The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill exposes a surface for indirect prompt injection by processing external inputs through various transformer models.
Ingestion points: The skill processes text, images, and audio through pipeline() calls in SKILL.md and references/pipelines.md, as well as dataset mapping in references/training.md.
Boundary markers: The instructions do not specify any delimiters or warnings to the model to ignore instructions embedded in the processed data.
Capability inventory: The skill utilizes subprocesses for package installation (uv pip install in SKILL.md), writes to the local filesystem (save_pretrained in references/models.md), and performs network operations (push_to_hub in references/models.md).
Sanitization: There is no evidence of input sanitization or filtering before passing data to the models.
[COMMAND_EXECUTION]: The skill provides instructions for installing necessary Python dependencies using the uv package manager in SKILL.md. These commands are standard for setting up the environment required for the Transformers library and its related machine learning tasks.
[EXTERNAL_DOWNLOADS]: The skill fetches pre-trained models, datasets, and evaluation scripts from the Hugging Face Hub (huggingface.co). These downloads are performed using the library's official APIs and target a well-known service in the machine learning community.

transformers