The Agent Skills Directory

[COMMAND_EXECUTION]: The skill requires the agent to run shell commands, notably yt-dlp for fetching YouTube video information and subtitles, and a provided Python script limpar_legendas.py to process the text.
[EXTERNAL_DOWNLOADS]: The agent uses yt-dlp to retrieve subtitles and metadata from YouTube. This is a functional requirement and targets a well-known service.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests and processes content from external YouTube subtitles.
Ingestion points: The agent reads subtitle files downloaded from YouTube (referenced in SKILL.md and handled by scripts/limpar_legendas.py).
Boundary markers: The instructions provide structural guidance for the output but lack explicit security delimiters or instructions to ignore embedded commands within the untrusted subtitle content.
Capability inventory: The agent has file system access for creating directories and writing articles, and it can execute shell commands like yt-dlp and python3 (SKILL.md).
Sanitization: The cleaning script scripts/limpar_legendas.py removes ASR technical tags and artifacts but does not specifically filter for malicious instructional content or prompt injection patterns.

teo-transcricao