The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads the nanobanana extension from https://github.com/gemini-cli-extensions/nanobanana and an AI model from hexgrad/Kokoro-82M on HuggingFace to enable image editing and offline TTS capabilities.
[COMMAND_EXECUTION]: Executes several system-level commands including ffmpeg for video assembly, soffice (LibreOffice) for PPTX to PDF conversion, and pdftoppm for image extraction from PDF files.
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: Research papers and user outlines are used as the primary data sources for drafting video scripts and slide content in Stage 1.
Boundary markers: The skill documentation notes that the "human drives all outline and visual decisions," but there are no technical delimiters or "ignore" instructions provided to the agent for processing untrusted paper content.
Capability inventory: The agent has the ability to execute shell commands (ffmpeg, soffice, pdftoppm) and network operations via TTS providers.
Sanitization: There is no evidence of sanitization or filtering of the research paper text before it is interpolated into the generation scripts.
[DYNAMIC_EXECUTION]: Performs runtime installation of CLI extensions using gemini extensions install and dynamically loads large model weights from HuggingFace via the kokoro library.

academic-presentations