The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill utilizes npx -y @midscene/ios@1 to download and run the Midscene automation package from the public npm registry. It also fetches reference images from external sources such as GitHub's asset domain for visual comparison tasks.
[COMMAND_EXECUTION]: The skill uses the Bash tool to execute various CLI commands for device connection, app launching, taking screenshots, and performing UI interactions. It also supports sending direct HTTP requests to a WebDriverAgent endpoint via the runwdarequest command.
[INDIRECT_PROMPT_INJECTION]: The skill processes live screen content through the take_screenshot and act commands to determine its next steps. This introduces a potential attack surface where an application being automated could display malicious text or UI elements designed to trick the agent's vision reasoning into performing unintended actions.
Ingestion points: Screen captures generated by the take_screenshot command (SKILL.md).
Boundary markers: None identified; the agent relies on its internal instruction following to differentiate between UI labels and actual task instructions.
Capability inventory: The skill has the ability to execute shell commands, interact with the device filesystem, and manipulate the UI of a connected iOS device.
Sanitization: None; the skill passes natural language descriptions directly to the Midscene CLI for execution.
[CREDENTIALS_UNSAFE]: The documentation guides users to configure sensitive environment variables such as MIDSCENE_MODEL_API_KEY. While it does not contain hardcoded secrets, it explicitly handles high-value credentials required for AI model access.

ios-device-automation