The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external sources (YouTube transcripts) and provides it to the agent for summarization or analysis. The transcript content is not wrapped in boundary markers, which could allow embedded instructions in the video subtitles to manipulate the agent's response.
Ingestion points: scripts/get_transcript.py fetches and prints the content of YouTube transcripts.
Boundary markers: None present. The agent is instructed to read the raw output of the script.
Capability inventory: The skill uses subprocess.run to execute yt-dlp and has the ability to read transcript files.
Sanitization: The clean_vtt function removes formatting tags and timestamps but does not sanitize the text for natural language instructions.
[COMMAND_EXECUTION]: The skill relies on the yt-dlp command-line tool. It executes this tool using subprocess.run in scripts/get_transcript.py. While the execution is handled via a command list (reducing shell injection risk), it still involves running external binaries with user-provided URL arguments.

youtube-watcher