The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads external binaries and machine learning models required for audio transcription.
Evidence: transcribe-captions.md uses installWhisperCpp to download the Whisper.cpp engine and downloadWhisperModel to fetch pre-trained models (e.g., medium.en). These operations are handled by the official @remotion/install-whisper-cpp library.
[REMOTE_CODE_EXECUTION]: The skill facilitates the execution of downloaded binaries to perform heavy computation tasks.
Evidence: transcribe-captions.md executes the transcribe function, which runs the locally installed Whisper.cpp binary against audio input files.
[COMMAND_EXECUTION]: The documentation instructs users to execute shell commands to manage the development environment.
Evidence: Multiple files (display-captions.md, import-srt-captions.md, transcribe-captions.md) suggest using npx remotion add or similar package manager commands to install necessary dependencies.
[PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection by ingesting and processing external data files.
Ingestion points: The skill reads external .srt files (import-srt-captions.md) and .json files (display-captions.md) via the fetch API.
Boundary markers: No explicit delimiters or warnings for the agent to ignore instructions embedded within the caption text are present.
Capability inventory: The skill is primarily focused on UI rendering and transcription; it does not currently expose dangerous capabilities (like eval or shell access) that could be triggered by injected data.
Sanitization: The skill performs parsing (e.g., parseSrt) but does not explicitly describe content sanitization for LLM safety.

remotion-captions