The Agent Skills Directory

[COMMAND_EXECUTION]: Orchestrates video processing by extensively invoking shell commands via Python's subprocess module and direct Bash scripts. It executes tools including ffmpeg, ffprobe, and node to manipulate media files, extract frames, and encode video segments.
[DYNAMIC_EXECUTION]: Uses __import__ in scripts/build_gfx.py and scripts/sheet.py to dynamically load and execute functions from local Python files based on generated tags. While this is part of the parallel subagent architecture, it involves executing code generated or discovered at runtime.
[EXTERNAL_DOWNLOADS]: scripts/setup.sh provisions the environment by downloading multiple dependencies through uv pip (including torch, mediapipe, and faster-whisper) and utilizes npx to fetch the hyperframes rendering engine. These originate from established registries and trusted vendors.
[INDIRECT_PROMPT_INJECTION]: The skill ingests raw audio data from user videos which is processed via speech-to-text. This untrusted data influences planning, graphics generation, and text overlays.
Ingestion points: transcript_full.json created in scripts/transcribe.py from user-provided video files.
Boundary markers: None. Instructions in the video audio are not isolated from the agent's processing logic.
Capability inventory: Extensive subprocess execution, file system writes, and package installation across all pipeline scripts (analyze.sh, finish.sh, compose.py).
Sanitization: Minimal; only punctuation removal and case-matching are performed during token alignment.
[REMOTE_CODE_EXECUTION]: Automated scanners identified a fetch pattern in scripts/eyesheet.py involving subprocess. Static analysis confirms this is a local utility function for extracting video frames via FFmpeg and does not involve remote network-based execution.

create-video-shorts