I used to keep the AI in a separate window. Ask a question, copy the answer, paste it back, fix what it got wrong. A lot of tab-switching to save a little typing.
Then I put Claude Code inside VS Code, and the job changed shape.
Now the agent lives in the same place as my files. I don’t describe my code to it. It reads the code. I say “the morning script is double-counting tasks, find it and fix it,” and it opens the right files, traces the logic, makes the change, and tells me what it touched. I review the diff like I’m reviewing a teammate’s work, not writing my own.
That’s the real shift. I stopped being the person who types and became the person who decides.
A few things that make it work for me:
- Keep the agent on a short leash for anything risky. It can read freely. Before it writes to anything I’d hate to lose, I want to see the plan first.
- Write instructions, not prompts. A clear file that says “here’s how this project works, here’s what good looks like” beats a clever one-off message every time. The agent reads it and stays on the rails.
- Let it run the boring loops. Renaming things across forty files, wiring up a new config, checking why a build broke. This is exactly the work I’m bad at staying patient for.
- Read every diff. “No errors” is not “correct.” I’ve watched it write something confident and wrong. The review step is not optional.
The part nobody warns you about: it’s a little disorienting at first. You spend years building the skill of typing the code, and then the value moves to knowing what to ask for and catching when the answer is off. That’s a different muscle. Worth building.
I run production for a touring show, alone, across a lot of cities. My whole life is held together by small systems. This is the first time the thing maintaining those systems can also edit them. That’s not a small change.