The Agent Skills Directory

[COMMAND_EXECUTION]: The documentation in references/transcripts.md recommends using sudo apt install for dependency management on Linux, which involves privilege escalation.
[COMMAND_EXECUTION]: The skill heavily relies on executing shell commands and external CLI tools (yt-dlp, eyeD3, whisper) via subprocess calls in both Bash and Ruby.
[EXTERNAL_DOWNLOADS]: The skill fetches and installs multiple external packages and libraries from PyPI, Homebrew, and APT. It also uses bundler/inline in scripts/ytmp3 to dynamically install Ruby gems (thor, gum) from RubyGems at runtime.
[EXTERNAL_DOWNLOADS]: The ytmp3 script uses curl to download content from arbitrary user-provided URLs.
[PROMPT_INJECTION]: The workflow examples in references/transcripts.md demonstrate insecure interpolation of shell-derived variables into Python command strings, creating a vulnerability to indirect prompt injection via malicious YouTube metadata.
Ingestion points: Video titles and subtitle filenames extracted via yt-dlp in references/transcripts.md and scripts/ytmp3.
Boundary markers: Absent; untrusted data is directly embedded into shell and Python strings.
Capability inventory: Extensive system access including file manipulation (rm, mv, mkdir), network requests (curl, yt-dlp), and shell execution (python3 -c, bash).
Sanitization: Relies on basic tr commands which do not fully sanitize against sophisticated injection attacks.

youtube