使用してLLMを8ビットまたは4ビット精度に量子化し、精度の低下を最小限に抑えつつ50〜75%のメモリ削減を実現します。限られたGPUメモリでより大規模なモデルを実行したり、推論を高速化するのに理想的で、INT8、NF4、FP4などのフォーマットをサポートしています。HuggingFace Transformersと統合され、QLoRAトレーニングや8ビットオプティマイザーを可能にします。

スキルを見る

dispatching-parallel-agents

その他

このClaudeスキルは、複数のエージェントを配備し、3つ以上の独立した問題を並行して調査・修正します。共有状態や依存関係がなく解決可能な、無関係な障害が発生するシナリオ向けに設計されています。中核となる機能は並列問題解決であり、効率を最大化するために独立した問題領域ごとに1つのエージェントを割り当てます。

スキルを見る