The Agent Skills Directory

[COMMAND_EXECUTION]: The orchestrator script build_video.py invokes manim and ffmpeg using subprocess.run with list-formatted arguments. This prevents shell injection by ensuring that user-provided inputs, such as video titles or CSV file paths, are treated as literal arguments rather than executable shell code.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from user-provided CSV files to generate video content and narration. While the execution flow is secure, the content itself is not sanitized as it is intended for literal rendering.
Ingestion points: parse_csv.py reads question and explanation text from the CSV input.
Boundary markers: No explicit delimiters are used to isolate CSV text from the agent's context during processing.
Capability inventory: File system access, network requests to Sarvam AI, and shell command execution.
Sanitization: Content is not sanitized; however, the impact is minimized by the skill's primary focus on media generation. A risk factor exists if the AI agent interprets instructions embedded in the CSV content as valid commands during the review phase.
[DYNAMIC_EXECUTION]: The skill uses Manim to render animations based on a JSON configuration file loaded at runtime by quiz_scene.py. This configuration is passed securely via an environment variable to a child Python process. This is a standard and necessary pattern for the Manim engine and is implemented without exposing the system to external code injection.
[DATA_EXFILTRATION]: The skill transmits text content to api.sarvam.ai for TTS synthesis. This is a documented functional requirement using a well-known service. The SARVAM_API_KEY is securely fetched from the environment or a local .env file, adhering to secret management best practices.

csv-quiz-video