The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of shell-based tools including yt-dlp, ffprobe, ffmpeg, and curl to download and process video data based on user-provided URLs and file paths. If the agent platform does not strictly sanitize these inputs, these command patterns are vulnerable to shell injection attacks.
[DYNAMIC_EXECUTION]: In Step 5, the skill constructs and executes a Python command string via python3 -c to perform transcription. Dynamically assembling executable code strings using variable file paths is a potential injection vector if the paths are maliciously crafted.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted content from external videos (transcripts, OCR, visual descriptions) and synthesizes this data into summaries and action items, creating a surface for adversarial content to influence agent behavior.
Ingestion points: SKILL.md (Steps 5 and 7) fetches transcripts and visual frame descriptions from external media.
Boundary markers: Absent; there are no instructions provided to help the agent distinguish between the skill's operational guidelines and potential instructions embedded within the processed video content.
Capability inventory: The skill possesses the ability to execute shell commands (Steps 3, 5, 7, 8) and write files to the local system (Steps 4 and 9).
Sanitization: While the skill includes an awk-based script to clean up YouTube VTT files (Step 5), it lacks comprehensive protection against adversarial text designed to manipulate LLM outputs.
[EXTERNAL_DOWNLOADS]: The skill fetches video files and metadata from various hosting platforms and uploads data to the Google Gemini API for analysis. These interactions represent data flows with external services that, while functional, involve the processing of remote content.

watch-video