The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from YouTube videos, including auto-generated or human-provided transcripts and visual frames.
Ingestion points: Data enters the agent context via yt-dlp JSON output (transcripts) and ffmpeg frame extraction (images) from external URLs.
Boundary markers: The instructions do not specify the use of clear delimiters or XML tags to wrap ingested transcript content to prevent the agent from obeying instructions embedded in the video text.
Capability inventory: The skill has access to the Bash tool, allowing for shell command execution, and the Read tool for file system access.
Sanitization: While the skill mentions "normalization" of output, it lacks explicit security sanitization to filter out prompt injection payloads from transcripts.
[EXTERNAL_DOWNLOADS]: The skill references and suggests the use of external software and repositories not belonging to a pre-verified trusted vendor.
Evidence: The skill suggests updating yt-dlp via pip install -U yt-dlp and mentions an external repository github.com/bradautomates/claude-video for execution, suggesting brew install for its dependencies.
[COMMAND_EXECUTION]: The skill utilizes the Bash tool to execute complex shell commands, including yt-dlp for metadata and transcript extraction.
Evidence: Execution of yt-dlp with various flags and piping output to files in /tmp/.
[DYNAMIC_EXECUTION]: The skill uses inline scripts to process data at runtime.
Evidence: A python3 -c one-liner is used to parse and regex-process JSON data produced by yt-dlp.

video-ingest