The Agent Skills Directory

[SAFE]: The skill defines a set of non-negotiable Refusal Rules that strictly prohibit the agent from performing potentially dangerous tasks, including system-wide package installations, bypassing CI/pre-commit checks, modifying files outside the repository (e.g., shell profiles), executing destructive commands like rm -rf or git reset --hard, and using elevated privileges via sudo.\n- [PROMPT_INJECTION]: The skill is hardened against instruction override attempts. It includes 40+ evaluation scenarios in evals/evals.json that specifically test and confirm the agent's refusal to follow malicious prompts, such as instructions to download and pipe scripts from untrusted domains to a shell or to incorporate unsafe eval() calls into the codebase.\n- [COMMAND_EXECUTION]: Execution of commands is strictly limited to project-specific development tasks like building the project (./build.sh), running test suites (pytest, ctest), and managing git state within a local fork. The skill explicitly requires the agent to seek user confirmation before any write operations, such as commits or pushes.\n- [REMOTE_CODE_EXECUTION]: A high-severity static analysis alert for eval with user input was determined to be a false positive. The pattern appears in a defensive evaluation test case (dev-038) designed to ensure the agent rejects such implementations, rather than as an instruction for the agent to follow.\n- [EXTERNAL_DOWNLOADS]: The skill references established development infrastructure, including NVIDIA's GitHub repositories and RAPIDS documentation. It includes specific defensive rules that block the downloading and execution of scripts from arbitrary or untrusted external URLs.

nvidia-cuopt-developer