The Agent Skills Directory

[COMMAND_EXECUTION]: Several Python scripts in the scripts/ directory (mapa_movimento.py, analisar_video.py, gravar.py) use the subprocess module to run system utilities such as ffmpeg and ffprobe for video processing and metadata extraction.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted video files provided by the user, which are then analyzed by a multimodal model (Gemini via OpenRouter). The model's description of the movement is used to inform the agent's code generation process, creating a vulnerability where a crafted video could attempt to inject instructions into the generated HTML/JS.
Ingestion points: User-provided video files (e.g., video-ref.mp4) are processed by analisar_video.py to generate analysis in gemini.md.
Boundary markers: None identified in the prompt templates or scripts to isolate the model-generated video description from functional instructions.
Capability inventory: The skill utilizes subprocess calls to ffmpeg, uv, and ffprobe, writes to the local file system (index.html, app.js), and performs network operations to the OpenRouter API.
Sanitization: No explicit sanitization or filtering of the multimodal model's output is performed before it is used for code generation.
[EXTERNAL_DOWNLOADS]: The documentation indicates the use of the rembg tool, which automatically downloads large AI model files (~1GB) from public repositories during initial execution.
[DYNAMIC_EXECUTION]: The gravar.py script uses Playwright's page.evaluate() function to execute JavaScript code within the browser context. While used for legitimate state management during video recording, this represents a dynamic execution pattern.

video2html