NVIDIA cuOpt User Rules for Codex: Verify a Small Optimization Model
Apply the cuOpt skill's input and result checks to a simple optimization prompt before relying on a solver or a larger routing data set.
中文Setup for Codex (NVIDIA plugin)
Codex with a toy model that can be solved by hand. Real cuOpt runs depend on the selected solver settings and compatible NVIDIA environment.
- Open the pinned GitHub source for this version. Copy the entire `skills/cuopt-user-rules/` folder with its references into your scratch project's `.agents/skills/cuopt-user-rules/`.
- Start Codex from that project and load the skill. State the decision variables, objective units and required bounds before running a solver. Ask for a direct check of the reported solution.
Example task
Minimize 3x + 5y given x + y = 4 and nonnegative x and y. Ask the agent to show the optimal assignment and plug it back into the objective, using this local problem as a solver-free fixture.Expected result
The minimum is 12 at x = 4 and y = 0. The assignment obeys the sum and both lower bounds. The expected status can be checked without installing cuOpt.
Upstream documentation
Original SKILL.md on GitHub · Apache-2.0 license · Source reviewed
Frequently asked question
Why check a cuOpt optimization result against the original constraints?
A reported assignment can be re-evaluated against the stated bounds and objective. This catches a mismatch in model inputs or interpretation.