Simulink Generate Embedded Code

Generate Embedded Code from Simulink Models

Configure Simulink models for Embedded Coder code generation using the configure_for_codegen function.

When to Use

• User asks to configure a model for code generation, production deployment, or Embedded Coder
• User mentions ERT, AUTOSAR, ARM, embedded target, or production code
• User wants to optimize generated code for speed or RAM
• User asks for MISRA C or MISRA C++ compliance
• User describes a deployment target in domain language ("deploy to ECU", "minimize flash")

When NOT to Use

• Simulation-only tasks (running a model, tuning parameters, viewing signals)
• Data dictionary or bus object configuration (standalone, not as part of code-gen setup)
• Test harness creation or coverage analysis
• Simulink Coder (GRT) targets — this skill only handles Embedded Coder (ERT) and AUTOSAR
• DDS (Data Distribution Service) or ROS targets
• Modifying or formatting generated code files after code generation

Rules

• Never expose the script's wrapper parameter names — describe the decision and its effect in domain terms. The configure_for_codegen wrapper parameter names (ConfigOnly, Build, Interface, OutputDir, Compliance, Objective, Target, Hardware, Language) are an implementation detail of the script — they must never appear in text you show the user, and neither should name="value" syntax. This prohibition covers only the wrapper arguments. It does not restrict describing what the configuration does at the model or ConfigSet level: you may and should explain the engineering substance of each choice in domain terms. Combine both — name the decision, then describe its effect. Tier-1 decision phrasing (safe as-is): "configured the model but did not build", "used a nonreusable function interface", "optimized for speed", "placed generated code next to the model". Concept-level effect phrasing (also safe, and expected on cascade-driven decisions): "optimizing for speed enables an execution-efficiency cascade — strength reduction, inlined parameters, and removal of division-by-zero protection"; "a reusable function interface produces multi-instance/reentrant code that passes state by pointer"; "a MISRA C profile drives specific style rules — casting mode, signed shifts, unreachable-default suppression". Caution: prefer engineering-concept language over raw Simulink ConfigSet parameter identifiers — say "inlines parameters," not "sets InlineInvariantSignals"; say "removes division-by-zero protection," not "sets NoFixptDivByZeroProtection". This applies to defaults, confirmations, follow-up questions, and error paraphrasing — everywhere except the phrase-mapping table below (which is your internal lookup, not user-facing).

• Model must be open. The model must already be open in MATLAB before calling configure_for_codegen. If it is not, open it directly with open_system('<model>') via evaluate_matlab_code — no separate skill is needed for this.

• Model must be saved to disk (or the user must supply a destination). The script writes generated code next to the model's .slx file. If the model is a fresh new_system/untitled window with no file on disk, the script cannot infer a location and returns success:false with a message asking the user to save the model or supply an explicit output directory. Relay that message verbatim and ask the user which they prefer — never guess a location on their behalf, and never call save_system without permission (see "Do not save" rule below).

• Model must be compilable. The model and all its dependencies (data dictionaries, referenced models, bus objects, MATLAB path entries) must be resolvable in the current MATLAB session. If configuration fails due to missing dependencies, inform the user what is unresolved and ask them to fix the environment — do not attempt addpath or other path manipulation on their behalf.

• Do not save; inform, then offer. The script does not call save_system and does not persist any bound data dictionaries. When configuration succeeds, explicitly tell the user nothing has been saved — model and dictionary changes exist only in memory, so they can review or discard freely. As a courtesy, ask whether they want to save now. Only save if the user says yes. Saving is outside the scope of this skill's core responsibility.

