The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external media (video files, URLs, RTSP/RTMP live streams, and desktop capture) using AI analyzers (VLM, transcript). This creates a surface for indirect prompt injection if the media contains malicious text or speech designed to influence the agent's behavior.
Ingestion points: Collection.upload() (URL/file), Collection.connect_rtstream() (RTSP/RTMP), and desktop capture sessions.
Boundary markers: No specific delimiters or 'ignore' instructions are provided in the example prompts when interpolating retrieved text or visual descriptions into higher-level LLM tasks like video.ask() or Collection.generate_text().
Capability inventory: The agent is instructed to execute Python scripts that interact with the videodb SDK, which performs network operations and file system reads (via .env).
Sanitization: There is no evidence of sanitization for the transcriptions or visual analysis outputs before they are processed by the agent.
[COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands and Python scripts to perform video processing tasks. It specifically uses a local utility script ws_listener.py to handle WebSocket events for desktop capture. These operations are within the scope of the skill's primary purpose and include instructions to maintain security, such as avoiding manual handling of the API key and scoping Python execution to the project directory.
[EXTERNAL_DOWNLOADS]: The skill downloads and installs the videodb Python SDK from standard registries. It also references a hosted MCP server at mcp.videodb.io. These resources are provided by the vendor (video-db) and are documented neutrally as they are required for the skill's core functionality.

videodb