The Agent Skills Directory

[COMMAND_EXECUTION]: Executes the agent command-line utility using the node:child_process.spawn function to run the Cursor sub-agent.
[COMMAND_EXECUTION]: Allows for the executable path to be overridden via the CURSOR_CLI_EXECUTABLE environment variable, enabling the use of alternative or local versions of the tool.
[COMMAND_EXECUTION]: The skill automatically includes the --trust and --approve-mcps flags when calling the CLI. These flags disable the underlying tool's built-in security confirmation prompts, which allows the delegated agent to execute shell commands and modify files without requiring user intervention.
[PROMPT_INJECTION]: The wrapPrompt function in scripts/lib/schema.mjs prepends specific instructions to the user's prompt to force a JSON response format, which overrides the sub-agent's standard conversational behavior.
[EXTERNAL_DOWNLOADS]: Fetches installation scripts from Cursor's official domain (cursor.com) as part of the setup process described in the documentation.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by ingesting external content and passing it to a sub-agent with significant capabilities.
Ingestion points: Input prompts are received via CLI arguments, the --prompt-file flag, or standard input (stdin).
Boundary markers: No specific delimiters or safety warnings are added to isolate the user-provided content within the prompt.
Capability inventory: The sub-agent has the capability to perform file system operations and execute shell commands, particularly when the --force flag is enabled.
Sanitization: The skill does not perform sanitization or validation on the input strings before they are interpolated into the prompt for the sub-agent.

cursor-cli