The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes UI trees and screenshots from a connected Android device using mobilerun device ui and mobilerun device screenshot. This data, which is untrusted and externally controlled (by the apps on the device), is used by the agent to decide on its next actions. There are no sanitization steps or explicit boundary markers found in the scripts to prevent instructions embedded in the device UI from influencing the agent's behavior.
Ingestion points: scripts/mr-snap.sh captures the UI tree and screenshot from the device.
Boundary markers: None present in instructions or scripts.
Capability inventory: The skill possesses extensive device control capabilities via mobilerun and adb, including tapping, typing, swiping, and launching applications.
Sanitization: No sanitization of UI text or content descriptions before processing.
[PROMPT_INJECTION]: The skill includes explicit behavioral instructions in references/stealth-and-human-pacing.md and references/writing-like-a-human.md that direct the agent to intentionally mask its AI nature. It provides a list of "model tells" to avoid (such as em-dashes, moralizing wrap-ups, and specific vocabulary like "delve" or "robust") and mandates specific pacing and jitter techniques to mimic human behavior and evade platform-level bot detection heuristics.
[COMMAND_EXECUTION]: The skill makes extensive use of Bash scripts to execute mobilerun and adb commands. While these are necessary for the skill's stated purpose of device automation, they allow for deep interaction with the connected mobile device's system settings (e.g., modifying accessibility services via adb shell settings put secure).

mobilerun-control