Boris, the creator of Claude Code at Anthropic, gave a live talk on how to actually use the tool — not as an autocomplete helper, but as a real agent tool. This summary distills every key point: installation, workflow, the most important slash commands, and four principles that make the difference.
Speaker: Boris · Anthropic · Format: Live talk · Hands-on · Core formula: Context + Plan + Loop
Installation
The only prerequisite is Node.js. After that, a single command does it:
npm install -g @anthropic-ai/claude-codeWhy Claude Code is different
- Editor-agnostic. Works with VS Code, Xcode, JetBrains, Terminal, SSH, Tmux — no workflow switch required.
- Local & private. Your code stays on the machine — no upload, no training on your own code.
- Agentic. Takes on whole tasks instead of completing single lines.
First steps: the most important slash commands
| Command | What it does |
|---|---|
/help | All commands available in the current session. |
/terminal-setup | Configure terminal shortcuts (Shift+Enter etc.) — for VS Code, Alacritty, Warp. |
/config | Theme, model, output style. Alias: /settings |
/theme | Switch color theme — Light, Dark, ANSI, colorblind-friendly. |
/init | Creates a CLAUDE.md in the project. |
/install-github-app | Set up Claude GitHub Actions — automatic PR reviews. |
/permissions | Manage Allow/Ask/Deny rules for tools. Alias: /allowed-tools |
/clear | Wipe the conversation and context — worth doing for every new task. |
Mac tip: Enable dictation via System Settings → Accessibility → Dictation. For long prompts, a genuine speed-up.
The four core tips
1. Ask first, code later
Don’t have it write code straight away — ask questions about the codebase first. Claude explains what’s already there, looks through the Git history, analyzes PRs.
- “How is this controller used?”
- “Why does this function have 15 arguments?”
- “What changed in this PR?”
The effect internally at Anthropic: onboarding time cut from weeks to days.
2. Plan before code
Have Claude brainstorm and lay out a plan before it starts writing. That keeps the model from running off in the wrong direction.
“Think first, make a plan, and get my approval before you write any code.”
3. Build in feedback loops
Give Claude tools to check its own work. Then it iterates on its own toward markedly better results:
- Run unit tests
- Screenshots via Puppeteer
- iOS Simulator
- Your own lint or build scripts
4. CLAUDE.md — set up a context file
Use /init to create a CLAUDE.md in the project root. It’s read automatically at the start of every session and is the single strongest lever for consistent results.
- The project’s key Bash commands
- MCP tools and how to use them
- Style guide and coding conventions
- Architecture decisions
- Important files and entry points
It can be shared project-wide or enterprise-wide.
Pro tips & shortcuts
| Shortcut / Feature | Function |
|---|---|
Shift + Tab | Permission mode: default → acceptEdits → plan |
Ctrl + C | Cancel input or generation |
Esc + Esc | Rewind to an earlier checkpoint |
Ctrl + O | Transcript viewer — output, tool calls, MCP calls |
Ctrl + R | Reverse-search through command history |
! Shell mode | Run directly in the shell, output lands in context |
# Memory | Quickly write notes / context into CLAUDE.md |
| Auto-accept mode | Shift+Tab to acceptEdits — edits without confirmation |
| Parallel instances | Multiple Claudes via tmux, SSH, or Git worktrees |
As a Unix tool: the Claude Code SDK
Not just interactive — Claude Code can be dropped in as a building block in CI pipelines, scripts, and incident-response workflows:
claude -p "Analysiere diese Logs und fasse kritische Fehler zusammen"Use cases: log analysis, automating incident response, pre-commit hooks, issue triage.
Security, multimodal & adoption
- Security around Bash: A sophisticated permission system — commands aren’t run blindly.
- Multimodal: Drag and drop images and screenshots straight into the terminal.
- Adoption: Roughly 80% of Anthropic’s technical staff use Claude Code daily.
Takeaway: three levers. No fourth needed.
- Context. A good
CLAUDE.mdbeats any clever prompting. - Planning. Let it think and plan first — then code.
- Loops. Build in tools for self-verification — so Claude can iterate.
A summary of a live talk by Boris (Anthropic) on Claude Code. Commands and shortcuts verified against the official documentation.
Leave a Reply