The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied CSV files where content is interpolated into generated Python scripts and TTS narration scripts, creating a surface for indirect prompt injection.
Ingestion points: The scripts/parse_csv.py script reads external CSV data from the input/ directory.
Boundary markers: No explicit delimiters or instructions are used to prevent the LLM or subsequent processing steps from interpreting data as instructions.
Capability inventory: The skill has the ability to execute shell commands (ffmpeg, manim), generate and run Python scripts, and perform network operations.
Sanitization: Basic whitespace cleaning and rudimentary LaTeX escaping are applied, but these do not fully sanitize against logical injection in the code generation phase.
[DYNAMIC_EXECUTION]: The skill uses a code generation pattern where it dynamically creates a Manim animation script (script.py) containing data from the CSV, which is then executed by the system.
Evidence: scripts/generate_scenes.py builds the script.py file using string concatenation and interpolation of CSV-derived data.
[COMMAND_EXECUTION]: The skill executes shell commands with shell=True, which is a recognized security risk if command arguments are not strictly controlled.
Evidence: scripts/add_audio.py uses subprocess.run(cmd, shell=True, ...) to invoke ffmpeg and edge-tts. While most arguments are derived from configuration, the pattern is susceptible to command injection if configuration values or filenames are manipulated.
[EXTERNAL_DOWNLOADS]: The skill interacts with external API services and CLI tools to generate media content.
Evidence: scripts/add_audio.py makes network requests to api.sarvam.ai for premium TTS and utilizes the edge-tts package for fallback audio generation.

csv-landscape-video-pipeline