The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the Bash tool to execute docker run commands for machine learning tasks. It utilizes GPU resources via the --gpus all flag and bind-mounts a host workspace directory to the container to facilitate data processing between the host and the containerized environment.
[EXTERNAL_DOWNLOADS]: The workflow involves pulling a Docker container image from NVIDIA's official registry (nvcr.io) and may fetch pre-trained model weights from Hugging Face repositories (e.g., google/siglip-base-patch16-224). These sources are recognized as well-known and trusted service providers.
[DYNAMIC_EXECUTION]: Several automated validation hooks, such as hooks/mining-artifacts-check.sh and hooks/mining-script-check.sh, dynamically execute Python scripts via the shell. These scripts are used to perform schema verification, check for row counts, and detect failure modes in the workflow's output.
[INDIRECT_PROMPT_INJECTION]: The skill ingests external data from user-supplied Parquet files containing target images and source pools. It extracts metadata—such as row counts, column schemas, and label values—to generate a structured Mining_Report.md. This creates an ingestion surface where external data content is processed and reflected back into the agent's execution context.
[DATA_EXFILTRATION]: A packaging hook in hooks/mining-package.sh accesses session transcript logs, including those located in the agent's project metadata directory at $HOME/.claude/projects/. These logs are copied into a timestamped results folder within the user's local workspace to provide a complete audit trail for the mining iteration.

tao-mine-aoi-images