The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script templates/split_55_45_compositor.py reads an external file render2.py and executes a portion of its content using exec(compile(ASSETS, "render2_assets", "exec")). This enables runtime code assembly which can be exploited if the source file is modified.
[COMMAND_EXECUTION]: Multiple scripts execute system-level commands using subprocess.run. scripts/fetch_giphy.py calls ffmpeg to process GIF frames. scripts/prep_talking_head.py executes ffprobe, ffmpeg, and whisper to analyze and process user-supplied video and audio clips.
[EXTERNAL_DOWNLOADS]: The skill fetches content from several external sources. scripts/fetch_figure.py retrieves portrait data and metadata from Wikipedia and Wikidata. scripts/fetch_giphy.py downloads media from Giphy's API. SKILL.md also describes the use of Apify for scraping Instagram reels.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data (Instagram captions, transcribed audio) and uses it to drive animation logic and on-screen text. While it includes a 'HARD GUARD' to filter specific identity names, the ingestion of external narration into the rendering pipeline creates an indirect injection surface.
Ingestion points: scripts/prep_talking_head.py (Whisper transcription of user clips), SKILL.md (Apify Instagram scraper for captions).
Boundary markers: The skill uses a 'HARD GUARD' regex to refuse processing if specific names are detected in the script.
Capability inventory: File system writes (frames, meta.json), network access (Giphy, Wikipedia), and command execution (FFmpeg, Whisper).
Sanitization: Basic string cleaning and keyword matching are performed on Whisper outputs.

split-animated-short