The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a bundled Zsh script scripts/window-shot.sh and the system utility screencapture to capture application windows.
[DYNAMIC_EXECUTION]: The scripts/window-shot.sh script uses the swift interpreter to execute dynamically provided code via a heredoc to interact with the macOS CoreGraphics API for window identification.
[INDIRECT_PROMPT_INJECTION]: The skill's primary function involves processing visual information from arbitrary user applications. This represents an attack surface where text content within those applications (e.g., task names, document content, or settings labels) could be interpreted by the agent as instructions.
Ingestion points: Visual input and screenshots from external applications via mcp__computer-use__* and window-shot.sh (SKILL.md).
Boundary markers: The instructions include a 'Look, do not touch' directive and state 'Type, create, change, or delete only when the user says so', which serves as a weak boundary for the agent (SKILL.md).
Capability inventory: The agent has the ability to execute shell scripts, take screenshots, and use computer-interaction tools (SKILL.md, scripts/window-shot.sh).
Sanitization: There is no explicit sanitization or filtering of the text content extracted from the captured screens before it is processed by the agent.
[PRIVILEGE_ESCALATION]: The skill requires the environment to have 'Screen Recording' and 'Accessibility' permissions enabled for the terminal or agent host. This grants the skill high-privilege access to the user's desktop environment and is a necessary prerequisite for its intended functionality.

clone