The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes transcription text (generated from user-supplied audio or external ASR) to create semantic editing plans. This creates a surface for indirect prompt injection if the audio content contains malicious instructions designed to influence the agent's highlight selection or cleanup logic.
Ingestion points: The agent reads transcription data from 语义审阅稿.md (generated in scripts/runtime/cli/newcut.mjs).
Boundary markers: The skill uses a structured U-indexing system (e.g., [U0001]) to reference transcript segments, which provides a degree of separation between instructions and data.
Capability inventory: The skill uses execFile to run ffmpeg, ffprobe, and whisper, and writes files using writeFile (as seen in scripts/runtime/cli/render-semantic-plan.mjs).
Sanitization: The skill does not explicitly sanitize the text content of the transcript, but relies on the LLM to perform semantic selection within the provided U-index constraints.
[COMMAND_EXECUTION]: The skill heavily utilizes shell commands via execFile, exec, and spawnSync to perform video editing tasks. It uses ffmpeg for audio extraction and video rendering, and ffprobe for media metadata inspection. A custom shellQuote utility is implemented in scripts/runtime/packages/newcut-core/src/asr/providers/doubao.mjs to mitigate command injection risks when using user-defined command templates for the Doubao ASR provider.
[EXTERNAL_DOWNLOADS]: The skill recommends installing openai-whisper via pip and requires ffmpeg. During runtime, it may fetch content from ByteDance's Doubao ASR API (openspeech.bytedance.com) as part of its core transcription functionality. These are well-known services and tools appropriate for the skill's purpose.

newcut-video-clipping