The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the Bash tool to interact with the nemotron CLI. This is the intended primary purpose of the skill for planning, dry-running, and executing retrieval recipe stages.
[EXTERNAL_DOWNLOADS]: The skill references downloading Docker images from the NVIDIA Container Registry (nvcr.io) and model checkpoints from the Hugging Face nvidia organization. These are well-known and vendor-controlled resources.
[DYNAMIC_EXECUTION]: The skill instructions specify the use of trust_remote_code=true when loading model checkpoints via the Transformers library. While this allows for the execution of code bundled with the model, it is a standard practice for these specific NVIDIA model architectures and is limited to the vendor's own verified repositories.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided text corpora for Synthetic Data Generation (SDG). This represents an attack surface where malicious instructions could be embedded in the training data. The skill mitigates this risk with explicit instructions for the agent to confirm data-governance policies with the user before processing and sending content to inference endpoints.
Ingestion points: User-provided corpus_dir or Hugging Face dataset URIs.
Boundary markers: Instructions warn the agent to confirm policy approval before Stage 0 SDG.
Capability inventory: The skill has Bash access, can read workspace files, and perform network requests to inference APIs.
Sanitization: The skill relies on user-governance confirmation and the use of dedicated inference APIs rather than direct prompt interpolation for command generation.

nemotron-retrieval-recipes