The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill configuration supports a prompts_module parameter which takes a dotted Python import path. This mechanism is designed for "domain adaptation," allowing the agent to load custom Python modules for specialized video annotation tasks. This creates a functional capability for dynamic loading and execution of code based on user-supplied paths, which could be exploited if an attacker manages to place a malicious module on the environment's PYTHONPATH in order to execute arbitrary code during the distillation process.\n- [INDIRECT_PROMPT_INJECTION]: The pipeline processes untrusted external video data and user-provided JSONL inputs to generate training datasets. This ingestion surface is susceptible to indirect prompt injection, where visual or embedded data in the videos could attempt to influence the distillation logic or the resulting reasoning traces.\n
Ingestion points: Video files located in video_root and entries in input_jsonl_files (referenced in SKILL.md and configuration.md).\n
Boundary markers: The system utilizes structured prompt templates to encapsulate data, though explicit sanitization of video-derived content before interpolation into downstream LLM prompts is not detailed.\n
Capability inventory: The skill possesses Write capabilities for result generation and Bash execution for running the auto_label tool.\n
Sanitization: Content filtering is implicitly performed by the VLM/LLM distillation process, but no pre-processing sanitization of the video data itself is noted in the provided documentation.

tao-generate-video-reasoning-annotations