The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documentation describes how to handle real-time data from participants, which constitutes an attack surface for indirect prompt injection if an AI agent processes this data.
Ingestion points: Instructions in examples/chat.md, examples/command-channel.md, and references/advanced.md describe how to listen for and process messages from the chat-on-message and command-channel-message events.
Boundary markers: The provided code examples for handling incoming messages do not include specific delimiters or markers to help an agent distinguish between system instructions and untrusted user input.
Capability inventory: The SDK allows for powerful operations such as media control, session lifecycle management, and participant manipulation, which are outlined in references/web-reference.md.
Sanitization: Examples shown focus on functionality (e.g., using JSON.parse for commands) but do not demonstrate content sanitization or safety filtering for participant data.
[EXTERNAL_DOWNLOADS]: The skill references various external packages and scripts required for the Video SDK implementation.
The skill instructs users to install the @zoom/videosdk and @zoom/videosdk-react packages from the standard NPM registry.
Examples in SKILL.md and examples/framework-integrations.md point to official Zoom repositories on GitHub for boilerplate authentication and framework starter projects.
SDK runtime assets are fetched from vendor-owned domains such as source.zoom.us, as described in the configuration guides in references/video.md and troubleshooting/common-issues.md.

video-sdk/web