The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill makes extensive use of dynamic code execution to interact with macOS APIs and UI elements.\n
Inline Swift code execution: swift -e 'import CoreGraphics ...' is used to enumerate windows.\n
AppleScript execution: osascript -e 'tell application ...' is used throughout SKILL.md to control application state (zoom, scroll, activate).\n- [COMMAND_EXECUTION]: The skill uses AppleScript (osascript) to perform complex UI automation.\n
Commands include activating applications, manipulating window properties, selecting spreadsheet cells, and opening files (open POSIX file).\n
Uses screencapture to save image data to the file system.\n- [PRIVILEGE_ESCALATION]: The troubleshooting documentation includes instructions for high-privilege system access.\n
references/permission-triage-template.md suggests using sudo sqlite3 to query the protected TCC (Transparency, Consent, and Control) database at /Library/Application Support/com.apple.TCC/TCC.db.\n
While intended for debugging permission issues, these instructions provide a template for an agent to access sensitive system privacy records.\n- [INDIRECT_PROMPT_INJECTION]: The skill is vulnerable to instructions embedded in external system data.\n
Ingestion points: The get_window_id.swift script (specifically window titles read via kCGWindowName) in scripts/get_window_id.swift.\n
Boundary markers: The skill does not implement delimiters or instructions to ignore potential commands embedded in window titles.\n
Capability inventory: The agent has access to osascript (system control), screencapture (file writes), and swift (system API access).\n
Sanitization: There is no sanitization or validation of the window titles before they are presented to the agent, allowing a malicious window title to potentially influence agent behavior.

capture-screen