本文へ移動
cccskills
無料GitHub で公開

socratic-mode

Use when the learner benefits from guided questioning to build reasoning and uncover misconceptions before receiving direct answers.

インストール方法を見る

含まれるファイル(2)

  • SKILL.md2.5 KB
  • examples.md1.2 KB

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

Purpose

Teach through strategically sequenced questions that reveal the learner's reasoning process and surface misconceptions before providing answers.

Activation

  • Learner asks for understanding, not just output. Learner is stuck on reasoning errors. Goal is interview readiness, architecture thinking, or debugging judgment. Learner requests hints.
  • Skip if: user explicitly wants an immediate final answer, or safety-critical urgency demands direct correction first.
  • Routing: prefer after teach-concept when understanding remains shallow. Combine with check-understanding to evaluate responses.

Inputs

  • Target problem or concept, learner goal and level, known misconceptions or error patterns.

Workflow

  1. Frame — State that guidance will be question-led. Define the target outcome.
  2. Elicit — Ask learner to explain their current understanding or plan.
  3. Probe — Ask about edge cases, constraints, tradeoffs. Use counterexamples to expose weak reasoning.
  4. Guide — Offer hints from broad to specific. Escalate hint specificity only if learner is blocked.
  5. Synthesize — Ask learner to restate corrected reasoning in their own words.
  6. Close — Assign one implementation or debugging task to apply the correction.

Rules

  • DO: ask one question at a time when confusion is high.
  • DO: include synthesis/help every 2–3 probes — don't just interrogate.
  • DO: keep tone supportive while holding high reasoning standards.
  • DO: end with a corrected model restatement and a concrete application task.
  • DON'T: give full answers before learner attempts reasoning.
  • DON'T: ask vague questions — include context and expected scope.
  • DON'T: leave detected misconceptions unclosed — always end with explicit correction.

Output

Responses should contain: context (concept + reasoning goal), guided questions (sequenced), hints (if needed, broad→specific), synthesis check (learner restates), and next step (application task). Format naturally.

Checklist

  • Questions sequenced from model-elicitation to correction.
  • Learner reasoning attempt required before full answer.
  • Misconception explicitly surfaced and corrected.
  • Transfer action assigned.

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

Use when critiquing a system design across correctness, scalability, reliability, maintainability, and observability using probe-first questioning.

日本語の概要は準備中です。原文の説明を表示しています。

yugash007/edu-agent-skills92026年5月19日 更新

Use when guiding a learner through real project implementation while preserving their ownership and teaching through the build process.

日本語の概要は準備中です。原文の説明を表示しています。

yugash007/edu-agent-skills92026年5月19日 更新

Use when generating personalized practice challenges calibrated to the learner's weak areas, level, and project context.

日本語の概要は準備中です。原文の説明を表示しています。

yugash007/edu-agent-skills92026年5月19日 更新

Use when verifying whether the learner truly understands a concept through reasoning and application, not recall trivia.

日本語の概要は準備中です。原文の説明を表示しています。

yugash007/edu-agent-skills92026年5月19日 更新

Use when coaching hypothesis-driven debugging that requires the learner to gather evidence and reason about root causes before applying fixes.

日本語の概要は準備中です。原文の説明を表示しています。

yugash007/edu-agent-skills92026年5月19日 更新

deep-dive

無料

Use when a learner has surface understanding and needs to explore mechanism, tradeoffs, edge cases, and production implications at depth.

日本語の概要は準備中です。原文の説明を表示しています。

yugash007/edu-agent-skills92026年5月19日 更新

yugash007 のスキルをすべて見る

このスキルの問題を報告する