完全版v2 Claude Code 実践ガイド Anthropic 開発責任者 Boris氏の28分講演を、
具体的プロンプト・コマンド・社内事例まで全部日本語で整理しました

Hello. Hey, everyone. I'm Boris. I'm a member of technical staff here at Anthropic, and I created Claude Code, and here to talk to you a little bit about some practical tips and tricks for using Claude Code. It's going to be very practical. I'm not going to go too much into the history or the theory or anything like this. And, yeah, before we start, actually, can we get a quick show of hands who has used Claude Code before? Yeah. All right. That's what we like to see. For everyone that didn't raise your hands, I know you're not supposed to do this while people are talking, but if you can open your laptop and type this, and this will help you install Claude Code just so you can follow along for the rest of the talk. All you need is Node.js if you have it. This should work. — What is Claude Code? — Claude Code is a new kind of AI assistant. Most prior generations have been about completing a line or a few lines of code at a time. Claude Code is fully agentic — meant for building features, writing entire functions, entire files, fixing entire bugs. It works with all your tools: VS Code, Xcode, JetBrains, Vim, Emacs — every IDE, every terminal, locally or over SSH/Tmux. It's general purpose, which can make "what do I type?" hard at first. — Setup — Run /terminal-setup (Shift-Enter for newlines), /theme (light/dark/daltonize), /install-github-app (today's announcement: @-mention Claude on any GitHub issue or PR). Customize allowed tools so you're not prompted every time. On macOS, enable dictation in Accessibility — hit the dictation key twice and just speak prompts. — Tip #1: Start with Codebase Q&A — The thing I recommend above everything else: ask questions about your codebase. We teach this on day one of technical onboarding at Anthropic. Onboarding used to take 2–3 weeks; now 2–3 days. There's no indexing — no remote DB with your code, code stays local, we don't train generative models on it. No setup, no waiting. Useful Q&A prompts: • How is this piece of code used? / How do I instantiate this? — goes beyond text search, finds real instantiation examples. • Why does this function have 15 arguments? — walks git history, finds who added them, links commits to issues, summarizes. • Look up GitHub issues with web fetch. • Every Monday: "what did I ship this week?" — reads git log by my username and gives a clean readout. We're not system-prompting Claude to use git history. The model just knows. We're lucky to build on such a good model. — Tip #2: Editing code — Small tool set: edit file, run bash, search files. Claude strings them together. Don't drop a 3,000-line ask cold. Best move: "before you write code, brainstorm, make a plan, run it by me, ask for approval." You don't need plan mode — just ask. Magic incantation: "commit, push, PR" — Claude reads git history, learns your commit format, branches, commits, pushes, opens the PR. No prompt engineering required. — Tip #3: Plug in your team's tools — Two kinds: bash CLIs (tell Claude the command, point at --help, drop frequent ones in CLAUDE.md) and MCP servers. Tell Claude about the tool and it starts using it. When entering a new codebase, give Claude everything the team already uses. Common workflows: 1. Explore → plan → confirm → code. 2. Give Claude a way to check its own work (unit tests, Puppeteer screenshots, iOS simulator). Then it iterates 2–3 times by itself and gets near-perfect output. Whatever your domain — give it a feedback loop. — Tip #4: Give Claude more context — CLAUDE.md is the special file. Project root CLAUDE.md → check in, share with team, auto-loaded each session. Local CLAUDE.md → personal, not checked in. Nested CLAUDE.md → pulled when working in that directory. Enterprise CLAUDE.md → org-wide. Contents: common bash commands, common MCP tools, architectural decisions, important files, style guide. Keep it short — long files eat context. Other context vectors: • .claude/commands/ slash commands (home dir or project). We use one in the Claude Code repo to label GitHub issues via a GitHub Action so humans don't have to. • @-mention files to pull them into context. Hierarchical config (project / global / enterprise) applies broadly: slash commands, permissions (allow/block lists), MCP server config. Example: block a URL company-wide via enterprise policy — users can't override. Auto-approve a common test command for all employees. Check an MCP JSON into the repo and every engineer gets prompted to install on first run. If unsure where to start: shared project context. Write once, share with the team, get network effects. /memory shows all loaded memory files; you can edit them. Pound-sign (#) to ask Claude to remember something — pick which memory file to write to. Real example: Anthropic's apps repo ships a Puppeteer MCP server config — every engineer can drive e2e tests and screenshots without installing themselves. — Pro tips: key bindings — Terminal is minimal so discoverability is hard. Quick reference: • Shift-Tab → auto-accept edits mode (bash still asks; edits don't). Use when Claude is on track. • # → remember this (writes to CLAUDE.md). • ! → bash mode. Runs locally; result goes into Claude's context. • @ → mention files/folders. • Esc → safely stop whatever Claude is doing. No corruption. • Esc Esc → jump back in history. • Ctrl-R → expand output (what Claude sees in context). • claude --resume / --continue → reopen prior session. — Claude Code SDK — Right after this talk, Sid does a deep dive on the SDK. The -p flag IS the SDK. Pass a prompt, allowed tools (specific bash commands too), and an output format (json or streaming-json). We use it in CI, incident response, all kinds of pipelines. Think "super-intelligent Unix utility" — pipe in, pipe out. Pipe in `git status` and use jq. Read a giant GCP-bucket log and ask Claude what's interesting. Pipe in Sentry CLI output. We've barely scratched the surface. — Advanced: parallel use — I'm a Claude normie — usually one running, maybe a few terminal tabs. Power users have SSH/Tmux tunnels into Claude sessions, multiple checkouts of the same repo with multiple Claudes, or git worktrees for isolation. We're actively making this easier. No limit on sessions. — Q&A — Q: Hardest part to implement? A: Safe Bash. Bash is inherently dangerous; manual approval on every command is unusable. We landed on: classifying read-only commands, static analysis to decide which combinations are safe, and a tiered permission system (project/global/enterprise allow- and block-lists). Hard because not everyone runs in Docker. Q: Multimodal? A: Fully multimodal from day one. Terminal makes it non-obvious. Drag-and-drop an image, give a file path, or copy-paste. I use it often: drag-drop a mock, give Claude a Puppeteer server, let it implement-and-iterate. Q: Why CLI not IDE? A: Two reasons. (1) At Anthropic engineers use VS Code, Zed, Xcode, Vim, Emacs — hard to build something for everyone, terminal is the common denominator. (2) We see up close how fast the model is improving — there's a good chance people aren't using IDEs by year-end. Want to be ready for that future and avoid over-investing in UI that may become irrelevant. Q: Use for ML / modeling / AutoML? A: A lot. Engineers and researchers at Anthropic both use Claude Code every day — roughly 80% of technical staff daily. That includes researchers using the notebook tool to edit and run notebooks. Thanks.