The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads model weights, configurations, and LoRA adapters from Hugging Face's official Model Hub. These downloads target well-known organizations including Stability AI, RunwayML, and Hugging Face's own repositories, which are recognized as trusted services for AI models.
[COMMAND_EXECUTION]: The installation guide includes standard shell commands for managing Python dependencies via pip. The commands install well-known libraries such as diffusers, transformers, accelerate, torch, and xformers from the official Python Package Index (PyPI).
[INDIRECT_PROMPT_INJECTION]: The skill establishes a data ingestion surface by processing untrusted natural language prompts from users to generate images. This is a standard architectural pattern for diffusion-based image generation.
Ingestion points: User-supplied strings are passed to the prompt parameter of the DiffusionPipeline in SKILL.md.
Boundary markers: No explicit boundary markers or "ignore instructions" warnings are utilized in the examples.
Capability inventory: The skill demonstrated capabilities for reading local image files using PIL (Image.open) and writing generated images to the file system (image.save).
Sanitization: The examples do not implement text sanitization or prompt filtering, relying instead on the underlying model's behavior and the library's optional safety checker.
[PROMPT_INJECTION]: The skill provides documentation on common troubleshooting steps, including how to bypass the internal safety checker using pipe.safety_checker = None. While this is an override of a safety feature, it is presented as a standard developer configuration for handling specific library behaviors (such as black image output) and is consistent with the skill's primary purpose for research and development.

stable-diffusion-image-generation