The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The script scripts/jump_cut_vad_singlepass.py uses torch.hub.load with trust_repo=True to download and execute code and models from the snakers4/silero-vad repository on GitHub at runtime.
[COMMAND_EXECUTION]: Several scripts, including auto_captions.py, compress_video.py, insert_3d_transition.py, jump_cut_vad_singlepass.py, simple_video_edit.py, vertical_crop.py, and youtube_clip.py, use subprocess.run to execute shell commands such as ffmpeg, ffprobe, yt-dlp, and npx. While these are standard for video processing, they involve executing commands with inputs derived from file paths and user-provided metadata.
[EXTERNAL_DOWNLOADS]: The scripts/youtube_clip.py script downloads video content from YouTube, a well-known service, using the yt-dlp tool. Additionally, insert_3d_transition.py requires running npm install to fetch dependencies for the Remotion rendering engine.
[DATA_EXFILTRATION]: The scripts/simple_video_edit.py script is designed to upload processed video files and generated metadata to the Auphonic web service, which is the intended functionality of the skill's upload feature.
[INDIRECT_PROMPT_INJECTION]: The skill generates video transcripts from untrusted external sources and interpolates them into prompts for the Anthropic Claude API in simple_video_edit.py and youtube_clip.py to generate summaries and chapters.
Ingestion points: Video transcripts generated via the Whisper model from potentially untrusted video files or YouTube downloads in scripts/simple_video_edit.py and scripts/youtube_clip.py.
Boundary markers: The prompts do not use explicit delimiters or "ignore instructions" warnings to isolate the transcript content from the system instructions.
Capability inventory: The skill has capabilities for file system access, network communication (API uploads), and shell command execution.
Sanitization: There is no evidence of sanitization or escaping of the transcript text before it is inserted into the LLM prompt template.
[DYNAMIC_EXECUTION]: The scripts/insert_3d_transition.py script uses npx remotion render to execute JavaScript-based rendering logic, and jump_cut_vad_singlepass.py dynamically loads Silero VAD code using Torch Hub.

video-edit