claude-code-playbook (toy-crane) — Summary

claude-code-playbook by toy-crane is a Korean-language Claude Code lecture site built with Next.js/Fumadocs (MDX) that doubles as a Claude Code plugin providing QA and migration skills for students. The project is primarily an educational content site (content/docs/ MDX) rather than a framework for general use; the plugin is a lightweight learning aid shipped alongside the docs site. It ships three specialized .claude/agents/: fact-check-reviewer (Opus model), style-reviewer (Sonnet), and korean-reviewer (Sonnet), each reading CLAUDE.md rules and returning table-formatted reports without making any edits. The CLAUDE.md enforces TDD for logic changes and portless proxy-based branch URLs for live preview; three worktree-lifecycle hooks (WorktreeCreate, WorktreeRemove, PreToolUse:Edit|Write) guard branch isolation. Two skills are published via a .claude-plugin/ manifest: /qna (lecture Q&A) and /migrate-to-copilot (bulk migration of Claude Code primitives to Copilot CLI format).

Differs from seeds: Closest to agent-os (Archetype 4 — markdown-scaffold + small command surface), but toy-crane adds a live web deployment, a publishing plugin manifest, and subagents with role-differentiated model selection (Opus for fact-checking, Sonnet for style/language). Unlike superpowers, which is a general-purpose skills framework, toy-crane's plugin exists purely to support students of a specific Korean lecture series and is tightly coupled to the site's MDX content structure.

claude-code-playbook (toy-crane) — Overview

Origin

A Korean-language Claude Code lecture course by toy-crane (GitHub: alwaysfun2183@gmail.com), hosted at https://docs.claude-hunt.com. The site serves students who are already comfortable with Cursor/Copilot and want to master Claude Code. Students share projects at https://claude-hunt.com.

Target audience

Developers with 2+ years of experience who have already used Cursor or GitHub Copilot; content is in Korean (한국어).

Philosophy

The project treats the repo as both a lecture-delivery platform (Fumadocs-based MDX site) and a Claude Code plugin. The plugin ships only the learning aids necessary to support the lecture content; no general-purpose agent framework is included. The CLAUDE.md makes this explicit:

"콘텐츠 작성 시 .claude/rules/ 의 규칙을 따릅니다." "새 레슨 작성·기존 레슨 수정 후에는 style-reviewer · korean-reviewer 에이전트로 일관성을 점검합니다." "로직 변경은 TDD로 진행 합니다 — 실패 테스트 → 최소 구현 → 리팩터 순서"

Distinguishing design choices

1. Reviewer-only agents: All three agents explicitly refuse to make edits (직접 수정하지 않는다); they return table reports only. This reduces the risk of agents silently corrupting MDX lecture content.
2. Model differentiation by role: fact-check-reviewer uses Opus for higher accuracy on API fact claims; style-reviewer and korean-reviewer use Sonnet.
3. Worktree isolation hooks: Hooks enforce that edits only happen on the correct worktree branch, preventing cross-branch contamination when multiple lectures are in flight simultaneously.
4. Plugin manifest for portability: The .claude-plugin/plugin.json + skill files let students install the Q&A assistant into their own projects without cloning the full lecture site.

claude-code-playbook (toy-crane) — Architecture

Distribution

• Type: Claude Code plugin (.claude-plugin/) + Next.js/Fumadocs site
• Install: Via /plugin UI in Claude Code session → Marketplaces tab, or by adding to .claude/settings.json:

  { "enabledPlugins": ["playbook-essentials@claude-code-playbook"] }
  

• Site dev: bun install && bun run dev (portless proxy → https://playbook.localhost)

Runtime requirements

• Node.js / Bun (for the site)
• Claude Code (for the plugin)
• portless proxy for branch-based preview URLs

Directory tree

.
├── .claude/
│   ├── agents/
│   │   ├── fact-check-reviewer.md   (Opus, WebFetch/WebSearch tools)
│   │   ├── style-reviewer.md        (Sonnet, Read/Grep/Glob)
│   │   ├── korean-reviewer.md       (Sonnet, Read/Grep/Glob)
│   │   └── dev.md                   (dev server skill)
│   ├── commands/
│   │   ├── worktree-create.sh
│   │   ├── worktree-guard.sh
│   │   └── worktree-remove.sh
│   ├── hooks/
│   ├── rules/
│   │   ├── course-content.md
│   │   ├── writing-principles.md
│   │   ├── writing-checklist.md
│   │   ├── writing-korean.md
│   │   └── lesson-frame.md
│   └── settings.json
├── .claude-plugin/
│   ├── plugin.json
│   ├── marketplace.json
│   └── skills/
│       ├── qna/
│       └── migrate-to-copilot/
├── content/docs/               (MDX lecture content)
├── skills/                     (legacy location for qna, migrate-to-copilot)
├── src/                        (Next.js app)
└── supabase/                   (DB for student project sharing)

Target AI tools

• Claude Code (primary, plugin-first)
• Portless for branch-based preview isolation

claude-code-playbook (toy-crane) — Components

Agents (3 in .claude/agents/, 1 additional)

Skills / Plugin skills (2, in .claude-plugin/skills/)

Hooks (3, in .claude/settings.json)

Rules (5 markdown files in .claude/rules/)

Scripts (3 bash scripts in .claude/hooks/)

claude-code-playbook (toy-crane) — Uniqueness

Differs from seeds

Closest seed is agent-os (Archetype 4 — markdown scaffold, minimal primitives), but toy-crane diverges in three important ways: (1) it ships a live web product (Fumadocs site + Supabase) alongside the CLAUDE.md conventions, making it a real product with agentic authoring assistance rather than a pure template; (2) it uses explicit model-role differentiation (Opus for fact-checking, Sonnet for style/language), which only taskmaster-ai among the seeds does for different task types; (3) its subagents are deliberately read-only reporters, which is a stronger safety constraint than any seed framework enforces. It does not overlap with superpowers (no skills-only behavioral framework) or openspec (no spec/command mirror pattern).

Positioning

A niche project: a Korean Claude Code lecture site that also functions as a deployable Claude Code plugin for its students. Not designed for general-purpose developer adoption; the plugin's value is specifically tied to the lecture content at docs.claude-hunt.com.

Observable failure modes

1. Single-author dependency: The repo is clearly a solo project; plugin quality depends on one maintainer's lecture updates.
2. Korean-only content: Non-Korean speakers cannot use the lecture content; the plugin's /qna skill references Korean-language source material.
3. Worktree guard is Claude Code-specific: The WorktreeCreate/WorktreeRemove hook events are Claude Code-specific; the /migrate-to-copilot skill exists precisely because students need to escape this lock-in.
4. No structured test coverage for MDX content: The fact-checker is an LLM agent, not a deterministic validator; it can miss errors or produce false positives at any time.