The Agent Skills Directory

[SAFE]: The skill behaves as described, providing a video creation workflow. It uses Model Context Protocol (MCP) tools for media manipulation and scraping, which are used appropriately within the skill's scope.
[PROMPT_INJECTION]: The skill ingests untrusted data from social media profiles to personalize content. This presents a surface for indirect prompt injection, which is a standard risk for skills processing external web data.
Ingestion points: Untrusted data enters the context via scrape_social on user-provided handles (e.g., in formats/talking.md Stage 1).
Boundary markers: The instructions do not specify explicit delimiters to separate untrusted data from system instructions.
Capability inventory: Includes media editing tools such as edit_trim, edit_concat, and add_captions.
Sanitization: The agent summarizes input data before use, providing some isolation.
[COMMAND_EXECUTION]: Video and audio editing are performed through specialized tools that do not allow arbitrary system command execution.
[EXTERNAL_DOWNLOADS]: The skill retrieves media for analysis and connects to the vendor's teleprompter service. These operations are essential to the workflow and use legitimate vendor domains.

content-director