Cursor AI Chat

Cursor AI Chat

Overview

Use Chat to reason about a bounded question or planned change while preserving human ownership of decisions, edits, review, and validation.

Prerequisites

• Approved privacy/model settings and a workspace with sensitive paths excluded.
• A clear question, relevant non-sensitive files, and the project's test/review expectations.

Instructions

1. Start a topic-scoped chat and attach only the files needed for the question.
2. Ask for assumptions, tradeoffs, and a testable proposal rather than accepting generated code blindly.
3. Verify suggestions against source, tests, security controls, and a human code review.
4. Start a new conversation when context becomes stale or the topic changes.

Output

• A bounded explanation or proposal with assumptions, affected files, and validation steps.

Error Handling

Examples

Ask Chat to compare SSE and WebSockets for the stated service, attaching only the API contract and architecture note. Require it to list assumptions and a rollout/rollback test plan; then review the proposal before authoring any change.

Master the Cursor AI Chat panel and inline edit for code assistance, debugging, and exploration.

Core Chat Interfaces

Chat Panel (Cmd+L / Ctrl+L)

The side-panel chat for conversational code assistance:

┌─────────────────────────────────┐
│  Chat                     [M]   │  ← Model selector
│─────────────────────────────────│
│  Context Pills:                 │
│  [main.ts] [utils/] [@Web]      │  ← Active context shown as pills
│─────────────────────────────────│
│  User: Explain the auth flow    │
│  in @src/auth/middleware.ts     │
│                                 │
│  AI: The middleware checks...   │
│─────────────────────────────────│
│  [Type message...]    [Send]    │
└─────────────────────────────────┘

Key actions:

• Select code in editor, then Cmd+L to add selection as context
• Cmd+Shift+L adds selection to existing chat without clearing
• Click model name in top-right to switch models mid-conversation

Inline Edit (Cmd+K / Ctrl+K)

Surgical edits within the editor. Select code (or place cursor), press Cmd+K, type instruction:

Selected: function calculateTotal(items) { ... }

Prompt: "Add TypeScript types and handle empty array edge case"

Result: Cursor shows diff inline -- green for additions, red for removals.
        Press Cmd+Y to accept, Esc to reject.

Best for single-function edits, type annotations, refactoring a specific block, adding error handling.

@-Symbol Reference

Type @ in chat to access context sources:

@Docs: Adding Custom Documentation

Register external docs for @Docs lookup:

1. Cursor Settings > Features > Docs
2. Click Add new doc
3. Enter URL: https://react.dev/reference/react (Cursor crawls and indexes it)
4. Reference in chat: @Docs React then ask questions

Works with any documentation site. Good for framework APIs, internal wikis, design systems.

Effective Prompting Patterns

Pattern 1: Context-First Prompts

@src/api/orders.ts @src/types/order.ts

The createOrder function doesn't validate the shipping address.
Add Zod validation that checks: street, city, state (2-letter), zip (5 or 9 digit).
Throw a typed ValidationError if invalid.

Pattern 2: Reference-Based Generation

@src/api/users.ts

Create a new file src/api/products.ts following the exact same pattern
as the users API. CRUD endpoints, same error handling, same response format.
Product fields: id, name, price (cents), category, createdAt.

Pattern 3: Debug with Context

@src/hooks/useAuth.ts @Lint Errors

The useAuth hook causes an infinite re-render loop when the token expires.
The useEffect fires repeatedly. What's the root cause and how do I fix the
dependency array?

Pattern 4: Architecture Discussion

@src/components/ @src/api/

I need to add real-time notifications. The app uses REST APIs currently.
Should I add WebSockets, SSE, or polling? Consider the existing architecture
shown in these directories.

Multi-Turn Conversation Management

Start new chats for new topics. Long conversations degrade response quality as context fills up.

When to start fresh:

• Switching to a different feature or bug
• After 10+ turns on the same topic
• When responses start repeating or losing accuracy

When to continue:

• Iterating on the same code change
• Follow-up questions about the same file
• Asking for tests after generating code

Model Selection for Chat

Switch models using the dropdown in chat header:

Chat vs Inline Edit vs Composer

Enterprise Considerations

• Privacy Mode: Enable in Cursor Settings > General > Privacy Mode for zero data retention
• Model governance: Teams can restrict which models developers access via admin dashboard
• Audit: Chat interactions are not logged server-side when Privacy Mode is on
• Cost control: Use Auto mode or default to faster models; reserve Opus/GPT-5 for complex tasks

Resources

• Cursor Chat Documentation
• @ Symbols Overview
• Keyboard Shortcuts