Claude Code

kept

The first AI tool that respects a terminal-first workflow. Earned its place day one.

|claude-code-ecosystem|tool →

signal: Anthropic launch post

The Claim

Claude Code is Anthropic's CLI tool for working with Claude directly from the terminal. No browser tab, no electron app, no VS Code extension required. It reads your project structure, understands file relationships, and can make multi-file edits with explicit diffs you approve before they land.

What I Tried

I threw it at the messiest part of my workflow: refactoring a Next.js app that had grown organically over six months. Components that should have been split weren't, shared types were scattered, and there were three different patterns for data fetching.

I started with small asks. "Move this type to a shared file and update all imports." It nailed it. Not just the move, but finding every import across 14 files and updating them correctly. Then I escalated: "Refactor this page component into a controller-view pattern with a custom hook." It generated the hook, the view component, and the container, and the split actually made sense.

I also used it for writing Foundry tests for a Solidity contract. It understood the test conventions, used the right cheatcodes, and structured the test file in a way I didn't have to rearrange.

What Surprised Me

The project context is the real differentiator. It doesn't just see the file you're pointing at. It reads your directory structure, your configs, your existing patterns, and it matches them. When I asked it to create a new API route, it followed the naming conventions and file structure I'd already established, without me specifying them.

The diff-based approval flow is also exactly right. I see every change before it touches disk. No magic, no black box.

Who It's For

Terminal-native developers who want AI assistance without leaving their environment. If you already live in tmux and neovim, this slots in without any workflow disruption. It's also strong for anyone doing multi-file refactors where context across the project matters.

Verdict

Kept. This is the first AI coding tool I've used that doesn't ask me to change how I work. It runs where I already am, shows me what it wants to do before doing it, and understands enough about my project to be genuinely useful. It's not perfect at everything, but the foundation is right, and that matters more than any individual feature.