Cursor Desktop Guide

Cursor Desktop Guide

Context Architecture

Understand the priority of context loading to effectively manage the agent's knowledge:

1. Team Rules (Highest priority)
2. Project Rules (.cursor/rules/)
3. User Rules (Cursor Settings)
4. AGENTS.md (Project root and subdirectories)
5. Skills (Automatically discovered)

Core Components

AGENTS.md

• Purpose: Global instructions and context for the project.
• Best Practice: Keep compact (<100 lines). Use nested AGENTS.md for subdirectories.
• Location: Project root (AGENTS.md) or subdirectories (e.g., projects/foo/AGENTS.md).

Project Rules (.cursor/rules/)

• Purpose: Specific, versioned rules.
• Format: .md or .mdc (with YAML frontmatter).
• Types:
  • Always Apply: alwaysApply: true
  • Apply Intelligently: alwaysApply: false (Agent decides based on description)
  • Apply to Specific Files: globs: ["pattern"]
  • Apply Manually: @rule-name mention

Skills (.cursor/skills/)

• Purpose: Portable, versioned packages for domain-specific workflows.
• Structure: Directory with SKILL.md, scripts/, assets/, references/.
• Usage: Multi-step workflows requiring resources or scripts.

Commands (.cursor/commands/)

• Purpose: One-off actions triggered by /command in chat.
• Format: Simple Markdown files.

Subagents

• Purpose: Isolated AI assistants for heavy tasks or parallel execution.
• Usage: Use Task tool.
• When to use: Context isolation needed, parallel work, or specialized long-running tasks.
• Cost: Each subagent has its own context window (token overhead).

MCP Servers & Integrations

• Local: Use MCP for stateful connections (e.g., Slack) or CLI tools via Shell for request-response (e.g., Jira, GitLab).
• Cloud Agents: Not used in this system (require remote repo).

Hooks (.cursor/hooks.json)

• Purpose: Scripts triggered on agent events (e.g., preToolUse, afterFileEdit).
• Types: Command-based (shell script) or Prompt-based (LLM evaluation).

Best Practices

1. AGENTS.md is the entry point, not an encyclopedia.
2. Use Rules with globs for file formats (e.g., ADRs).
3. Use Skills for complex workflows (reviews, reports).
4. Use Commands for simple rituals (/morning).
5. Use Subagents only when context isolation is strictly necessary.
6. Use .cursorignore to exclude archives and cache from Semantic Search.
7. Use Max Mode (1M tokens) for heavy workflows.

References

• Full documentation: references/REFERENCE.md
• Online up-to-date documentation: https://www.cursor.com/docs