The Agent Skills Directory

[COMMAND_EXECUTION]: The skill includes a Python script scripts/slice_stream_frame.py that executes yt-dlp and ffmpeg to process video data. These operations are performed using subprocess.run() with a list of arguments, which is the recommended secure method for executing shell commands as it prevents command injection vulnerabilities.
[DATA_EXPOSURE]: The helper script uses the --cookies-from-browser flag with yt-dlp to access streaming content that may require authentication (e.g., age-restricted videos). While this accesses local browser cookie data, it is a standard feature of the yt-dlp tool for its intended purpose, and the skill does not contain any logic to exfiltrate this sensitive information to external servers.
[INDIRECT_PROMPT_INJECTION]: Because the skill processes external data (video frames and LiveChat transcripts), it is theoretically susceptible to indirect prompt injection. However, the skill explicitly includes defensive instructions, such as 'Zero-Thinking JSON Enforcement' and 'Main Display vs Chat Isolation,' to guide the Vision Language Model (VLM) to ignore potentially malicious or distracting content in the chat overlay and respond only in structured data formats.
[EXTERNAL_DOWNLOADS]: The skill downloads small sections of video content from YouTube using yt-dlp. These downloads are targeted (using --download-sections) to minimize bandwidth and are directed from a well-known service (YouTube) to the local environment for analysis, which is consistent with the skill's stated purpose.

verifying-stream-ground-truth