The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/generate_srt.py uses subprocess.run to execute ffmpeg for embedding or burning subtitles into video files. The implementation safely passes arguments as a list rather than a single string, which effectively prevents shell injection attacks.
[EXTERNAL_DOWNLOADS]: Upon first use, the skill downloads large-scale machine learning models (approximately 3GB) from Hugging Face, including OpenAI's Whisper models and fine-tuned Hebrew models from the ivrit-ai organization. These are fetched from a well-known service for AI model hosting.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a surface for indirect prompt injection because it processes user-provided audio and video files containing untrusted speech data.
Ingestion points: scripts/generate_srt.py reads data through the WhisperModel.transcribe function.
Boundary markers: The system does not use specific delimiters for spoken content within the audio stream.
Capability inventory: The skill uses subprocess.run to call ffmpeg and writes subtitle files to the local disk.
Sanitization: Arguments for external commands are correctly tokenized in a list to prevent execution of injected shell commands. The risk is minimized as the output is restricted to transcription text and video rendering.

video-subtitles