The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from untrusted external sources, creating an attack surface for indirect prompt injection.
Ingestion points: The skill uses markitdown to extract text from PDF and DOCX files and mentions using webfetch for URLs in references/long-reads.md.
Boundary markers: There are no explicit delimiters or instructions provided to the agent to ignore potentially malicious content within the ingested text.
Capability inventory: The skill utilizes subprocess calls for ffmpeg and ffprobe and performs file system operations (read/write) via Python scripts in references/long-reads.md and references/podcast.md.
Sanitization: While the skill includes a clean_for_narration helper, it is designed for editorial formatting (expanding abbreviations) rather than security sanitization.
[EXTERNAL_DOWNLOADS]: The skill instructions involve downloading and installing multiple third-party packages and cloning repositories from GitHub.
Evidence: Mentions git clone for repositories like FunAudioLLM/CosyVoice, fishaudio/fish-speech, index-tts/index-tts, myshell-ai/MeloTTS, and neuphonic/neutts-air across various backend documentation files.
Context: These are established open-source projects in the text-to-speech community, and the downloads are consistent with the skill's purpose.
[COMMAND_EXECUTION]: The skill frequently uses shell commands for audio post-processing, which is standard for audio tasks but represents an active execution surface.
Evidence: Extensive use of ffmpeg and ffprobe CLI tools for concatenation, normalization, and format conversion throughout the reference guides, such as in references/ffmpeg-audio.md.

audio-generation