First Day with Claude Code — Initial Impressions

After years of hopping between Cursor, Copilot, and various LSP setups, I decided to give Claude Code a serious shot today. Here’s what happened.

The Setup

Getting started was straightforward — install the CLI, authenticate, point it at a project. I started with a mid-sized React project I’ve been meaning to refactor.

First impression: the context window is real. It actually reads your entire codebase without choking. That sounds obvious, but after years of context limits that force you to manually feed files, it feels almost unfair.

What Worked Well

Architecture discussions. I spent about 30 minutes talking through a refactor plan with Claude Code. Not coding — just thinking out loud. It tracked the conversation, asked good clarifying questions, and by the end had a coherent plan I could actually use.

Debugging sessions. Pointed it at a gnarly race condition. It traced through the code, identified the likely culprit (a useEffect with a missing dependency), and explained why it was wrong. More useful than just pointing at the error message.

Documentation generation. Asked it to write docs for a module I’d been putting off. The output was decent — not great, but good enough that I only had to edit 20% of it.

Where It Still Struggles

Large refactors. When I asked it to restructure three interconnected files, it got confused about which version of the code was current. The agent would reference stale context despite clear file changes.

Terminal state. It doesn’t always know what’s in your terminal. I had a build error, pasted it, and got advice that assumed a different error had occurred. Some context bleed.

Speed. Running with Claude 3.5 Sonnet via the CLI is noticeably slower than local Copilot. Not unusable, but noticeable enough to break flow occasionally.

The Verdict After Day One

Useful. Genuinely useful for architecture work, debugging, and documentation. But it’s not at the point where I can hand it a feature spec and come back to working code.

My hypothesis: it’s best as a thinking partner, not an autonomous agent. Give it a constrained task with clear boundaries and it delivers. Give it open-ended work and it wanders.

More testing needed. Check back tomorrow.