The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/main.py uses the os.system() function to call ffmpeg. The command string is constructed using f-strings that include the wav path and the vo['audio'] path, the latter of which is sourced from the user-provided brief.json. This pattern is vulnerable to shell command injection if the input file contains malicious metacharacters.
[EXTERNAL_DOWNLOADS]: The skill performs several external network operations to download assets and data. It connects to api.openai.com for text-to-speech, transcription, and image generation. Additionally, it fetches SVG icons from cdn.jsdelivr.net and retrieves an icon registry from raw.githubusercontent.com. These destinations are recognized well-known technology services and official repositories.
[INDIRECT_PROMPT_INJECTION]: The skill ingests data from local configuration files (brief.json, assets.json) and scripts (vo.txt). This data is used as prompts for AI image generation and is rendered directly as text into the video output. There are no security boundaries to prevent malicious instructions within this data from influencing the agent's behavior.
Ingestion points: brief.json, assets.json, and vo.txt in the example and workspace directories.
Boundary markers: None identified in the prompt construction logic.
Capability inventory: Subprocess calls to ffmpeg, ffprobe, chromium-browser, and convert (ImageMagick).
Sanitization: No sanitization or escaping is performed on user-controlled strings before their use in shell commands or prompt interpolation.
[DYNAMIC_EXECUTION]: The skill implements a dynamic video rendering engine. It also uses a headless Chromium process to rasterize SVG assets via subprocess.run(), which includes the use of the --no-sandbox flag and a temporary staging directory in the user's home folder.

illustrated-explainer