The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill provides explicit and safe guidance for managing NVIDIA_API_KEY and NGC_API_KEY. It instructs users to avoid echoing keys in shell history, warns against storing them in world-readable shell startup files, and recommends using chmod 600 environment files or secret managers.
[EXTERNAL_DOWNLOADS]: The skill references official NVIDIA resources for downloads. It fetches container images from nvcr.io, the Riva Python client via pip install nvidia-riva-client, and sample code from official GitHub repositories under the nvidia-riva organization. These are recognized as trusted vendor resources.
[COMMAND_EXECUTION]: The skill uses standard shell commands and Docker for deployment, system readiness checks (nvidia-smi, docker info), and model optimization (riva-build, riva-deploy). These operations are necessary for the skill's primary purpose of managing NIM deployments and are performed using official vendor tools.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied text and audio files for transcription, synthesis, and translation tasks. While this presents an ingestion surface for potentially untrusted data, the skill relies on official NVIDIA NIM containers and client libraries for processing.
Ingestion points: User-provided text for TTS/NMT and audio files for ASR (references/asr.md, references/tts.md, references/nmt.md).
Boundary markers: The skill uses standard CLI flags and gRPC request fields to delimit data.
Capability inventory: Subprocess calls for Docker and system checks, network operations via gRPC/HTTP to NIM endpoints.
Sanitization: Relies on the underlying Riva/NIM service architecture to process speech and text data safely.

nemotron-speech