The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses ffmpeg and ffprobe to perform scene detection, frame extraction, audio extraction, and silence detection. These commands are executed using subprocess.run with list-based arguments, which prevents common command injection vulnerabilities. Numeric parameters like timestamps and thresholds are validated or cast to appropriate types before use.
[EXTERNAL_DOWNLOADS]: The skill makes network requests to the MiMo AI API (via api.xiaomimimo.com and related subdomains) to perform ASR (Speech-to-Text) and VLM (Vision-Language Model) analysis. These requests are necessary for the skill's stated purpose of video understanding.
[CREDENTIALS_SAFE]: The skill manages API keys through environment variables. Notably, it includes a sanitization function (_sanitize_api_error in scripts/lib.py) that uses regular expressions to redact API keys and data URLs from error messages before logging them, which is a strong security best practice.
[INDIRECT_PROMPT_INJECTION]: The skill has an indirect prompt injection surface as it processes untrusted video data (transcripts and visual descriptions) and user-supplied background research, which are interpolated into prompts for subsequent LLM stages.
Ingestion points: Untrusted data enters the agent context via asr_result.json, vlm_analysis.json, and background_research.json.
Boundary markers: The skill uses structured prompts with clear section headers and identifiers (e.g., [visual:...], [asr:...]) to help the LLM distinguish between instructions and data.
Capability inventory: The skill can perform network requests, write to the local work directory, and execute shell commands via ffmpeg.
Sanitization: The skill performs text normalization and clipping, but primarily relies on the downstream LLM's inherent safety guardrails to handle potentially malicious content extracted from the video.

video-understanding