Cursor AI System Prompt

Overview

Cursor (by Anysphere) is a code editor built on VS Code that uses Claude and GPT-4o for code completions, chat, and agent mode. The community has reverse-engineered significant portions of its system prompt through responses that “leak” it under certain prompting conditions.

This entry documents the known structural patterns — useful for building your own coding assistant or understanding how Cursor constrains model behaviour.

Key behavioral Instructions

Code generation mode

Cursor instructs the model to:

• Output code without explanation unless asked
• Prefer minimal diffs over rewrites
• Match the indentation, style, and conventions of the surrounding file
• Never add comments unless they exist in the original

Chat mode

In chat (Cmd+L / Ctrl+L), the prompt shifts to:

• Allow longer explanations
• Reference specific file locations using the @file syntax
• Prefer code blocks with language identifiers

Agent mode

In agent/composer mode, the system prompt enables:

• Multi-step planning before execution
• Tool calls for terminal, file creation, and search
• Verification loops (“run the tests, fix any failures”)

Patterns worth borrowing

1. Mode-specific behavior — switch instructions based on the context (inline completion vs chat vs agent). Cursor does this explicitly.
2. Style mirroring — instruct your model to adapt to the user’s existing code rather than imposing its own style.
3. Diff-first output — for code edits, a targeted diff is almost always better than a full file rewrite.

Community resources

• nicholmikey/cursor-tools — community-documented prompt analysis
• Cursor changelog — Cursor announces behavior changes here; significant ones reflect prompt changes