The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill's transcription script (scripts/transcribe.py) utilizes the funasr library, which automatically downloads required pre-trained models (paraformer-zh, fsmn-vad, and ct-punc) from ModelScope. ModelScope is an established and well-known model repository provided by Alibaba. This download behavior is standard for the skill's primary function.\n- [PROMPT_INJECTION]: The skill implements a multi-agent workflow to reconcile and proofread transcription results, which involves processing external text data derived from ASR and OCR sources. This creates a surface for indirect prompt injection if the source audio, video, or visual material contains adversarial instructions.\n
Ingestion points: scripts/transcribe.py (ASR output) and references/ground-truth.md (text extracted from images via OCR).\n
Boundary markers: The reconciliation prompts in references/reconciliation.md include specific behavioral constraints and instructions to be "extremely conservative" and avoid rewriting speaking content.\n
Capability inventory: The Python transcription script writes results to the local filesystem (funasr_result.json), and the skill instructions suggest the use of shell commands for ffmpeg and whisper-cpp.\n
Sanitization: No programmatic sanitization or filtering of transcribed text is performed before it is passed to the agent for reconciliation.

accurate-transcription