The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external speech audio, transcriptions, and web search results, which creates a surface where maliciously crafted data could attempt to influence the agent's logic.\n
Ingestion points: Audio data processed via process_audio_stream and web content retrieved via mcp__firecrawl__firecrawl_search and WebFetch in references/mellifluo-platform.md and SKILL.md.\n
Boundary markers: The instructions do not define specific delimiters or instructions to ignore embedded commands in the processed audio/text data.\n
Capability inventory: The skill has access to bash execution, file editing, and network search tools across all scripts.\n
Sanitization: No specific sanitization or filtering of speech-to-text transcriptions is implemented before processing.\n- [DYNAMIC_EXECUTION]: The skill uses torch.load to load neural network weights, which relies on the pickle serialization format.\n
Evidence: torch.load(model_path) is called in RealTimePERCEPTR and RealTimePhonemeRecognizer within /references/ai-models.md to load model weights.\n
Context: This is the standard mechanism for loading weights in the PyTorch ecosystem and is tied to the skill's primary function of speech analysis.\n- [EXTERNAL_DOWNLOADS]: The skill downloads pre-trained AI models from external repositories at runtime.\n
Evidence: Implementation in /references/ai-models.md fetches models from HuggingFace, specifically facebook/wav2vec2-xls-r-300m and vitouphy/wav2vec2-xls-r-300m-timit-phoneme.\n
Context: These models are hosted on HuggingFace and use well-known provider organizations.

speech-pathology-ai