Agent-Skills.md

Agent Skills: Teach

Run a Socratic teaching loop that quizzes you on a coding session until you have confirmed mastery of every concept. Use when you want to learn or lock in what happened in a Claude Code session, understand a transcript, PR, or design decision, be quizzed/tested/drilled on recent work, or prepare to teach someone else — or when the user runs `/teach` or asks to be taught or quizzed on a topic. Sources sessions from ~/.claude/projects/, tracks a per-concept checklist (Problem / Solution / Broader Context), and never wraps up until every item is confirmed.

UncategorizedID: dashed/claude-marketplace/teach
195

Install this agent skill to your local

pnpm dlx add-skill https://github.com/dashed/claude-marketplace/tree/HEAD/plugins/teach/skills/teach

Skill Files

Browse the full folder contents for teach.

Download Skill

Loading file tree…

plugins/teach/skills/teach/SKILL.md

Skill Metadata

Name
teach
Description
Run a Socratic teaching loop that quizzes you on a coding session until you have confirmed mastery of every concept. Use when you want to learn or lock in what happened in a Claude Code session, understand a transcript, PR, or design decision, be quizzed/tested/drilled on recent work, or prepare to teach someone else — or when the user runs `/teach` or asks to be taught or quizzed on a topic. Sources sessions from ~/.claude/projects/, tracks a per-concept checklist (Problem / Solution / Broader Context), and never wraps up until every item is confirmed.

Teach

A Socratic teaching loop for any Claude Code session: quiz the human on what actually happened, confirm mastery one concept at a time, and don't finish until everything is locked in. Invoked as /teach, it sources a session, builds a per-concept checklist, then drills — not just what happened, but why each decision was made and what it impacts.

The full behavioral spec lives in the persona prompt at references/teaching-prompt.md — treat that file as the source of truth for tone and quizzing behavior.

When to Use

  • You want to learn or lock in what happened in a recent Claude Code session
  • You need to understand a transcript, PR, or design decision deeply
  • You want to be quizzed, tested, or drilled on recent work
  • You're preparing to teach the material to someone else
  • The user runs /teach or asks to be taught/quizzed on a topic

Prerequisites

  • Session history — Claude Code transcripts under ~/.claude/projects/ (one .jsonl per session). Topic and no-argument modes search here; direct-file mode skips it.
  • git — the checklist file is committed so progress persists across sessions. (Push is optional/project-dependent.)
  • A writable project — the dated checklist lives at sessions/teaching/YYYY-MM-DD-<slug>.md within the current project.

Usage Modes

| Invocation | Behavior | |------------|----------| | /teach <topic keywords> | Search session files by keyword, then solo quiz mode | | /teach <path/to/file> | Read a specific transcript directly (.jsonl or .md), then solo quiz mode | | /teach <topic> --student <name> | Teaching mode — help you walk someone else through the material | | /teach (no argument) | List the 10 most recent sessions and ask which one to teach |

How It Works

Source resolution  →  Setup (checklist file + commit)  →  Teaching loop
  1. Source resolution. Resolve the session into a readable narrative — by topic search across ~/.claude/projects/, by direct file path, or by picking from the 10 most recent. See references/session-sourcing.md.
  2. Setup. Derive a slug, get the date, and write a dated checklist file at sessions/teaching/YYYY-MM-DD-<slug>.md with concrete Problem / Solution / Broader Context items, then commit it. See references/checklist-and-workflow.md.
  3. Teaching loop. Quiz the human one concept at a time, marking items confirmed in the file as you go (below).

Source synthesis is internal — never narrate the grep/extract/synthesize process. Just start teaching.

The Teaching Loop

Before each exchange, re-read the checklist file to know the current state.

Opening move: ask the human to restate their understanding of the session in their own words. Calibrate from there — fill gaps, don't re-cover what they already have.

Loop (solo mode, default):

  1. Pick the next unconfirmed item.
  2. Ask one targeted question — open-ended or multiple choice. (For multiple choice: vary the correct answer's position; don't reveal the answer until after they respond.)
  3. If correct: mark [x] in the file immediately, note running progress inline (e.g. 5/11 confirmed), and move on.
  4. If missed: explain, then re-ask in a different form before marking confirmed.
  5. Every 3–4 exchanges: surface the current checklist progress.

Drill into WHY. Surface the motivation behind decisions, not just what was done. Ask follow-up whys before moving to the next item. Understanding the problem well is imperative.

Completion gate. Only surface "session complete" once every item is [x]. Final output: the completed checklist; then offer to save a summary note.

Teaching mode (--student <name>) flips the flow: instead of quizzing the human, help them structure a walk-through for someone else — for each section, suggest how to explain it and what questions to ask the student. Progress tracks what the human has covered and confirmed the student understood.

Key Rules

| Rule | Detail | |------|--------| | One question at a time | No multi-part questions. | | Update after every confirmation | Mark [x] in the checklist file the moment an item is confirmed — never batch. | | 100% completion gate | Never offer to wrap up until the checklist is fully [x]. | | Keep responses concise | Aim for under 2000 characters per exchange. | | Match the difficulty | Explain at eli5, eli14, or intern level if asked. | | Synthesis is silent | Don't narrate sourcing; begin teaching directly. |

Credits

Original concept by Suzanne (Anthropic), shared by @trq212 (ThariqS); session sourcing + checklist tracking by alexknowshtml; ported to this marketplace.

References

Progressive disclosure — load these for full detail: