The Agent Skills Directory

[COMMAND_EXECUTION]: The skill leverages uv run to execute repository-specific Nemotron steps. It includes several safety guardrails, such as defaulting to --dry-run and requiring explicit user confirmation before performing remote submissions, destructive changes, or expensive launches.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a significant attack surface as it is designed to ingest and process external datasets (e.g., JSONL, Parquet, and Hugging Face repositories). It mitigates this risk by requiring the agent to validate artifact wiring during the planning phase and by encouraging the use of structured data and explicit field mapping to maintain data integrity.
[DYNAMIC_EXECUTION]: The skill includes functionality to generate Python scaffolding (e.g., cli.py, run.py, and deployment DAGs) for multi-step pipelines. This dynamic code generation follows established templates, wraps official library APIs, and is governed by a multi-phase workflow (Orient-Plan-Act-Verify) that requires user approval before implementation.
[CREDENTIALS_SAFE]: The skill contains explicit instructions to handle authentication tokens (such as NVIDIA_API_KEY and HF_TOKEN) securely. It mandates that secrets be referenced only by their environment variable names, prohibits inlining or committing values, and requires the redaction of sensitive fields during any file inspection tasks.

nemotron-customize