The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses ffmpeg for video frame extraction and video concatenation. It also executes a local shell script (extract-frames.sh) from a related skill to perform the extraction process.
[EXTERNAL_DOWNLOADS]: The skill instructs the user to install well-known dependencies if missing, including ffmpeg via Homebrew and openai-whisper via pip.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data (video, audio, and product descriptions) which serves as an injection surface.
Ingestion points: Source video files and user-provided product descriptions are used to generate prompts and dialogue scripts.
Boundary markers: The skill does not use explicit boundary markers to isolate transcribed audio or analyzed visual content, although it follows a structured prompt template.
Capability inventory: The agent can execute shell commands (bash, ffmpeg) and make network requests to the Arcads API.
Sanitization: There is logic to filter 'forbidden words' (e.g., 'cinematic', '8k') from the final generation prompt to avoid API rejections, but no explicit sanitization for instructions embedded in the input media. This is mitigated by mandatory human review gates (Steps 4 and 7).

clone-ad