  • Name every file that would be saved. The script's JSON output includes a pendingSaves field listing the exact files whose in-memory state must be persisted for the configuration to reload correctly — always at least each configured .slx, plus any .sldd when a model's active ConfigSet is a ConfigSetRef backed by a data dictionary (the script writes the applied settings into a new dictionary entry, and saving the .slx alone would persist a reference to an entry that is not yet on disk — on reload, get_param(model, 'SystemTargetFile') would fail with "Unable to get parameter ... from referenced configuration set ... in data dictionary.").
  • Save via the mechanism named by kind. Read pendingSaves from the script output, name every entry when offering to save, and — on user consent — save each file using the mechanism named in its kind field (model → save_system, dictionary → Simulink.data.dictionary.open(...).saveChanges()). Never use Simulink.data.dictionary.saveAll; it would persist unrelated dictionaries that happen to be open in the session. If the user declines, save nothing.

• All configuration is performed via configure_for_codegen (in the skill's scripts/ directory), called through evaluate_matlab_code with project_path set to SKILL_DIR/scripts — where SKILL_DIR is the absolute filesystem path this SKILL.md was loaded from. Do not guess a workspace-relative path (MCP rejects paths under .claude/); resolve SKILL_DIR to whichever absolute location the loader used. Do not addpath the user's model or dictionary folders — the script handles CWD/dictionary resolution internally and self-cleans. Pass model as an open model name — with or without the .slx/.mdl extension; the script strips it either way (e.g. 'mbasic' or 'mbasic.slx' both work). The model must already be open (see rule above). The agent never calls set_param directly.

• Resolve MISRA ambiguity in your response. "MISRA" alone is ambiguous — MISRA C and MISRA C++ are distinct standards. If the user says only "MISRA" without specifying, pick the one matching the source language (MISRA C for C, MISRA C++ for C++) and explicitly state the assumption in your final response — e.g., "Assuming MISRA C since the target language is C. Let me know if you meant MISRA C++."

• State your inference for every required choice you didn't get verbatim from the user. The four required inputs — target framework, source language, hardware device, and optimization objective — must each either come directly from the user or be a silently-safe inference that you explicitly narrate back. Never expose the script's wrapper parameter names. Common inferences to narrate (same treatment as the MISRA ambiguity rule):

  User phrasing	Inferred choice	How to say it back
  "embedded C", "production C" (no C++ mention)	source language = C	"Using C as the source language since you said 'embedded C' — let me know if you meant C++."
  "AUTOSAR" (bare, no Classic/Adaptive)	target = AUTOSAR Classic	"Using AUTOSAR Classic since you didn't specify — say the word if you meant AUTOSAR Adaptive."
  "embedded", "production", "ECU" (no shared-library / AUTOSAR mention)	target = ERT executable	"Building for an ERT executable target — tell me if you want a shared library instead."
  "ARM Cortex-A", "Cortex-M4", "x86-64" (short form)	full hardware device string	"Mapped 'ARM Cortex-A' to the ARM Compatible ARM Cortex-A device — let me know if you meant a different variant."

• Ask one batched clarifying question for what you cannot safely infer — never guess silently. Among the four required inputs, hardware and objective are never silently defaulted — the wrong device string picks the wrong word sizes and endianness, and the choice between speed, memory footprint, and step-through debuggability is a real engineering tradeoff the user must own. target may be inferred as ERT executable when the user says "embedded", "production", "ECU", or nothing target-related; language may be inferred as C in the same conditions. If either or both of hardware / objective is missing after the phrase-mapping pass, ask a single consolidated question naming exactly what you need, then proceed once. Do not ask one parameter at a time. Example: "To configure this I need two things: (1) which processor family — ARM Cortex-A, ARM Cortex-M, x86-64, PowerPC, or something else? (2) should I optimize for execution speed, memory footprint, or step-through debuggability (preserves block-to-code traceability)?" This is the one allowed exception to the "one call, not two" rule below — a single upfront clarifying question, then exactly one script call.

