The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on the agent-device CLI tool to perform operations such as listing devices, capturing UI snapshots, and simulating user interactions (press, scroll, navigation). These commands are used to gather data about mobile app structures.
[EXTERNAL_DOWNLOADS]: The documentation directs users to install the agent-device package via npm. This tool appears to be a core component provided by the author to enable the skill's functionality.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface as it ingests and processes text data (labels, identifiers, and placeholders) from the UI of external mobile applications. While the skill defines 'Safety Rules' to prevent the agent from performing dangerous actions like clicking 'Delete' or 'Purchase' buttons, there are no specific sanitization steps to prevent malicious text within a mobile app's UI from being interpreted as instructions by the underlying AI model.
Ingestion points: UI data is ingested via the agent-device snapshot --json command in SKILL.md.
Boundary markers: No explicit delimiters are used to separate ingested UI text from the agent's internal instructions.
Capability inventory: The skill can execute various device interaction commands through the agent-device CLI.
Sanitization: Action-level sanitization is present via 'Blocked identifiers' and 'Blocked labels' lists to prevent high-risk interactions during autonomous mapping.

cartographer