The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the user to clone the LLaVA source code from its public GitHub repository (https://github.com/haotian-liu/LLaVA) and install it locally using pip install -e ..
[COMMAND_EXECUTION]: The skill and its references include commands to run inference servers, Gradio web interfaces, and DeepSpeed-based training scripts (llava.serve.cli, llava.serve.gradio_web_server, and llava/train/train_mem.py). These are standard operations for model deployment and research.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a surface for indirect prompt injection as it processes external images and user-provided text queries.
Ingestion points: Images are loaded via Image.open() or CLI arguments in SKILL.md.
Boundary markers: No explicit boundary markers or instructions to ignore embedded visual text are present in the example prompts.
Capability inventory: The documented capabilities are limited to text generation and terminal output; no high-risk actions (like file writes or network requests) are triggered by the model's output in the provided code.
Sanitization: No sanitization of the generated output is shown before it is printed to the console.

llava