How Do You Keep Cursor's Agent on Track With a Plan and Checkpoints?
Press Shift+Tab in the chat input until the mode shows Plan, describe the task, answer its questions, edit the plan, and click Build. If the result misses, revert to a checkpoint, make the plan more specific, and run it again. Cursor's docs say this is often faster than fixing it with follow-up prompts.
[In Plan mode] Task: [the goal, in one or two sentences] Scope: [files, folders or systems in scope, and anything off-limits] Follow these existing patterns: @[reference file] Done when: [tests pass / this command succeeds / this behavior works] Ask me anything you need to know. In the plan, list the files you'll change and make each step something you can verify by running a command. Flag any step where you're choosing between valid approaches.
- 1Clarify and research
- 2Plan
- 3You edit it
- 4Build
- 5Revert, refine, rerun
Try it now
- 1Open Agent with Cmd+I (Ctrl+I), press Shift+Tab until you reach Plan, and send the prompt above for a real multi-file task.
- 2Answer its questions, then read the plan and delete or change one step.
- 3Click Build. If the result is wrong, hover over an earlier message and click Restore Checkpoint, edit the plan, and build again. It worked if the second run reflects your edit.
Use it when
Cursor names complex features with several valid approaches, tasks touching many files or systems, unclear requirements, and architectural decisions. It also suggests Plan mode automatically when your prompt includes keywords that indicate a complex task.
Skip it when
Quick changes or tasks you've done many times. Cursor's docs say jumping straight to Agent is fine there. Skip it too when the edit fits in one sentence and you already know the file, since the questions and research would take longer than the work.
Why it works
Plan mode researches your code and asks clarifying questions before it writes anything, so a wrong assumption shows up in a document you can edit rather than across 20 files. Cursor's docs add that the hard part is often deciding what change to make. A plan also breaks work into steps Agent can check, and it opens as a file, so you can change it directly. Checkpoints, which Agent creates automatically before significant changes, make it cheap to roll back and try again.
Example
Say you want notification preferences in your settings page. Plan mode asks which notification types exist and where to store them. Cursor's own example plan then follows src/pages/Settings.tsx, adds a column to UserPreferences, reuses the Toggle component, adds an API route and lists tests. You cut a step you don't need before clicking Build. Since each task is separate, you can see where it went wrong if something fails.
Going further
Plans are saved to your home directory by default, so they vanish from the team's view. Click Save to workspace to keep one with the code for sharing and documentation. From a plan, Build in Parallel runs independent steps at once while keeping dependent ones in order. For larger work, a Project assigns a coordinator agent to plan and delegate. And note that a checkpoint is not a commit: it's local and separate from Git, so commit each stage you want to keep.
Common mistakes
- Fixing a wrong build with a stream of follow-ups. Cursor's advice is to revert, refine the plan and rerun.
- Clicking Build without reading the file list, which is where scope creep is easiest to catch.
- Giving no done condition, so "finished" is whatever the agent decides.
- Leaving a useful plan in your home directory, where teammates can't read it.
Sources
- Plan Mode, Cursor Docs
- Plan mode, Cursor Help
- Developing features, Cursor Learn
- Cursor Agent, Cursor Docs
- What is multi-agent coding?, Cursor Help
- Projects, Cursor Docs
Checked 6 November 2026. Products change often, so confirm details with the vendor.