The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes 3D ground truth and USD scene data to generate descriptions via an external NVIDIA NIM LLM. This creates a surface for indirect prompt injection where adversarial scene content could influence the generated captions.
Ingestion points: Scene USD files and prim data processed via the isaacsim.replicator.caption.core extension.
Boundary markers: None explicitly defined in the provided scripts or instructions to delimit scene data from instructions.
Capability inventory: Writing output files to the output_path and performing network requests to the NVIDIA NIM API endpoint.
Sanitization: The skill does not mention explicit validation or sanitization of the scene content before it is processed by the LLM.
[COMMAND_EXECUTION]: The skill provides scripts/starter_irc_config.py, a local utility script used to generate configuration YAML files. The documentation suggests running this script using the run_script helper.
[EXTERNAL_DOWNLOADS]: The skill makes API calls to https://integrate.api.nvidia.com/v1 to perform captioning tasks. This corresponds to the official endpoint for NVIDIA NIM services.

vlm-scene-captioning