The Agent Skills Directory

[COMMAND_EXECUTION]: The skill extensively uses Bash and FFmpeg to process video files. Commands include file search (find), metadata extraction (ffprobe), and complex video editing loops. While standard for video editing, these are powerful tools that interact with the local filesystem.
[EXTERNAL_DOWNLOADS]: The skill utilizes yt-dlp to download subtitles/transcripts from remote URLs and npm to install Remotion and its dependencies. It also calls the ElevenLabs API for voiceover generation and references fal.ai for media generation.
[CREDENTIALS_UNSAFE]: The skill requires an ELEVENLABS_API_KEY for Phase 5. However, it correctly follows security best practices by instructing the user to set this as an environment variable (export ELEVENLABS_API_KEY=your_key) rather than hardcoding it.
[PROMPT_INJECTION]: Phase 2 (AI STRUCTURE) involves reading transcript.txt and using an LLM to generate an Edit Decision List (cuts.txt). This represents an indirect prompt injection surface where malicious content inside a video transcript could attempt to influence the agent's editing decisions.
[INDIRECT_PROMPT_INJECTION]: The skill reads untrusted external data (video transcripts via yt-dlp and Whisper).
Ingestion points: transcript.txt and VTT files processed in Phase 2.
Boundary markers: None explicitly mentioned in the instructions for the EDL generation step.
Capability inventory: Significant capabilities including Bash, FFmpeg (subprocess calls), and filesystem writes across all phases.
Sanitization: The skill includes a vtt_to_paragraph.py script to clean transcripts, which provides a layer of sanitization for the text data before LLM processing.

video-editing