Best Practices

Use AI tools as fast drafting and review assistants. Keep human control over physics, numerical methods, validation, and final code acceptance.

Recommended Workflow

1. Start from an existing Exasim example.
2. Provide the AI with exact file paths and relevant source files.
3. Specify geometry, units, parameters, boundary conditions, and outputs.
4. Ask for incremental changes.
5. Review assumptions before accepting edits.
6. Compile and run small cases frequently.
7. Validate against analytical, manufactured, or benchmark results.
8. Iterate with narrow follow-up prompts.

Good Habits

• Tell the AI what not to change.
• Ask for affected files before coding.
• Ask for explicit assumptions.
• Ask for verification commands.
• Keep a clean git diff.
• Review generated array sizes and dimensions.
• Prefer simple examples before complex multiphysics.
• Preserve existing Exasim APIs unless intentionally extending them.

Common Mistakes

Mistake	Consequence
"Create a solver" without details	Underspecified physics and unusable code.
Missing boundary conditions	Ill-posed or incorrect PDE problem.
Missing units	Wrong scale, stiffness, or nondimensional parameters.
Too many changes at once	Hard-to-review diffs and hidden regressions.
No reference example	Wrong frontend style or obsolete API usage.
No validation criteria	Code may run but solve the wrong problem.
Ignoring GPU/MPI data movement	CPU-only code or stale device data bugs.

Best Prompt Style

Use direct, testable instructions:

Modify only <file>. Preserve existing behavior unless <condition>. Add a small
test or example. Run <command>. Explain residual risks.

When To Stop And Ask For Human Review

Request expert review when a change affects:

• conservation or consistency;
• HDG static condensation;
• nonlinear or linear solver logic;
• preconditioner storage;
• MPI communication;
• CUDA/HIP memory ownership;
• production physics models;
• benchmark results.