The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes transcripts and subtitle data from external URLs (YouTube and other platforms). This content is untrusted and could potentially contain malicious instructions designed to manipulate the agent's behavior during or after processing the transcript.
Ingestion points: In scripts/vtd.js, transcript data is ingested via the youtube-transcript-plus library or by parsing subtitle files generated by yt-dlp (cmdTranscript function).
Boundary markers: The skill does not add boundary markers or "ignore instructions" warnings to the transcript text before returning it to the agent.
Capability inventory: The skill has the ability to write files to the local file system and execute subprocesses (yt-dlp, ffmpeg) via scripts/vtd.js.
Sanitization: The script performs basic cleanup of HTML-like tags and deduplicates subtitle segments, but it does not specifically sanitize the content for prompt injection patterns.
[COMMAND_EXECUTION]: The script scripts/vtd.js executes external binaries (yt-dlp and ffmpeg) using child_process.spawn. It provides a mechanism to pass arbitrary additional arguments to these binaries via the -- separator in the command-line interface.
Evidence: In scripts/vtd.js, the extra arguments parsed from the CLI are appended directly to the arguments array passed to spawn in the cmdDownload, cmdAudio, and cmdTranscript functions. While the use of spawn without a shell mitigates direct shell injection, the underlying binaries (particularly yt-dlp) possess extensive flags that could be misused if provided with malicious input.

video-transcript-downloader