This skill now lives in AgentGameLab/skills, together with our other agent skills, as one installable Claude Code plugin —
/plugin marketplace add AgentGameLab/skills. This repo stays up and still works standalone, but new versions land there. 本 skill 已迁入 AgentGameLab/skills 单仓，与其余 skill 合并为一个可安装的 Claude Code plugin。本仓保留可用，但后续更新只发新仓。
A Claude Code skill that upgrades "give the answer" into "teach until understood." 一个把「给答案」升级为「教到会」的 Claude Code 教学陪练 skill。
learn-coach turns the assistant from an answer-vending machine into a patient tutor. When you want to understand and retain something — an algorithm, a system, a piece of math, a domain concept — it walks you through it across turns instead of dumping a 1500-word lecture that you forget by tomorrow.
The single biggest failure mode of LLM teaching is saying too much at once — any pedagogy gets crushed under an 800-word monologue. Every rule in this skill fights that instinct.
- Feynman recall(费曼回讲) — you explain it back in your own words; gaps surface where the words get fuzzy.
- Socratic questioning(苏格拉底式提问) — one question at a time, with a hard-coded escalation ladder so it never traps you in endless "guess again."
- Scaffolding(脚手架分层) — start from an 80%-accurate model that actually runs, add precision layer by layer, and fade the scaffold as you gain independence.
- Productive failure(生产性失败 · 先猜后讲) — you commit a guess before the explanation; the wrong guess becomes the diagnostic that shapes what gets taught next.
Any "I want to learn / understand / master X" intent — even without the word "teach." Examples: 教我 X / 我想学 X / 帮我搞懂 Y / 给我讲讲 Z 的原理 / 一直没搞懂 / 带我过一遍.
It deliberately stays out of the way for lookup questions (an API name, a syntax detail), production firefighting, or when you say "just tell me the answer" — those get a direct answer, no Socratic detour.
- One step per turn, then stop and wait for you — cramming two steps into one turn defeats the whole pipeline.
- ≤300 words + 1 diagram + 1 question per turn.
- Every core concept gets a diagram (SVG/ASCII/Mermaid) — time-sequences, structure, and comparisons belong in pictures, not prose.
- Every term is glossed in plain language on first use.
- An escape hatch always wins: say "just tell me" and it explains directly, no insistence.
Content is Chinese-first (中文为主), with technical terms kept in English.
Clone into your Claude Code skills directory as a folder named learn-coach:
git clone https://github.com/AgentGameLab/learn-coach.git ~/.claude/skills/learn-coachClaude Code auto-discovers the skill from its SKILL.md front-matter — no extra registration needed.
learn-coach/
├── SKILL.md # the skill: routing + 7-step teaching pipeline + rules
├── references/
│ └── teaching-playbook.md # worked good/bad dialogue examples & phrasing library
└── evals/
└── evals.json # trigger/behavior eval cases
MIT — see LICENSE.