The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from YouTube video metadata (titles, descriptions) and transcripts, which are interpolated into LLM prompts for qualification and review.
Ingestion points: scripts/search_and_summarize.py fetches video metadata via the DuckDuckGo Search CLI and transcripts via the youtube-transcript-api library.
Boundary markers: The skill uses triple backtick delimiters in references/review_prompt.md to wrap transcript content and structured fields in references/qualify_prompt.md.
Capability inventory: scripts/search_and_summarize.py utilizes subprocess.run for searches and Path.write_text for exporting research notes to local files.
Sanitization: The script performs transcript cleaning (removing timestamp-only lines) and chunks long transcripts to manage input size and potential injection complexity.
[COMMAND_EXECUTION]: The skill executes external commands using subprocess.run in scripts/search_and_summarize.py to call the ddgs CLI for video searching. The command is passed as a list, which mitigates standard shell command injection vulnerabilities.
[DYNAMIC_EXECUTION]: The test suite in tests/test_search_and_summarize.py uses the __import__ function for dynamic module loading. This is restricted to the standard library re module for transcript cleaning tests.

youtube-topic-research