Cursor Agent can search code, edit files and run terminal commands. Treat it like a fast collaborator with tool access—not an authority that should make unchecked changes.
1. Open the repository and establish the baseline
Open the real project, make sure it builds/tests before the change, and know how to revert with version control.
2. Ask Agent to inspect before editing
Open Agent and ask it to locate the relevant files, explain the current flow and propose a minimal change plan before touching code.
3. Give a bounded implementation request
Specify desired behavior, constraints, files/areas that must not change, compatibility requirements and acceptance tests.
4. Let it edit and run targeted checks
Allow necessary file edits and terminal commands, but keep destructive, deployment or secret-handling actions behind explicit human review.
5. Review the diff and tests yourself
Read the changed code, inspect generated tests, run the project and check for unintended edits before committing.
First inspect the codebase. Do not edit yet. Find the files responsible for [behavior], explain the current flow, list risks, and propose the smallest change that satisfies: [acceptance criteria]. After I approve the plan, implement it and run the relevant tests.
Starting from an uncommitted dirty working tree; asking for a vague “fix everything”; allowing broad refactors during a bug fix; approving terminal commands without reading them; merging because tests are green without reviewing the diff.
Finish check
The finished task should have a small understandable diff, relevant tests, and a clear explanation of what changed and what remains uncertain.