The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Fetches site favicons from Google's service and page screenshots from Microlink's API. It also initiates a WebSocket connection to Soniox for real-time dictation transcription. These integrations target well-known services and follow secure implementation patterns, such as using single-use server-side keys for authentication.
[INDIRECT_PROMPT_INJECTION]: The skill manages a significant ingestion surface for untrusted data as part of its core functionality.
Ingestion points: Processes assistant responses, tool outputs, and scraped content from external sites (Facebook Marketplace, Ads, and Job listings).
Boundary markers: Employs distinct UI components (cards, chips) to separate agent content from external data and explicitly treats all quoted information as untrusted.
Capability inventory: Includes user-triggered actions such as task creation based on suggestions and approval/denial of tool calls.
Sanitization: Implements strict sanitization using rehype-sanitize and a customized rehype-harden configuration within the streamdown renderer. This setup explicitly blocks dangerous protocols like javascript:, data:, file:, and vbscript:, while requiring explicit user confirmation before opening external URLs.
[DYNAMIC_EXECUTION]: Uses dynamic imports to lazily load the @streamdown/code highlighter module and initializes a browser AudioWorklet for PCM audio processing. The worklet source is defined as a static internal string, and these patterns are standard practices for the intended UI features.

agent-chat-ui