Claude Code Setup, How Boris and the Anthropic Team Actually Use It
Claude Code is Anthropic's agentic coding CLI. It runs in your terminal, has direct access to your file system, and can read, write, execute, and reason about your entire codebase in a single session. Boris Cherny, who led much of the early Claude Code work at Anthropic, has shared a lot about how the internal team uses it, and the mental model is different from how most people approach AI coding tools.
The key shift: Claude Code is not an autocomplete tool. It's a collaborator that can hold context across an entire project. You don't use it to generate one function. You use it to execute multi-step tasks, refactor across files, and reason about architecture decisions while seeing the actual code.
How to set it up properly:
- Install via npm:
npm install -g @anthropic-ai/claude-code - Authenticate with your Anthropic API key
- Run from your project root, it ingests your directory structure immediately
- Use
CLAUDE.mdfiles to give it standing instructions about your project (conventions, what not to touch, how you like things structured)
The CLAUDE.md file is the unlock. Most people skip this. If you put a CLAUDE.md in your project root and tell Claude Code your stack, your conventions, and your current goals, every session starts with full context. No re-explaining.
Boris's observed workflow pattern:
- Open long-running sessions for complex tasks
- Give it the full goal, not just a step
- Let it run, interrupt only if it goes somewhere obviously wrong
- Review the diff, not every intermediate action
The mental model is closer to a junior engineer you trust than a sophisticated autocomplete. You wouldn't tell a junior engineer "write line 47 of this file." You'd tell them what you're trying to build and let them figure out the steps.
The biggest mistake I see: people use it like Copilot, one function at a time. That's leaving most of the value on the table.
I built this entire portfolio site in a single Claude Code session. The velocity is real.