The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill accesses a sensitive authentication token located at ~/.openkbs/projectJWT to authorize requests sent to the vendor's transcription proxy.
[COMMAND_EXECUTION]: The script uses execFileSync to run the ffmpeg and ffprobe binaries for converting media files and extracting duration metadata.
[PROMPT_INJECTION]: The VOCAB environment variable is interpolated into the transcription prompt without sanitization. While this is intended to provide a list of terms to the model, it could theoretically be used to inject instructions if the environment variable is populated from an untrusted source.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted audio/video content which is subsequently processed by a language model. It attempts to mitigate accidental instruction following by including specific formatting instructions in the transcriptionPrompt (e.g., "Output only the transcript, nothing else").
Ingestion points: Media files provided via the inputPath argument in transcribe.mjs.
Boundary markers: The script includes explicit instructions to the model to only output the transcript and maintain specific formatting, acting as a boundary for the generated text.
Capability inventory: The skill possesses capabilities for subprocess execution (ffmpeg), network communication (fetch to Google and vendor proxy), and file system writing (fs.writeFileSync for transcripts).
Sanitization: The input file is converted to a 16 kHz mono MP3 using ffmpeg before processing, which serves to normalize the audio stream and remove non-audio metadata.

file-transcribe