Claude Code first impressions
It's the first time (after trying Aider and Cursor with SOTA OpenAI and Anthropic models) when it was genuinely useful in a big repository (this website), and it implemented several features from the very high-level description.
I used the default model (Sonnet 4).
Started with the /init command to let it scan the repo and generate a CLAUDE.md file.
Then I asked it to implement some features in a really concise form.
For example, In this app, every note might have a tag, which is just a word that starts with a # and can be located anywhere in the note body. Implement a popup to that would suggest the most common tags to select. While user typing, it should filter out. It should allow using keyboard arrows to select a tag.
My repo is Rails backend plus TypeScript/React frontend. It was able to figure out what files need to be updated on its own, and did what I asked flawlessly in just one attempt.
Then the job left for me was only to review the code, and commit it.
I didn't even change the looks of it. It correctly used proper colours from the theme defines with CSS vars, and length units defined as SASS variables.
I guess the part of my success with it, is good source code to start with: it is properly organised - Rails has rigid file structure (convension over configuration). Typescript guards agains missing types. In CSS, I limit myself to using unit and color variables.
In total, I implemented 4 new features, and closed 2 small tech debt tasks. I ended up spending 50 minutes and around $4 on something that would take me the entire day. That would include installing Claude Code and figuring out how it works.
It's interesting, that Claude Code does not resort to indexing, and instead heavily relies on grep, find, and other CLI tools.
They also made it terminal-based, and didn't go the way of forking VS Code. As a Vim user, I really appreciate that.