The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the agent to install the watch-skill package via pip or uv. It also uses a 'doctor' command to bootstrap binary dependencies such as ffmpeg and yt-dlp from the internet.
[INDIRECT_PROMPT_INJECTION]: This skill provides a surface for indirect prompt injection by extracting and presenting untrusted data from external videos (transcripts, captions, and OCR text) to the agent.
Ingestion points: Video transcripts (via Whisper or cloud STT), OCR text extracted from video frames, and video metadata processed by the watch-skill engine.
Boundary markers: No specific delimiters or warnings to ignore embedded instructions are specified when the extracted text is passed to the agent.
Capability inventory: The skill uses Bash to execute the CLI tool and Read to access local frame files.
Sanitization: The skill relies on the underlying engine's processing and the agent's internal safety filters, with no explicit sanitization of the extracted text described in the instructions.

watch