The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill describes workflows where an AI agent ingests untrusted scene and data text to programmatically build scene hierarchies. This creates a surface for indirect prompt injection attacks.
Ingestion points: Workflow 11 in SKILL.md and the builder module documentation describe the ingestion of untrusted strings representing scene structures.
Boundary markers: The skill requires the use of explicit delimiters (<<<CONTEXT_START>>> and <<<CONTEXT_END>>>) to wrap untrusted context.
Capability inventory: The agent is granted file-writing capabilities through the builder module scripts (create_scene.py, add_node.py, save_scene.py).
Sanitization: The instructions mandate sanitizing node names to be alphanumeric/underscored and ensuring all paths are prefixed with res://.
[EXTERNAL_DOWNLOADS]: The agent-vision diagnostic module requires the installation of external Python dependencies to facilitate its screen capture and processing features.
The documentation identifies several standard Python libraries used for cross-platform window and screen capture, including mss, pywin32, dxcam, and pyobjc-framework-Quartz.
The skill also references OS-level utilities such as wmctrl, xdotool, and grim for Linux-based environments.

godot-master