The Agent Skills Directory

[COMMAND_EXECUTION]: The skill interacts with system utilities and CLI tools including ffmpeg, ffprobe, claude, and gh to process audio files and generate summaries. These interactions are implemented using Python's subprocess.run with argument lists, avoiding shell interpretation and mitigating command injection risks.
[UNVERIFIABLE_DEPENDENCIES]: The skill manages its own dependencies by offering to install several Python packages (faster-whisper, openai-whisper, tqdm, rich) and system tools via Homebrew. These packages are well-known and fetched from official registries (PyPI, Homebrew). The skill also downloads pre-trained transcription models from established sources like OpenAI and Hugging Face.
[INDIRECT_PROMPT_INJECTION]: By transcribing user-provided audio files and passing the resulting text to an LLM for processing, the skill maintains an inherent exposure to indirect prompt injection. Malicious audio content could potentially contain spoken instructions designed to override the summarization logic. The skill uses structured templates and separators to reduce the risk of accidental instruction following.
Ingestion points: Transcribed audio text is interpolated into LLM prompts within scripts/transcribe.py.
Boundary markers: The skill uses --- delimiters and explicit headers (e.g., Transcrição:) to separate the data from the processing instructions.
Capability inventory: The skill reads local audio files, writes Markdown reports, and invokes external LLM CLI tools to generate content.
Sanitization: No specific sanitization or filtering of transcribed text is performed prior to its inclusion in the LLM prompt.
[DYNAMIC_EXECUTION]: The scripts/transcribe.py script features a self-healing mechanism that attempts to install missing UI libraries (rich, tqdm) at runtime using pip. Additionally, the skill dynamically generates and executes prompts for LLM summarization based on the detected transcript type, which is a core feature for its intelligent processing capabilities.

audio-transcriber