The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill uses npx @midscene/android@1 to download and execute the Midscene Android automation package from the npm registry. This is the standard method for running the Midscene CLI.
[COMMAND_EXECUTION]: The skill provides tools to execute adb shell commands on connected Android devices via the runadbshell command. This level of access is necessary for low-level device control such as inspecting battery status or launching specific activities.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection (Category 8) as it processes untrusted visual data from external applications and websites.
Ingestion points: Screen content is captured via take_screenshot and processed by the AI to determine actions.
Boundary markers: None identified in the instructions for isolating screen content from agent instructions.
Capability inventory: The agent can execute arbitrary ADB shell commands, interact with UI elements, and access the local file system to save/read screenshots.
Sanitization: No specific sanitization is performed on the visual data before it is interpreted by the vision model.
[CREDENTIALS_UNSAFE]: The skill instructions provide best-practice guidance for managing sensitive API keys for AI models (e.g., Gemini, Qwen) using .env files. No hardcoded credentials were detected in the skill content.

android-device-automation