The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute various system commands, including ffmpeg for audio compression and segmenting, ffprobe for media duration measurement, and curl.exe for interacting with the Groq API. These operations are performed locally on the user's system to prepare data for transcription.
[EXTERNAL_DOWNLOADS]: The skill performs network operations to api.groq.com to send audio files for transcription (Whisper) and text for processing (LLM). Groq is a well-known service, and the skill includes comprehensive warnings about the privacy implications of uploading voice data to external servers, including specific guidance on data retention settings.
[INDIRECT_PROMPT_INJECTION]: The skill processes transcription results derived from external audio or video sources through an LLM for text cleaning and formatting. This presents a potential surface for indirect prompt injection, as malicious instructions spoken in the source media could theoretically attempt to manipulate the LLM's cleaning behavior.
Ingestion points: Transcription text received from the Groq Whisper API, which is originally sourced from potentially untrusted user-provided audio/video files.
Boundary markers: The clean.py script uses a system prompt to define cleaning rules but does not employ specific delimiters or boundary markers to isolate the untrusted transcription text within the LLM request.
Capability inventory: The skill has capabilities to read and write local files, extract media metadata, and execute shell commands.
Sanitization: There is no automated sanitization or filtering of the transcription content before it is processed by the cleaning LLM.
[DYNAMIC_EXECUTION]: The skill relies on the creation and execution of local Python scripts (make_srt.py and clean.py) to manage SRT file generation and chunked API calls. These scripts are executed at runtime using uv run python.

opencode-groq-api