The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies heavily on executing shell commands, including yt-dlp, ffmpeg, and ffprobe. It incorporates user-provided variables such as $URL and local file paths directly into these command strings. While the instructions use double quotes (e.g., "$URL"), there remains a risk of command injection if the underlying agent does not properly escape or validate these inputs before shell execution.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external video sources, which creates a vulnerability to indirect prompt injection.
Ingestion points: Video content (frames processed by vision models) and metadata JSON files generated from external URLs (Step 2 and Step 4 in SKILL.md).
Boundary markers: The instructions do not define explicit delimiters or instructions to ignore embedded commands within the extracted visual text or metadata descriptions during the synthesis phase.
Capability inventory: The skill has broad capabilities, including file system read/write access and the ability to execute system commands like ffmpeg and yt-dlp.
Sanitization: There is no mention of sanitizing or filtering the metadata fields (like title or description) or the text transcribed from video frames before they are used to generate the final analysis JSON.
[EXTERNAL_DOWNLOADS]: The skill instructs the agent to download and install external software packages (ffmpeg, yt-dlp) using system package managers like brew, apt, or pipx. While these are well-known and reputable tools, the instruction to install software at runtime is a significant capability that should be monitored.

analyze-video