Celeris Core Agent Instructions

A task-based path from no prior model knowledge to defensible, reproducible Celeris-WebGPU work.

Open the authoritative instructions

Give this prompt to Codex

Use this once at the beginning of a Celeris task. The agent will load deeper material only when the request requires it.

Read https://celeris.usc.edu/docs/core/AGENT_INSTRUCTIONS.md and follow its task router. Work only with the main two-dimensional Celeris Core simulator; exclude transect_version. First classify my request as explain, create, run, analyze, validate, or modify Core. Read only the references required for that task, state material assumptions, and use https://celeris.usc.edu/docs/core/reference/README.md when deeper technical knowledge is needed.

Choose the user's task

Explain physics or parameters

Route directly to the relevant equations, algorithms, controls, source evidence, and validation limits.

Create or revise a case

Select compatible physics, then use configuration, input, example, and scripting guidance as needed.

Run a first case

Complete one known-working run; experienced users can go straight to running and automation.

Analyze output

Decode fields, reconstruct variables, retain metadata, and build reproducible analysis.

Validate a result

Use benchmarks, sensitivity tests, conservation checks, and the correct level of evidence.

Modify Core implementation

Enter through architecture, then inspect only the affected handlers, texture contracts, shaders, and source.

Deep knowledge, on demand

Open the Deep-Knowledge Index for any detailed question about equations, tunable parameters, numerical methods, input conversion, output encoding, diagnostics, or implementation.

The index covers water, granular, and coupled flow; all auxiliary physical processes; boundaries and forcing; numerical algorithms; configuration; input/output pipelines; rendering; examples; and validation. Its machine-readable component catalog lets an agent route a question without scanning every document.

Open the component parameter index.

Download the complete Core Markdown documentation ZIP for local agent use. SHA-256 checksum.

Core operating rules

  1. Treat current JavaScript and WGSL as authoritative when older prose disagrees.
  2. Configure a case in its own directory; do not edit the solver merely to make a simulation.
  3. Start from the closest working example with compatible physics, coordinates, and forcing.
  4. Preserve units, datum, CRS, grid orientation, provenance, and explicit versus inferred choices.
  5. Validate inputs, perform a short browser smoke test, then assess conservation, sensitivity, and reference evidence.
  6. Report assumptions, limitations, and validation status; a completed run is not automatically a validated prediction.

Optional reference shelves

Operational guides
Architecture and source contracts
Examples, benchmarks, and utilities