The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads a Whisper GGML model file from a third-party mirror of Hugging Face if the file is not already present in the user's cache.
Evidence: In scripts/bili_transcript.py, the constant HF_MIRROR_BASE points to hf-mirror.com, and the ensure_whisper_model function performs the download using urllib.request.urlopen.
[COMMAND_EXECUTION]: The script executes several external command-line tools to process audio files and perform transcription.
Evidence: scripts/bili_transcript.py calls subprocess.run() for ffmpeg, afconvert (on macOS), whisper-cli, and whisper. These executions use argument lists rather than raw shell strings, which is a secure practice to prevent shell command injection.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data (subtitles and transcribed text) sourced from external Bilibili videos, which could potentially contain malicious instructions aimed at the agent.
Ingestion points: The functions fetch_subtitle_body and whisper_fallback in scripts/bili_transcript.py ingest content from the Bilibili API and transcription tools respectively.
Boundary markers: Not present in the raw data, but the SKILL.md instructions explicitly tell the agent to deliver the text to the user without altering the original meaning.
Capability inventory: The script maintains the ability to write to the local filesystem, perform network operations, and execute system commands.
Sanitization: The video identifier (BVID) is strictly validated using the regular expression BV[0-9A-Za-z]{10} before any processing occurs.

bilibili-transcript