The Agent Skills Directory

[SAFE]: The analyzed skill is composed entirely of technical documentation and instructions for the agent to retrieve information from local snapshots. No malicious patterns, hardcoded credentials, or unauthorized access attempts were identified in the provided content.
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: The skill workflows involve reading markdown files within the references/ directory (e.g., ptx-docs, cuda-guide).
Boundary markers: Not explicitly used in the querying instructions; the agent is expected to parse standard markdown documentation.
Capability inventory: The skill defines workflows using search and development tools like rg (ripgrep), nsys, ncu, and compute-sanitizer on user-provided code and local documentation.
Sanitization: None present; the skill treats the static documentation as an authoritative and safe knowledge base.

cuda-skill