Claude Code: Three Months In
A quick reflection on how Anthropic's terminal agent has changed my day-to-day life as an engineer.
2025-09-10
Loading content...
A quick reflection on how Anthropic's terminal agent has changed my day-to-day life as an engineer.
Three months after my first impressions post, Claude Code is still open in a terminal tab somewhere on my machine. That was not guaranteed. Most AI tools get a two-week honeymoon and then become another subscription I forget to cancel.
What stuck is boring in the best way. I cd into a repo, start a Claude Code session, and hand off work I already know how to do but do not want to type again.
My calendar did not get better. I still have back-to-back meetings most days. The difference is what happens in the gaps.
Before a design review last week I kicked off a doc pass on an API module I had been avoiding. Wrong env var names, a route that moved two refactors ago, a test file that still imported the old helper. Claude Code fixed all of it while I was in the meeting. I came back to a diff and green tests, not a todo list.
Same pattern for boilerplate I hate: new endpoint, matching test, same error-handling shape as the file next door. I describe the done state. It runs the commands. I review the diff like any other PR.
That payment-service race condition from my first post? Still the example I use when someone asks if this is real. Close the loop, do not hand me homework. Three months in, that bar has held.
Architecture bets. Cross-repo refactors with fuzzy scope. Anything where "done" is a judgment call instead of a test suite.
It still guesses wrong when the task is vague. I learned that the expensive way in month one. Now I start with something bounded, same advice I gave in the terminal evolution write-up.
I have not picked one winner. Cursor Composer is still where I do multi-file UI work. Claude Code is where I want shell access, git state, and my actual project scripts without acting as the human API between chat and repo.
If you have been curious about terminal agents, pick one repo and one small task with a clear definition of done. A README fix. Tests on one gnarly module. You will know in an afternoon whether it earns a permanent tab.