The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the agent to clone the project from its official GitHub repository and install dependencies using pip. It also specifies downloading pre-trained model weights from Hugging Face (e.g., liuhaotian/llava-v1.5-7b). These actions are standard for the deployment of this specific open-source architecture.
[COMMAND_EXECUTION]: The skill includes shell commands for launching a Gradio web server, running a command-line interface for model interaction, and executing training pipelines via bash and deepspeed. These commands are essential for the primary functionality of using and fine-tuning the model.
[INDIRECT_PROMPT_INJECTION]: The model is designed to process external images and user-provided text instructions, which represents an inherent attack surface for multimodal AI. Malicious instructions could potentially be embedded within images or prompts processed by the model.
Ingestion points: External images loaded via PIL.Image.open and user text prompts appended to conversation templates in SKILL.md.
Boundary markers: The skill uses conversation templates (conv_templates["llava_v1"]) to structure inputs, which provides some structural separation between user input and model instructions.
Capability inventory: The skill facilitates repository cloning, package installation, shell script execution for training, and writing local JSON configuration files.
Sanitization: There is no explicit evidence of input sanitization or filtering for the image or text content before processing.

llava