The Agent Skills Directory

[COMMAND_EXECUTION]: The skill documents usage of the z-ai command-line interface for performing audio transcription tasks, including file-based and base64-encoded processing.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted audio data from external files which could contain malicious spoken instructions designed to influence the behavior of an agent processing the transcribed text.
Ingestion points: Files are read using fs.readFileSync in the provided scripts/asr.ts and various examples in SKILL.md (e.g., transcribeAudio, transcribeDirectory).
Boundary markers: None identified in the prompt templates or code examples.
Capability inventory: The skill uses z-ai-web-dev-sdk for network-based ASR processing and fs.writeFileSync for saving results to the local filesystem.
Sanitization: No specific sanitization or filtering of transcribed text is implemented before usage or storage.

ASR