The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill installs verified machine learning dependencies, including torch, transformers, datasets, evaluate, accelerate, timm, pillow, librosa, and soundfile from official registries using the uv package manager.
[CREDENTIALS_UNSAFE]: Documentation correctly handles authentication by advising the use of environment variables (HF_TOKEN) or interactive login prompts (huggingface_hub.login()), which avoids the risk of hardcoded credentials.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-provided text, images, and audio through transformer models, which is an inherent behavior for machine learning tasks.
Ingestion points: Untrusted inputs provided to the pipeline() API and model generate() methods across all documented scripts.
Boundary markers: No explicit delimiters or instructions to ignore embedded commands are included in the provided code snippets.
Capability inventory: The skill is configured with Read, Write, Edit, and Bash permissions to facilitate local model execution and file management.
Sanitization: No explicit input filtering or output validation logic is present in the inference examples.

alterlab-transformers