The Agent Skills Directory

[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection through its core functionality of analyzing external YouTube videos.
Ingestion points: The youtube_url provided by the user is passed directly to the Gemini API in scripts/youtube_to_heygen_video.py for content analysis.
Boundary markers: Absent. The SCRIPT_EXTRACTION_PROMPT provides strict instructions to the model but does not wrap the external data in delimiters or include instructions to ignore embedded commands within the video content.
Capability inventory: The skill has the capability to write to a PostgreSQL database (psycopg2) and send generated data to the HeyGen API (requests.post).
Sanitization: Absent. The script performs basic character replacement (removing newlines) but does not validate the LLM output for adversarial instructions before using it in subsequent API calls.
[DATA_EXFILTRATION]: The skill performs network operations to external services to fulfill its purpose.
Evidence: It sends requests to the HeyGen API (api.heygen.com) and interacts with the Google Gemini API. While these are well-known services, they constitute external data flow from the agent's environment.
[CREDENTIALS_UNSAFE]: The skill follows secure practices for credential management.
Evidence: Both SKILL.md and the Python implementation correctly instruct the user to store sensitive tokens like GEMINI_API_KEY, HEYGEN_API_KEY, and DATABASE_URL in environment variables or a .env file, avoiding hardcoded secrets in the code.

youtube-to-heygen-video