The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/media_to_transcript.py executes several system binaries including ffmpeg, ffprobe, and curl via the subprocess module. These tools are used for media conversion, metadata probing, and as a network transport fallback. The commands are constructed using list-based arguments which mitigate common shell injection risks, but they provide the skill with significant system-level capabilities.
[INDIRECT_PROMPT_INJECTION]: The skill ingests metadata from external media URLs, including titles and descriptions provided by platforms like YouTube or Bilibili. This untrusted content is incorporated into both the ASR context and the final AI correction prompt without sanitization to prevent adversarial instructions from influencing the AI's output.
Ingestion points: External media metadata processed in scripts/media_to_transcript.py via the probe_and_download function.
Boundary markers: None identified; untrusted metadata is concatenated directly into prompt templates.
Capability inventory: The skill has the ability to read local files, perform network operations, and execute subprocesses.
Sanitization: Content is truncated for length but not filtered for potential prompt injection payloads.
[DATA_EXPOSURE]: The pipeline automatically reads and loads configuration from environment variables and various .env files, including those located in the user's home directory (e.g., ~/.skills/.env). While this is used for legitimate credential management for Volcengine and R2 services, it demonstrates a pattern of accessing sensitive configuration files outside the immediate project scope.
[DYNAMIC_EXECUTION]: The script utilizes importlib to dynamically load and execute logic from a separate local skill named video-transcript. This architecture introduces a dependency on the integrity of the external skill's code at runtime.

media-to-transcript