The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The setup.sh script downloads an installation script from https://astral.sh/uv/install.sh. This is the official distribution endpoint for the uv package manager, a well-known tool in the Python ecosystem.- [COMMAND_EXECUTION]: The youtube_handler.py script executes the yt-dlp utility using subprocess.run to fetch video metadata. The command uses sys.executable to ensure it runs within the established virtual environment.- [DATA_EXFILTRATION]: The skill performs legitimate network requests to fetch page titles and YouTube transcripts using urllib.request and the youtube-transcript-api. These operations are required for the skill's primary purpose of analyzing external content.- [PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface due to its core function of ingesting untrusted data for processing and outline generation.\n
Ingestion points: External content enters the agent's context through scripts/convert_to_markdown.py (which handles files and general URLs) and scripts/youtube_handler.py (which handles YouTube transcripts).\n
Boundary markers: The instructions in SKILL.md do not implement explicit delimiters or instructions to treat the converted content as data rather than instructions, potentially allowing embedded malicious prompts to influence the agent.\n
Capability inventory: The skill scripts can execute shell commands via subprocess.run (in youtube_handler.py), perform network fetches, and write files to the project directory.\n
Sanitization: While the skill uses regex to sanitize YouTube video IDs, it relies on the underlying parsing libraries (docling, yt-dlp) for the safe handling of document and media structures.

planning-content