Agent skills
The RunMat plugin includes two skills: one for running self-contained MATLAB-syntax code and another for working across the files in a RunMat Project. Both use the same RunMat tools, with instructions tailored to each kind of work.
Run MATLAB-syntax code
run-matlab-code is for self-contained execution and result inspection. It applies when you ask the model to:
- execute or verify MATLAB-syntax code;
- debug a script using actual runtime diagnostics;
- benchmark a numerical operation;
- inspect values, shapes, or types; or
- create an interactive 2D or 3D figure.
The skill checks that RunMat is ready, executes the relevant code in the selected Project, and inspects diagnostics and results before answering. Conceptual questions that need no execution stay in the conversation.
Example:
Use RunMat to compare the numerical stability of these two matrix formulations and plot the error.Work in a RunMat Project
work-in-runmat-project is for durable or multi-file work. It applies when you ask the model to:
- inspect or repair an existing
.mProject; - work across multiple files or a
runmat.tomlmanifest; - use local data or a repository root;
- make reviewable patches;
- run a file or source selection; or
- continue work in an open RunMat Desktop Project.
The skill identifies the intended Project, reads the relevant revision before editing, applies bounded changes, runs the appropriate entrypoint, and handles revision conflicts explicitly.
Example:
Open my RunMat Project, find why main.m fails, apply a reviewable fix, and run it again.Automatic and explicit use
Natural-language requests are usually enough. The plugin descriptions help the model load the matching skill only when it is relevant.
For explicit selection:
- In Codex, use
$run-matlab-codeor$work-in-runmat-project. - In Claude Code, use
/and select the namespaced RunMat skill supplied by the plugin.
Selecting a skill explicitly chooses how the model approaches the task. Project permissions and execution limits still apply.
Where the skills can work
The skills work in the Projects open in RunMat Desktop, from Claude Code or Codex on the same computer. The first time a tool works in a Project, RunMat asks you to allow it.
The skill keeps using the same Project. It tells the model if RunMat Desktop closes or disconnects, or Project permission expires.
Toolbox questions
RunMat implements more than 2,100 core builtins and all core language grammar and semantics. When a request depends on a niche toolbox function, the skill checks the RunMat documentation before claiming support.
You can ask a conceptual question about unsupported functionality without consuming an execution. An execution is counted only when RunMat starts a runtime job; see Usage and limits.