“This was our ‘secret weapon’ internally. But it was so powerful that we decided to release it externally.”
— Boris, Anthropic Technical Staff (Claude Code Developer)
In February 2024, Anthropic released “Claude Code.” This is not just a coding assistance tool. It provides an entirely new development experience where you can call Claude, an AI, as a “resident team member” on your local machine.
This article explains the essence of Claude Code, how to use it, and its future possibilities based on talks by Anthropic developers.
What is Claude Code?
Claude Code is an “AI agent” that you can call Claude from the terminal.
It can be launched simply by running the claude command in your usual shell (macOS Terminal, iTerm2, SSH environments, tmux, IDE terminals, etc.).
npm install -g @anthropic-ai/cloud-codeclaudeClaude reads the codebase in that directory and autonomously performs specified tasks (refactoring, test creation, fixes, structural changes, etc.).
You may also want to read
What is Claude Code? A Complete Guide to Pricing, Features, and Security
Why the Terminal Was Chosen
On this point, developer Boris says:
“The development environments used by our internal engineers vary greatly. Some use VSCode, others are Vim devotees. That’s why we chose the terminal as the entry point, which is the most common denominator.”
As a result, Claude Code has gained the following advantages:
- Works in any development environment (not dependent on GUI or specific IDEs)
- Lightweight with rapid improvement cycles (development cycle is faster without UI)
- Seamlessly compatible with existing IDEs
Especially when used in an IDE terminal like VSCode, Claude’s edits are reflected directly on screen, creating a feeling of humans and AI collaborating on the same screen.
From Internal Tool to External Release
Claude Code was originally a non-public tool for internal use at Anthropic. However, when deployed across the company, internal DAU (Daily Active Users) skyrocketed in just 3 days.
“That was the moment we were convinced we should release it,” Boris recalls. Even after the external release, the response was strong, especially from users such as:
- Enterprise developers working with large codebases
- Individual developers particular about their IDEs
- DevOps engineers integrating with CI/CD environments and GitHub Actions
With a clear picture of the actual user base, it became evident that Claude Code is not just an internal tool but a widely sought-after development support infrastructure.
Evolution Realized with Claude 4
Initially, Claude Code ran on Claude 3.7 Sonnet, but has now migrated to the Claude 4 series (Opus/Sonnet). This transition dramatically improved the precision of understanding intent and generating code accurately in one shot.
For example:
- Unit test generation → Nearly 100% automatically and accurately completed
- Refactoring and fix instructions → Expected output possible in one attempt
- GitHub Actions integration → Can call and request fixes/reviews with
@claudeon PRs
As a result, tasks that used to be “a bit of a hassle, so let’s postpone” have now transformed into tasks that are completed instantly.
Development Style Leveraging Claude (Instructing Rather Than Writing)
Development using Claude Code is significantly different from the traditional style of “humans writing code.”
An agent-directed style, where you have Claude create plans and make corrections as needed, becomes mainstream.
Anthropic developers actually use it like this:
- First, ask Claude to “plan how to implement this feature”
- Have it present several options and select the approach to adopt
- Then, ask Claude to generate code and automatically edit files
- Have humans perform only final fine-tuning (if necessary)
In this way, Claude handles “design → thinking → coding,” allowing humans to focus on decision-making and reviews.
Sharing Work Rules with Claude MD
In Claude Code, you can describe instructions like “I always want you to do this” or “This directory is special” in a Markdown file called .claude.md.
For example:
## Always consider the following- Run the linter after editing- Output test files to `__tests__`- This project uses TypeScriptThese settings can be managed with three types of files:
- Project-wide settings:
claude.md - Personal user notes:
claude.local.md - Settings common to all projects: Placed in the
.claude/folder under the home directory
This allows Claude to start working with an understanding of “your context”.
Claude Becomes a Team Member with GitHub Integration
Claude Code enables you to “directly assign work to Claude in Issues and PRs” by integrating with GitHub.
Usage is very simple.
claude /install github actionJust run this command in the terminal, and Claude will be integrated into the repository.
Then, simply comment on PRs or Issues like this:
@claude fix this@claude write testsClaude will automatically modify files and create new branches or PRs.
Summary
Claude Code is not an “AI-assisted coding tool.” It’s more like adding a “super-competent colleague” who resides in your team. Introduction is very simple.
If you have a Node.js environment, you can start with these two lines:
npm install -g @anthropic-ai/claude-codeclaudeWhy not start by trying small fix tasks or unit tests and “pair programming” with Claude? In the future, we’re seeing a vision where Claude is invited not only to terminals and GitHub but also to chat apps and everyday communication environments. The act of developers “setting up an environment to write code” itself may gradually become a thing of the past.
[Reference]
This article was written based on the full transcript of an official talk (YouTube) by Boris, Anthropic Technical Staff.