The Agent Skills Directory

[COMMAND_EXECUTION]: The skill enables the execution of Nemotron customization steps using uv run. This is managed through a multi-stage workflow that prioritizes safety by enforcing opt-in execution via a --run flag, keeping commands scoped to repo-safe operations, and explicitly forbidding the use of environment dumps or secret logging.
[PROMPT_INJECTION]: The skill handles external inputs such as user-provided datasets (JSONL, Parquet) and environment profiles, which presents a surface for indirect prompt injection. To mitigate this, the skill requires a planning phase where the agent must document and validate the artifact graph, matched strategies, and hardware constraints before any code is generated or executed.
[SAFE]: No evidence of malicious behavior, obfuscation, or credential exfiltration was found. The skill includes proactive security measures for handling sensitive data (redacting tokens/passwords) and verifying the integrity of Hugging Face snapshots and remote endpoints.

nemotron-customize