Claude Code Training Curriculum — A One-Day Program Built Around Parallel Sessions and Plan Mode

1. One-day curriculum

Four to ten attendees per session. Plan Mode and parallel sessions are the two pillars (morning / afternoon). MCP and hooks in the afternoon are weighted according to the company's conventions.

2. CLAUDE.md — the home for team conventions

Claude Code auto-loads CLAUDE.md at the start of every session. Run it in two tiers: global (~/.claude/CLAUDE.md) and project-root. In training we hand attendees a template and they swap in their company's conventions.

# CLAUDE.md

## 最重要原則
1. Plan Mode で開始する。曖昧な指示はプランで合意してから実行
2. main / develop に直接コミットしない。必ず feature ブランチ + PR
3. 破壊的コマンド (rm -rf, git push -f, db migrate) は実行前に必ず確認

## コーディング規約
- pnpm を使う (npm/yarn 禁止)
- src/server/ は src/app/ から直接 import しない
- env 値を console に出力しない

## レビュー
- PR は実装エージェント名 (Claude Code / Cursor / Codex) を本文に明記
- テストなしの PR はマージしない

• Keep "top principles" to 3–5 lines. Too many and nobody follows them.
• For each prohibition, state the reason (e.g., "no direct push to main: PRs are required for review").
• List destructive commands by concrete name, not by category.

3. Plan Mode — plan, approve, execute

Plan Mode forces the agent to write the plan in prose before doing anything. We enforce these rules:

1. Any instruction longer than two lines enters Plan Mode via shift+tab
2. Always read the plan and push back ("change just this part")
3. Exit Plan Mode and execute only after agreement
4. For large changes, save the agreed plan to a separate file before executing

During the afternoon hands-on, the rule "Plan Mode at least five times each, implement only what has an agreed plan" cements the habit.

4. Slash commands and sub-agents

Encode frequent workflows as slash commands in .claude/commands/*.md. In training, attendees write three: PR review, library investigation, and commit-message generation.

# .claude/commands/review-pr.md
---
description: 現在のブランチの PR をレビューし、修正案を提示する
---

# レビュー手順
1. `gh pr view` で本文を取得
2. `git diff main...HEAD` で差分を確認
3. 以下の観点でレビューする:
   - テストカバレッジ
   - エラーハンドリング
   - 命名 / 抽象度
4. 結論: Approve / Request Changes / Comment を 1 行で示す

Heavy exploration and independent review go to a sub-agent via theTask tool. The parent gets a short summary back, which keeps the main context lean. We practice with "grep all markdown in this repo for legacy terms" to give attendees a feel.

5. Automation and safety nets via hooks

# .claude/settings.json (抜粋)
{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Bash",
        "hooks": [
          {
            "type": "command",
            "command": ".claude/hooks/block-destructive.sh"
          }
        ]
      }
    ],
    "PostToolUse": [
      {
        "matcher": "Edit|Write",
        "hooks": [
          { "type": "command", "command": "pnpm lint --fix" }
        ]
      }
    ]
  }
}

• PreToolUse blocks destructive Bash (rm -rf, git push -f)
• PostToolUse runs lint + typecheck after Edit / Write
• SessionStart loads prior notes; Stop persists the summary
• PreCompact archives a summary before long sessions get compacted

The training has attendees encode their company's forbidden commands both in CLAUDE.md and in a PreToolUse hook — defense in depth.

6. MCP integrations

MCP exposes external systems (Linear, Postgres, Slack, GitHub) as agent tools. The training adds Linear MCP and runs "read my ticket in Plan Mode → edit the linked files → open a PR" end-to-end. API keys are passed through 1password run, not pasted into.env.

7. Parallel sessions — 15 tasks a day

The peak of the day. We set up tmux + git worktrees so "one task = one branch = one terminal" is mechanically enforced.

1. Create a worktree: git worktree add ../<repo>-task1 -b feat/task1
2. Split tmux panes and launch claude in each worktree
3. One cycle: Plan Mode → agree → execute → Apply, finishing in around 30 minutes
4. When done, open the PR and remove the worktree immediately

That structure gives attendees the firsthand sense of running 10–15 tasks a day with zero wait. We also cover sharing .cursorignoreand hook settings at the worktree level for safety.

8. PR workflow alongside Codex and Cursor

Claude Code

Runs many tasks in parallel. Owns long exploration, refactors, and end-to-end PR creation.

Cursor

Single-file or scoped finishing. The "shape it up before review" pass.

Codex

Cross-reviews Claude Code's PRs. Independent reimplementation lifts decision quality.