Cursor Training Curriculum — Half-Day Program Design for Engineering Teams

1. Half-day curriculum

Four to eight attendees per session works best. As an in-house program, the productive pattern is "non-coding AI use" in the morning and this Cursor half-day in the afternoon, so non-engineers and mixed teams interlock.

2. Tab and Inline Edit habits

Tab autocomplete is heavily context-coupled. It is strong on one-line completions and weak when it proposes a long diff — in that case it is faster to stop and use Inline Edit (⌘K) with a scoped instruction. We share these heuristics with attendees:

• Stop and use ⌘K whenever the completion is five lines or longer
• Accept import additions; read logic additions before accepting
• Trust completions more in test files; be strict in production code

3. Composer — workhorse for multi-file edits

Composer (⌘I) is the "instruct → diff preview → Apply" loop for multi-file edits. We run three hands-on patterns:

1. Scope the work with @Folder to a feature directory
2. Use @File as anchor plus @Code for symbol-level references
3. Read the diff before Apply; iterate by saying "rewrite hunk N as {...}"

Enforce git add -p after every Apply. That single rule removes incidents where AI-introduced changes slip into a commit unnoticed.

4. Agent Mode — drawing the boundary

Agent Mode loops through file discovery, terminal execution and self-correction. Powerful, but easy to wreck a repo if the boundaries are wrong. We always teach four rules:

• Auto-run commands are an allowlist. Never include rm, git push, or migrate.
• Always a fresh branch: run git checkout -b agent/xxx before launching Agent Mode.
• Cap any single agent task at 30 minutes; force an Apply / Reject decision before continuing.
• Never let it read production .env: add .env* and secrets/ to .cursorignore.

5. Cursor Rules — shared team style

The core of the training is writing a rules file and using it through Apply. We hand attendees this template and they swap in their company's conventions.

# .cursor/rules/team.mdc
---
description: チーム共通ルール (TypeScript / Next.js / pnpm)
globs: ["**/*.ts", "**/*.tsx"]
alwaysApply: true
---

# コーディング規約
- pnpm を使う。npm / yarn コマンドは提案しない
- 新規ファイルは Server Component を既定にする。"use client" は必要時のみ
- as any 禁止。型が複雑なら unknown + narrowing
- console.log は提案しない。logger を使う

# 禁止
- src/server/ の Server-only API を src/app/ から直接 import しない
- 環境変数の値をログ出力しない (process.env.* を直接 console に渡さない)

• Only one file should set alwaysApply: true; if all rules are always-on they fight each other
• For each prohibition state the reason; without "why," the model ignores it
• Split into 3-5 files rather than one giant rule; let globs auto-select

6. Adding MCP

// ~/.cursor/mcp.json (グローバル) または .cursor/mcp.json (リポ単位)
{
  "mcpServers": {
    "linear": {
      "command": "npx",
      "args": ["-y", "@linear/mcp"],
      "env": { "LINEAR_API_KEY": "lin_api_xxx" }
    },
    "postgres": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-postgres", "postgres://..."]
    }
  }
}

The hands-on adds Linear MCP and walks through "use @Linear to pull my ticket, then implement it in the linked files with Composer." API keys live in env only for the exercise; in production we pass them through 1password run or an OS keystore.

7. Review workflow and sensitive data

1. The review bar is the same whether the PR is AI-generated or not. But list the agent (Cursor / Claude Code / Codex) in the PR description.
2. No tests, no merge — for humans or AI.
3. Enable Privacy Mode. On enterprise plans, double-check per repository.
4. Protect privileged data (employee records, customer PII) with three layers: .cursorignore, Privacy Mode, and input moderation.
5. When someone leaves, remove them from the Cursor team immediately — cached copies can still reach prior code.

8. Coexisting with Codex and Claude Code

Cursor

Stay in the editor for single-file or scoped edits. The "make it review-ready" pass.

Claude Code

Terminal-based, multiple chats in parallel. Long-running exploration, larger refactors, and templated repetitive work via PR templates.

Codex

Second-opinion review and parallel reimplementation. Use it to audit what Cursor produced.