The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill requires installing external components from various sources, including the nanobanana extension from a GitHub repository (https://github.com/gemini-cli-extensions/nanobanana). It also downloads configuration and assets from well-known services, such as the Kokoro model from HuggingFace (hexgrad/Kokoro-82M) and standard packages through official registries like PyPI and Homebrew.
[REMOTE_CODE_EXECUTION]: The skill installs and executes third-party code through the gemini extensions system and Python environment. Specifically, the nanobanana extension and the Kokoro model are downloaded from external repositories and executed at runtime to perform image and audio generation.
[COMMAND_EXECUTION]: The skill makes extensive use of shell commands to manage media conversion and assembly. This includes using ffmpeg for video production, soffice (LibreOffice) for converting PPTX files to PDF, and pdftoppm for slide image extraction. It also includes instructions for process management via pkill.
[INDIRECT_PROMPT_INJECTION]: The skill processes external research papers and user-provided outlines to generate narration scripts and visual prompts, creating an attack surface for indirect prompt injection.
Ingestion points: Research papers and user-provided slide plans processed in Stage 1 as described in SKILL.md.
Boundary markers: The skill does not define specific delimiters or instructions to isolate the untrusted paper content from the agent's core processing logic.
Capability inventory: The skill has access to file writing, shell command execution (ffmpeg, soffice), and network access via multiple TTS APIs and Gemini extensions.
Sanitization: No explicit sanitization or filtering of the research paper content is documented before the data is interpolated into the agent's prompts.

academic-presentations