The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input data (text captions and images) from JSONL files, which are subsequently interpreted by a Vision-Language Model (VLM).
Ingestion points: The data.input_jsonl file processed in SKILL.md and defined in references/configuration.md.
Boundary markers: No explicit prompt delimiters or "ignore embedded instructions" warnings are provided in the pipeline steps to separate instructions from input data.
Capability inventory: The skill executes the auto_label tool via Bash and writes results to the local filesystem.
Sanitization: No sanitization or filtering of the input JSONL content is described in the provided instructions.
[COMMAND_EXECUTION]: The skill is built around executing CLI tools and managing containerized services.
Executes the auto_label generate command within the TAO Toolkit container.
Provides instructions for running Docker commands (docker run --runtime nvidia) to host local inference servers using the vllm-openai image.
[EXTERNAL_DOWNLOADS]: The skill references and downloads resources from well-known and trusted sources.
Pulls official Docker images from the NVIDIA Container Registry (nvcr.io/nvidia/tao/tao-toolkit).
Recommends pulling the vllm/vllm-openai image from Docker Hub and installing the vllm package via pip from the official vLLM project repository.

tao-generate-image-grounding