Quick answer

Use Tab completions for routine typing, an inline edit shortcut for local changes, Agent mode for multi-file tasks, and project rules to encode conventions. Give the agent small, well-specified tasks; review diffs; run tests; and use checkpoints and version control to undo mistakes.

Core features

Cursor lets you pick among several models, including its own, and has been reported to offer models from OpenAI, Anthropic, Google and others. Availability and naming change often.

Prompting the agent

A well-scoped task

Add rate limiting to the /login endpoint: max 5 attempts per IP per 15 minutes, returning HTTP 429 with a Retry-After header. Reuse the existing middleware pattern in src/middleware. Add tests for the limit and reset. Don't modify other routes. Show me a plan first.

Guidelines

• Point at context with file and symbol references so the agent doesn't guess.
• Ask for a plan and correct it before edits begin.
• Keep tasks atomic: one feature or bug per run.
• State the definition of done: tests pass, linter clean, no unrelated changes.
• Feed back errors verbatim when something fails.

Project rules

Cursor supports persistent rules that tell the AI about your stack, patterns and preferences (for example testing framework, style conventions, folders to avoid). Put stable conventions there instead of repeating them in every prompt, and keep them concise and specific.

Workflows

Fix a bug you don't understand

1. Paste the error and ask chat to trace the likely path through the code.
2. Have it propose two hypotheses and a test that would distinguish them.
3. Apply the fix in agent mode, run the test, and inspect the diff.

Build a small feature

1. Describe the feature and ask for a plan and file list.
2. Approve, then let the agent implement in steps, running tests as it goes.
3. Review the diff, commit, and repeat for the next piece.

Refactor safely

1. Commit your working state first.
2. Ask for the refactor in small stages.
3. Run the full test suite after each stage; revert to a checkpoint or commit if it fails.

Review and safety habits

• Use Git branches and commit often. Checkpoints help, but they don't replace version control.
• Read diffs before accepting. Watch for deleted code, changed tests that hide failures, and new dependencies.
• Be careful with terminal commands the agent proposes, especially destructive ones, and with what secrets are in your environment.
• Test manually as well as automatically for user-facing behavior.
• Check licensing and security implications of any generated code you ship.

Cost and limits

Reviews report that Cursor moved from fixed request counts to usage-based credits on paid plans in mid-2025, and that heavy agent sessions, especially with premium models or expanded-context modes, can cost noticeably more than the base subscription. Monitor the usage dashboard, choose lighter models for simple tasks, and check current pricing before relying on any figure.

Common mistakes

• Handing the agent a vague, sprawling task.
• Accepting changes without reading the diff.
• Using the most expensive model for trivial edits.
• Skipping tests because the agent "ran them."
• Letting context files and rules grow stale and contradictory.

Limitations and alternatives

• It requires coding knowledge to supervise effectively.
• It does not provide hosting or deployment by itself.
• AI can hallucinate APIs or produce subtly wrong logic.
• If you prefer to stay in another IDE or in the terminal, other assistants and agents fit that workflow better.

FAQ

What is Cursor?

An AI-first code editor built on the VS Code codebase, with chat, multi-file editing, agent mode and completion.

What is Agent mode?

A mode where Cursor plans, edits files, runs commands and iterates on errors, creating checkpoints you can revert.

How does Cursor pricing work?

Paid plans have used usage-based credits, so costs vary. Check current pricing and your usage dashboard.

Can a beginner use Cursor?

Yes, but you still need to read, test and understand what it writes.