The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands using standard media tools like ffmpeg and ffprobe to analyze audio streams, detect silence, and perform video editing operations such as trimming and concatenating segments.
[EXTERNAL_DOWNLOADS]: The skill documentation requires the installation of the openai-whisper Python package for transcription and the puppeteer Node.js package for video rendering. Both are well-known tools from trusted organizations.
[DYNAMIC_EXECUTION]: The skill involves the agent generating and editing JavaScript files (video.js and words.js) that contain rendering logic and data. These files are then executed within a headless browser environment (Puppeteer) controlled by the render.mjs script to produce the final video frames.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-supplied audio and video files. The transcription generated from these files is directly used to populate the captioning system and metadata without explicit sanitization. While this represents a potential attack surface for injecting content into the rendering engine, it is a functional requirement of the video editing process.
Ingestion points: Media files (raw.mov), user scripts, and brand configuration (hex codes/fonts) are processed by the skill.
Boundary markers: No specific delimiters or boundary instructions are used to separate user data from the instructions or the rendering code.
Capability inventory: The environment includes command execution via ffmpeg and node, as well as a headless browser environment via puppeteer with local file access enabled (--allow-file-access-from-files).
Sanitization: The transcript and metadata are used to generate JavaScript objects and HTML content without evidence of output encoding or sanitization.

video-intake