The Agent Skills Directory

[COMMAND_EXECUTION]: Executes shell commands via ffmpeg and ffprobe to concatenate video blocks, crop platform variants, and verify media file metadata.
[COMMAND_EXECUTION]: Utilizes the vidjutsu command-line interface to perform file uploads to cloud storage and manage service authentication.
[EXTERNAL_DOWNLOADS]: Communicates with external API endpoints at api.wavespeed.ai, api.vidjutsu.ai, and generativelanguage.googleapis.com to perform image generation, video rendering, and compliance scanning.
[DATA_EXFILTRATION]: Transmits locally generated image frames and video clips to Wavespeed and VidJutsu cloud infrastructure. This behavior is documented as a requirement for the skill's media generation and compliance verification features.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection in recipes/04-define-clips.md and references/clip-spec.md. It interpolates untrusted user dialogue directly into the instructions for the Seedance video model using simple single-quote delimiters ('<line>').
Ingestion points: Dialogue lines from script.md and clips/block_*.json.
Boundary markers: Uses single quotes for dialogue, but lacks explicit "ignore embedded instructions" warnings for the model.
Capability inventory: The skill has broad capabilities including subprocess execution (ffmpeg, vidjutsu) and authenticated network operations.
Sanitization: No escaping or validation is performed on the dialogue strings before interpolation into the prompt template.

wide-cam-podcast