the cursor-to-claude pipeline that doubled my throughput
use cursor for the chunky edits, claude code for the long-running tasks. one pattern, three weeks of trial, two specific rules that made it stick.
i’ve been bouncing between cursor and claude code for three weeks. last week i caught myself shipping twice as much as i did in early may, and i sat down to figure out why. the answer was a specific pipeline:
- cursor for tactical edits — anything where i need to see the diff land in my editor in under 5 seconds. inline edits, “fix this function,” small refactors.
- claude code for strategic runs — anything that touches more than one file or takes more than a minute. multi-file refactors, migrations, exploratory builds.
the trick was deciding up front which one a task belongs to. mid-task switching ate hours.
the two rules that made it stick
rule one: claude code gets the AGENTS.md. the project conventions, the testing strategy, the names i use. cursor doesn’t need it — cursor’s working memory is the open file. claude is making decisions across the whole repo, so it gets the whole repo’s contract.
rule two: cursor gets the keyboard, claude gets the terminal. when i’m in cursor i type. when i’m in claude code i talk in checklists. mixing the two interaction modes — typing into claude, talking to cursor — felt fine but produced worse output every time.
what i’d change
i’d like a third tier for “let it cook overnight” — a longer-running, more autonomous agent for the kind of work that’s too big for a focused claude session but too cohesive to break up. i haven’t found the right tool for that yet. open to suggestions.