The Agent Skills Directory

[COMMAND_EXECUTION]: The skill helper scripts (render.py, transcribe.py, timeline_view.py, grade.py) make extensive use of subprocess.run() to invoke ffmpeg and ffprobe. These commands are used for audio extraction, frame sampling, color grading, and final video rendering. While this is the intended functionality, it involves executing shell commands with arguments derived from project metadata.
[EXTERNAL_DOWNLOADS]: The skill transmits audio data to Volcengine's Recording File Recognition service at openspeech.bytedance.com for transcription. This is a well-known service used for the skill's primary ASR functionality.
[DATA_EXFILTRATION]: Audio extracted from source videos is base64-encoded and sent to the Volcengine API. This constitutes a transfer of local media data to a remote service for processing.
[REMOTE_CODE_EXECUTION]: The skill facilitates the generation and execution of Python scripts for Manim animations and uses npx to execute Node.js-based rendering tools (HyperFrames and Remotion) via sub-agents. This dynamic code generation and execution is a core feature for creating video overlays.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection from video transcripts. Malicious instructions could be included in the speech of a video, which is then transcribed and processed by the editor sub-agent.
Ingestion points: Video transcripts are consolidated into takes_packed.md and provided as context to the editor sub-agent.
Boundary markers: The editor sub-agent brief defined in SKILL.md does not utilize delimiters or specific instructions to ignore embedded commands within the transcript text.
Capability inventory: The skill has broad capabilities including filesystem writes in the project directory, shell command execution via helper scripts, and spawning sub-agents for script execution.
Sanitization: No sanitization or validation is performed on the transcript content before it is ingested by the LLM.

video-use