The Agent Skills Directory

[SAFE]: The skill provides documentation and examples for using the physicalai-train library, a project by the vendor open-edge-platform. The instructions are standard for machine learning workflows, involving model training, validation, and testing.
[EXTERNAL_DOWNLOADS]: The skill mentions downloading datasets using the LeRobotDataModule and repo_id (e.g., lerobot/pusht). This is a standard operation in robot learning frameworks for fetching training data from public repositories like Hugging Face. These are legitimate resources within the stated domain of the skill.
[INDIRECT_PROMPT_INJECTION]: The skill has a standard vulnerability surface as it processes external configuration files (YAML) and datasets. However, there are no specific instructions that attempt to bypass safety filters or exploit the agent's behavior. The use of structured YAML configs with jsonargparse suggests standard configuration management practices.
[COMMAND_EXECUTION]: The skill provides example CLI commands using the physicalai tool. These commands are used for training (fit), validation, and testing, which is the intended and transparent purpose of the skill.

pai-train-policy