The Agent Skills Directory

[COMMAND_EXECUTION]: The skill facilitates the invocation of local command-line tools for media analysis and processing.
Evidence: Instructs the agent to use ffmpeg for read-only inspection of source file metadata.
Evidence: References the use of whisper.cpp, mlx_whisper, and faster-whisper for local audio transcription.
[INDIRECT_PROMPT_INJECTION]: The skill processes transcription data which serves as a potential vector for indirect prompt injection from audio content.
Ingestion points: The skill ingests transcripts generated by the ChatCut plugin or local ASR models (e.g., Whisper) as described in references/ChatCut实操.md.
Boundary markers: While the skill includes strict behavioral constraints like "只删和重排，不加词" (only delete and reorder, no adding words), it lacks explicit prompt delimiters to segregate transcription data from system instructions.
Capability inventory: The skill possesses capabilities to modify project timelines (apply_script), interact with the ChatCut plugin API, and execute shell-based media tools.
Sanitization: No specific input sanitization or validation logic is defined to filter malicious prompts embedded within processed transcripts.
[EXTERNAL_DOWNLOADS]: The skill references external resources for supplementary functionality.
Evidence: Fetches face detection models (YuNet) from media.githubusercontent.com/media/... to facilitate privacy-preserving blurring of bystanders. This source is a standard distribution platform for GitHub-hosted assets and is handled neutrally.
[DYNAMIC_EXECUTION]: The skill supports a fallback execution path involving local runtime execution of media models.
Evidence: Provides instructions for a "本机转写旁路" (local transcription bypass) which involves executing Whisper-based models on the host machine when primary cloud-based plugins are unavailable.

zmm-cut