The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/media_optimizer.py script utilizes subprocess.run to call ffmpeg and ffprobe for media file analysis and optimization. These calls are implemented using list-based arguments without shell=True, which is a secure practice to prevent shell injection.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external files (PDFs, images, audio, video) that are subsequently passed to the Gemini API for analysis or generation.
Ingestion points: Files are ingested through CLI arguments in scripts/gemini_batch_process.py, scripts/document_converter.py, and scripts/media_optimizer.py.
Boundary markers: The scripts use instructional prompts (e.g., in document_converter.py) to request specific output formats like clean Markdown, which provides a weak boundary against malicious content embedded in the processed files.
Capability inventory: The skill has the capability to write files to the local disk and execute subprocesses (ffmpeg).
Sanitization: There is no explicit sanitization or filtering of the content within the media or document files before they are sent to the AI model.
[DYNAMIC_EXECUTION]: The skill uses sys.path.insert to dynamically load utility modules (resolve_env.py, api_key_rotator.py) from standard locations associated with the author's environment. While this involves dynamic path modification, it aligns with the vendor's established development patterns for resource sharing across skills.
[SAFE]: The skill relies on official libraries (e.g., google-genai) and standard industry tools (e.g., ffmpeg, Pillow). No malicious obfuscation, persistence mechanisms, or credential exfiltration patterns were detected.

ai-multimodal