Cursor agent best practices

Takeaways

Agent harness = instructions (rules/prompts) + tools (edit, search, terminal) + model; Cursor tunes per-model behavior from evals.
Plan before code—UChicago study: experienced devs plan more; Plan Mode researches, clarifies, drafts reviewable markdown plans.
Save plans to .cursor/plans/ for team docs and resumable work; edit plans directly before execution.
Context management matters: long threads degrade; start fresh sessions for new tasks; use @ references over pasted blobs.
Review agent output like a PR—rules, hooks, and linters extend the verification loop beyond the model.

Cursor-side counterpart to Claude Code + AGENTS.md guidance in this vault—Plan Mode and rules align with commit/skill conventions already in use.

How should vault .cursor/rules split always-on vs. path-scoped rules as the wiki grows?