The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes external binaries, specifically yt-dlp and ffprobe, to download and inspect media files. These calls use list-based arguments to prevent command injection vulnerabilities, avoiding the use of a shell for execution.
[EXTERNAL_DOWNLOADS]: The skill fetches content from external URLs provided by the user. It relies on the yt-dlp binary to handle the connection to hundreds of supported media hosting platforms.
[DATA_EXPOSURE]: Media files are downloaded to a designated directory within the user's workspace (.mediaskills/downloads) or a path defined by the MEDIASKILLS_DATA_DIR environment variable. This is a standard practice for media processing agents.
[INDIRECT_PROMPT_INJECTION]: The skill has a vulnerability surface for indirect prompt injection via remote media metadata.
Ingestion points: Remote media titles and IDs are fetched from external URLs via yt-dlp in scripts/url.py.
Boundary markers: The yt-dlp template limits the title length to 200 bytes, but no explicit sanitization for LLM instructions is present.
Capability inventory: The skill and its siblings have capabilities for shell command execution (subprocess.run in scripts/url.py and scripts/_mediaskills_common.py) and file system access.
Sanitization: There is no evidence of metadata sanitization before the output is returned to the agent context, meaning an adversarial title could influence downstream agent behavior.

download