Cursor is VS Code with three AI entry points: Tab for completions, Cmd/Ctrl K for an edit on the
lines you selected, and the Agent sidepane on Cmd/Ctrl I. Ten minutes is enough to land one real
change with all three.
1. Open your project folder
File → Open Folder and pick the repository root, not a single file: Agent searches and edits inside the folder you opened.
2. Learn the four keys
| Action | macOS | Windows and Linux |
|---|---|---|
| Accept a Tab suggestion | Tab | Tab |
| Accept only the next word | Cmd → | Ctrl → |
| Inline edit on the selection | Cmd K | Ctrl K |
| Toggle the Agent sidepane | Cmd I or Cmd L | Ctrl I or Ctrl L |
| Add the selection to chat | Cmd Shift L | Ctrl Shift L |
Cursor’s shortcuts reference lists the
macOS bindings; Windows and Linux swap Cmd for Ctrl.
3. Make one inline edit
Select the lines, press Cmd/Ctrl K, describe the change:
add a guard clause that returns early when userId is missing
What it does: scopes the request to the selected lines rather than the whole file, so the result is small enough to read.
4. Give Agent a job that spans files
@auth.ts add rate limiting to the login route, with a test beside the existing ones
What it does: @-mentions pull the named files into context. Agent can search
the codebase, read and edit files, and run terminal commands — real
tool use — and it applies edits automatically, so review happens after the
write, not before it.
5. Verify it worked
git diff --stat
What it does: lists every file Agent touched and by how much — the fastest check that it stayed inside what you asked for. Run the tests next.
Cursor also takes checkpoints: snapshots made automatically before significant changes in a session. Click one in the chat timeline to preview it, or restore it to put every file back to that state — local, and separate from Git.
Failure mode: an edit you did not want, already on disk
Because Agent writes as it goes, a vague instruction lands before you see it. Restore the checkpoint from just before the run, then ask again scoped to one file and one function — a session safety net, not a substitute for committing first.
Standing rules — framework, test style, paths to leave alone — belong in a rules file rather than in every prompt, and the outside tools Agent reaches are worth pinning down too: see MCP servers in Cursor.