I switched from Cursor to Claude Code six weeks ago. The first two weeks felt like a regression — the inline autocomplete was a habit I didn’t realize I had. The next four weeks felt like a step change: a different model of what an editor is for.

What changed

Cursor treats the editor as the surface. The model is in the IDE, the chat is a sidebar, and the diff is the unit of work. The agent proposes; you accept or reject. This is a fast loop for surgical edits and small refactors, but it caps out when the work is “I don’t know exactly what I want, but I know what the next step should be.”

Claude Code treats the terminal as the surface. The model is the editor. You describe an outcome; the model proposes a plan; you approve or redirect; the model writes and runs. This is a slow loop for line-edits, but a much faster loop for the “I have a vague shape in my head, help me make it concrete” work.

The two models are optimizing for different things. Cursor optimizes for the moment of editing. Claude Code optimizes for the moment of building.

What I miss

  • The inline autocomplete. There is no replacement for “I type a function name, the body appears.” That is a separate product (Copilot, Continue) and using it alongside Claude Code works fine, but it is friction.
  • The “apply to file” diff. Claude Code writes files whole, which is fine, but a unified diff you can scan in 200ms is something I miss.
  • The visual presence of the model in the IDE. There is something about the chat panel that makes it easier to remember that the agent is there. In a terminal, you have to remember to invoke it.

What I gained

  • The plan-then-execute loop. Most of my work now starts with a 5-paragraph plan I would not have written by hand. The plan is the thinking; the code is the output.
  • The model can run my tests, my build, my linter. I do not have to copy errors from a terminal into a chat.
  • The model can be wrong in private. If it generates bad code, I tell it “no” and it revises. The diff is in my editor; the bad version is in the conversation scroll. I can scroll back and see what it tried and why.
  • The cost of starting a project is dramatically lower. “Build me an Astro site with these properties” is one prompt away from a running site.

Where this goes

I do not think “Cursor vs Claude Code” is the right question for the next year. I think the right question is “what is the smallest set of tools I can use to keep my own thinking honest while the agent does the building?” The agent-first workflow is the better default. Cursor-style IDEs are the better escape valve for when I need to think with my hands.