The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of the Bash tool to orchestrate complex workflows, including building Docker images, managing containers, and executing Python training scripts. This behavior is fundamental to the skill's purpose of local model fine-tuning.
[EXTERNAL_DOWNLOADS]: The skill fetches large assets, including container images from NVIDIA's NGC registry (nvcr.io) and datasets/models from the Hugging Face Hub. These operations target well-known, reputable services essential for the machine learning lifecycle.
[DYNAMIC_EXECUTION]: The skill dynamically generates a complete Python project (scripts, configuration, and Dockerfiles) at runtime based on the user's task and live documentation research. While this involves executing generated code, it is the primary functional delivery of the skill.
[INDIRECT_PROMPT_INJECTION]: The skill implements a 'research' phase that fetches content from external sources like model cards on Hugging Face and example scripts on GitHub. This data is used to inform code generation, creating a surface for indirect influence, although the skill includes prioritization and conflict-resolution rules to manage this risk.
[DATA_EXFILTRATION]: By default, the skill is configured to push trained model checkpoints, evaluation metrics, and reports to the Hugging Face Hub. This represents an outbound transfer of user-generated assets to an external platform, which is the intended delivery mechanism for the fine-tuning workflow.

tao-finetune-huggingface-model