The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/volc_asr_stream.py invokes ffmpeg via subprocess.run to transcode various media formats into PCM audio. While the command uses a list-based argument structure to mitigate shell injection, it processes user-provided file paths. The skill also suggests using ffmpeg and uvx commands in SKILL.md for chunking and running scripts.
[EXTERNAL_DOWNLOADS]: The skill facilitates network connections to external API providers, specifically OpenAI (api.openai.com) and Volcano ASR (bytedance.com). It also recommends fetching Python packages such as httpx, socksio, and openai-whisper at runtime using uvx. These services and packages are well-known or trusted, aligned with the skill's primary purpose.
[INDIRECT_PROMPT_INJECTION]: The skill implements an 'AI polish pass' where the agent (Claude) reads and edits the full SRT transcript to correct homophones. This presents an indirect prompt injection surface where malicious instructions embedded in the audio source could be transcribed and subsequently interpreted by the agent during the polishing phase.
Ingestion points: Data enters the context from API-generated JSON results which are processed into SRT files and then read by the agent in-session.
Boundary markers: The instructions do not specify explicit delimiters or 'ignore' instructions for the polishing prompt.
Capability inventory: The agent has the ability to write files (build_srt_from_asr.py) and execute shell commands via subprocess (volc_asr_stream.py).
Sanitization: There is no evidence of sanitization or filtering of the transcript text before the agent processes it.

wjs-transcribing-audio