The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill performs dynamic compilation of source code at runtime to interact with OS-level APIs. On macOS, it uses the swift interpreter to run .swift files (e.g., macos_permissions.swift, macos_window_info.swift) for permission checks and window metadata retrieval. On Windows, the take_screenshot.ps1 script uses Add-Type to compile C# code for P/Invoke access to the Win32 API.
[COMMAND_EXECUTION]: The skill executes multiple system utilities via shell and subprocess calls to perform screen captures. This includes screencapture on macOS, scrot, gnome-screenshot, and import (ImageMagick) on Linux, and powershell.exe on Windows. The PowerShell helper specifically uses -ExecutionPolicy Bypass to allow the execution of the local script.
[INDIRECT_PROMPT_INJECTION]: The skill provides a surface for indirect prompt injection by capturing visual data from the user's environment. If the resulting screenshots contain malicious text and are subsequently processed by an agent with vision capabilities, the agent could potentially follow instructions embedded in the image.
Ingestion points: Screen captures generated by take_screenshot.py and take_screenshot.ps1 from the user's desktop, applications, or windows.
Boundary markers: None; the captured data is written directly to standard image formats (PNG/JPG).
Capability inventory: The skill possesses capabilities to write files to the file system and execute arbitrary shell commands through its helper scripts.
Sanitization: The Python script includes basic escaping for application names when using osascript, but otherwise relies on the agent to provide valid paths and identifiers.

screenshot