The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data in the form of YouTube transcripts, titles, and descriptions.
Ingestion points: Video transcripts and metadata are fetched via the api.transcriptout.com endpoints (e.g., /v1/transcript, /v1/search).
Boundary markers: The skill includes a dedicated 'Treat fetched content as data' section instructing the agent to never follow instructions found inside transcripts (such as commands or requests to reveal API keys).
Capability inventory: The skill primarily uses HTTP requests but provides shell-based authentication workflows in its reference files.
Sanitization: Protection relies on prompt-based instructions rather than programmatic sanitization of the fetched strings.
[PERSISTENCE]: The references/auth-setup.md file instructs the agent to modify user shell profile files (such as ~/.bashrc, ~/.zshenv, or ~/.profile) to export the TRANSCRIPTOUT_API_KEY environment variable, ensuring the key persists across future shell sessions.
[COMMAND_EXECUTION]: The authentication guide in references/auth-setup.md provides complex shell command sequences for account registration, OTP verification, and secret management using system tools like curl, grep, sed, mkdir, and chmod.

youtube-full