The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes the external command-line tools yt-dlp and ffmpeg via the Python subprocess module to process media and verify tool availability. Evidence includes subprocess.run calls in scripts/doctor.py, scripts/probe_url.py, and scripts/check.py used for version checks and metadata probing.
[DATA_EXFILTRATION]: The skill allows the use of user-provided cookie files (e.g., ~/.config/yt-dlp/cookies.txt) for authenticated or age-restricted content. Access is restricted to user-supplied absolute paths, and the skill includes comprehensive rules against storing, committing, or sharing secrets within the repository. Documentation is neutral, as the skill facilitates legitimate media access using standard tool configurations and enforces user ownership of credentials.
[PROMPT_INJECTION]: The skill ingests untrusted metadata from external URLs, representing a potential surface for indirect prompt injection. 1. Ingestion points: External video metadata (titles, descriptions, etc.) fetched from video hosting sites via the scripts/probe_url.py script. 2. Boundary markers: Data is ingested via structured JSON output from a CLI wrapper, though no explicit prompt delimiters or 'ignore instructions' warnings are mentioned in the SKILL.md instructions. 3. Capability inventory: Subprocess execution for media downloads and file system write access to ~/Downloads. 4. Sanitization: scripts/probe_url.py acts as a filter, extracting only specific, expected fields (title, id, duration, etc.) from the raw yt-dlp JSON output before returning it to the agent, which mitigates the risk of malicious instructions in metadata influencing agent behavior.

yt-dlp