The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill handles untrusted external data by fetching transcripts and metadata from YouTube.
Ingestion points: Transcripts and video metadata are ingested via yt-dlp in references/fetch.md and processed by the subagent.
Boundary markers: The instructions in SKILL.md and references/fetch.md explicitly command the agent to treat transcript cues as data and never as instructions, specifically highlighting that phrases like "ignore previous instructions" found in the text must be ignored.
Capability inventory: The skill has the capability to run shell commands (yt-dlp, grep, awk), execute a Python cleaning script, and manage file operations in temporary directories.
Sanitization: The scripts/clean_vtt.py script uses regular expressions and html.unescape to strip formatting and tags, ensuring the resulting transcript is plain text and not interpreted as code or structured directives.
[COMMAND_EXECUTION]: The skill utilizes system utilities including yt-dlp, grep, and awk for metadata extraction and transcript processing. The implementation follows secure practices by using double-quoting for shell variables and reading untrusted metadata (like video titles) from files using cat within command substitutions, which prevents the shell from re-evaluating the content as executable commands.

youtube-transcript