AI agent skill#
An AI coding agent can read and edit code and run tests with the tools you make available. The examples on this page assume a recent frontier coding or reasoning model with access to your project, source documents, data, and Julia execution tools. Choose a model suitable for the complexity of your scientific task; a chat session without those tools cannot execute or validate the code.
PlantSimEngine ships an optional skill: a folder of instructions, references, and executable examples that an agent can read as needed. It covers model authoring, coupling, debugging, observations, calibration, and evaluation.
You choose the scientific question, assumptions, references, and validation evidence. The agent can help turn those choices into code and check that the declared models connect as intended.
Set up the skill for your package version#
Find the skill shipped with the PlantSimEngine version you actually use:
julia
using PlantSimEngine
package_root = pkgdir(PlantSimEngine)
skill_root = joinpath(package_root, "skills", "plantsimengine")
(Base.pathof(PlantSimEngine), Base.pkgversion(PlantSimEngine), skill_root)Copy or link that complete directory into your coding agent's skill directory, following its installation instructions. Keep its SKILL.md, references, assets, scripts, and evaluations together. Update the skill when you update PlantSimEngine so its instructions describe the API you actually use.
The skill uses the Julia tools available in your agent's environment. It does not require a particular editor or connector. We recommend installing Kaimon as an optional way to give your agent a live Julia session for execution, inspection, and debugging. Its Quick Start explains installation and connecting your coding agent.
Describe the result and provide the evidence#
Start with the scientific goal, relevant files, constraints, and the result you want to inspect. Let the agent choose how to investigate and implement it. Detailed instructions help when a scientific protocol requires them or an observed mistake needs correcting; a long sequence of generic reasoning or verification instructions is rarely a useful starting point.
Current OpenAI guidance on skills and prompts and Anthropic's skill authoring guidance recommend focused instructions and references loaded as needed. Check their effect on your own tasks; scientific requirements still need to be explicit.
Tell the agent to use the skill shipped with your loaded PlantSimEngine package. The following prompts are starting points to adapt, with square brackets marking information from your study.
Implement a model from a paper#
Provide the paper and relevant supplements, the formulation you want, and any reference implementation or known results. State the intended object scale and time interval.
Implement [formulation] from [paper and supplements] as a reusable PlantSimEngine model for [scale and interval]. Keep equations and parameter sources traceable to their pages, equation numbers, or tables. Identify unresolved scientific choices without inventing values. Deliver readable model code, a minimal scenario, and checks against [reference results], with the assumptions and remaining limitations.
See Implement a model from a paper, Implement a basic model, and Port an existing model. The skill's references/paper-to-model.md supplies the corresponding scientific guidance.
Couple component models#
Provide the model packages, objects and topology, weather or spatial sources, units, and the outputs of interest. Mention any known conversions or timing requirements.
Couple [component models] for [objects and topology] using [input data] at [cadences]. Explain the selected sources, units, physical bases, and time meanings of the connections. Keep conversions explicit and flag missing inputs or incompatible interfaces. Deliver the scenario and evidence that it connects the intended objects and preserves [expected relationship].
See Coupling models, Model catalog and related packages, and Model compatibility and replacement. The skill includes references/scenario-coupling.md and runnable coupling examples.
Debug an unexpected result#
Provide a runnable scenario, its inputs, the error or unexpected output, and what you expected and why. Include the relevant observation or reference result, rather than only a plot showing disagreement.
Investigate [error or discrepancy] in [scenario] with [inputs]. The expected result is [result and evidence]. Find where the discrepancy first appears and distinguish configuration, numerical, and scientific causes. Show the evidence for your diagnosis and a verified correction where supported; leave unsupported hypotheses explicit. Keep parameter tuning separate.
See Inspect a simulation, Common Errors, and Debugging Growth And Resource Ordering. The skill's references/diagnostics.md maps symptoms to the public inspection tools.
Force a model with observations#
Identify which measured variable should replace a predicted input, its units, object identity, timestamps, sampling interval, and missing-data rules. State which downstream results should still be predicted.
Use measured [variable] from [dataset] to drive [consumer models] instead of predicting that input. Respect [units, object IDs, and sampling policy] and document any time alignment or conversions. Show the values supplied at model updates and the downstream predictions. Keep the imposed values distinct from results used to assess model performance.
See Use Observed Values In A Simulation. Supplying observations may require removing the model that ordinarily predicts the same input; choosing two writers is a different experiment.
Calibrate selected parameters#
Provide the calibration data and its protocol, target variable, parameter names and bounds, objective, and the data reserved for evaluation. Make clear whether other parameters and initial conditions are fixed.
Calibrate [parameters within bounds] against [target observations] using [calibration subset and objective]. Keep [fixed parameters and initial conditions] fixed and reserve [evaluation subset]. Report the parameter estimates, fitting diagnostics, objective value, and limitations. Save the configuration and data identifiers needed to reproduce the fit.
See Parameter Fitting. PlantSimEngine's Evaluation.fit is an extension point; the model package or study supplies its fitting method and optimizer.
Evaluate on independent observations#
Provide frozen parameter sets, the evaluation subset, the prediction target, matching rules, and meaningful metrics. Keep any model selection or tuning data separate from the independent evaluation data.
Evaluate [frozen baseline and calibrated models] on [independent dataset] for [target, scale, and interval] without retuning. Match predictions to observations using [matching rules]. Report [metrics], residual patterns, and differences by [site or treatment], with sample counts and exclusions. Explain which conclusions the evaluation supports and where uncertainty or missing coverage limits them.
See Collecting And Plotting Outputs and Parameter Fitting. The skill's references/data-calibration-evaluation.md covers observation roles, data alignment, calibration, and independent evaluation.
Review what the agent returns#
Expect a concrete result you can inspect:
the loaded package path and version;
the selected process and hypotheses, with the evidence supplied for them;
readable model code and scenario configuration;
checks that the required inputs are available, their units and meanings match, and the models receive values from the intended sources;
tests actually run, their results, and any remaining limitations.
For a data-based study, also expect the observation roles and alignment rules, the calibration/evaluation split, parameter provenance, and reproducible metrics. An optimizer converging or a declaration check passing does not establish predictive accuracy.
PlantSimEngine's Authoring reports describe models and compare interfaces. Diagnostics explains inputs, connections, timing, and execution. The reports show which information comes from the model author, which was found by examining the code, and which is missing. These checks help find implementation and coupling errors; scientific validation still needs suitable observations or reference results.
Check the packaged examples#
The skill includes a minimal model, alternative hypotheses, a physical conversion model, coupling examples, and their tests. You or your agent can check the copy supplied by the installed package with:
julia
include(joinpath(
pkgdir(PlantSimEngine), "skills", "plantsimengine",
"scripts", "check-examples.jl",
))Use Model repository layout and tests when you are ready to organize your own reusable model package.
On this page