• Extract all parameters before the first call — one call, not two. Read the user's request once and map every phrase to a configure_for_codegen parameter, then invoke exactly once. Making a second corrective call to add a parameter you forgot (e.g. ConfigOnly, Compliance, Build) is a rule violation. Common phrase mappings:

  User phrase	Parameter
  "just configure", "apply settings only", "prepare the model but don't generate code"	ConfigOnly=true
  "codegen", "generate code", "generate the code but don't build/compile", "generate code without building an executable"	(no override — defaults ConfigOnly=false, Build=false generate source without a toolchain build)
  "build it", "compile it", "generate and build", "produce an executable"	Build=true
  "MISRA" (bare)	Compliance="MISRA C" or "MISRA C++" per Language (see MISRA ambiguity rule)
  "speed", "fast", "execution efficiency"	Objective="Speed"
  "small flash", "minimize RAM", "memory-constrained", "low memory"	Objective="RAM"
  "debug", "debuggable", "step through", "step-through", "traceability", "preserve block structure"	Objective="Debug"

  Three distinct build modes — do not conflate them. The words "don't build" / "don't compile" appear in two of these buckets and are ambiguous on their own; disambiguate by whether the user mentions generating code:

  Intent	User signal	Mode
  Apply ConfigSet only, produce no code	user says "just configure" / "apply settings" and does not mention generating code	ConfigOnly=true (skips rtwbuild entirely)
  Generate source files, skip toolchain build	user says "generate code" or "codegen" while also saying "don't build/compile" (or says nothing about building)	defaults — rtwbuild(..., 'GenerateCodeOnly', true)
  Generate source and produce an executable	user says "build" / "compile" / "produce an executable"	Build=true (calls slbuild)

  If the user mentions "generate code" anywhere in the request, ConfigOnly=true is wrong — code generation is what they asked for. "Don't build/compile" in that same sentence means skip the executable build, not skip codegen.

• Never modify generated code files. After code generation, the output files (.c, .cpp, .h, .arxml, etc.) are read-only artifacts. Do not edit, reformat, or overwrite them.

• Relay success:false verbatim — never invent workarounds. The script returns JSON with a success flag and a populated errors[] list when a prerequisite is missing (e.g. AUTOSAR Blockset not installed or unlicensed, model not loaded, invalid MISRA/language combination). When success:false, quote the error message to the user and stop. Do NOT retry with ConfigOnly=true to mask the missing dependency, do NOT fall back to direct set_param() calls, and do NOT rerun with different flags hoping the second try succeeds. The refusal is the correct outcome — the user needs the specific error to fix their environment.

• By default, generated code is placed next to the model file. Only pass OutputDir if the user explicitly requests a different location.

• Report assumed defaults in plain language — name every one, don't cherry-pick, and never expose the script's wrapper parameter names. When the user does not specify a value, tell them which default you are applying for each of these decisions: function interface style, MISRA compliance, whether to build, whether to configure only, and output location. Describe every choice in user-facing prose — do not name the underlying script wrapper arguments (no ConfigOnly, Build, Interface=, OutputDir=, etc.) or show name=value pairs. Engineering-substance descriptions of what those choices do at the model/ConfigSet level are fine and encouraged where relevant. If the user didn't name a choice, you must still name your default — even for the "none" case (e.g. no MISRA compliance). Example: "Since you didn't specify, I applied these defaults: nonreusable function interface, no MISRA compliance profile, configured the model but did not run a build, and placed generated code next to the model file. Let me know if you want any of these changed."

• Only itemize parameterChanges when the user explicitly asks for it — otherwise summarize in domain terms. The script's JSON output includes a parameterChanges array with parameter, description, from, and to fields for every ConfigSet parameter it touched. Do NOT enumerate this by default — a domain-phrased summary of what the configuration accomplishes is the expected response shape. Render the table only when the user's prompt explicitly requests structured detail, e.g. "list every parameter you changed", "show me the changes as a table", "which parameters did the configuration modify?", "produce a report of the applied settings". When asked, render as a 4-column markdown table:

  Parameter	Description	From	To

  Populate directly from the script fields — parameter (raw ConfigSet identifier such as SystemTargetFile), description (the ConfigSet's own human-readable prompt, e.g. "System target file"), from (previous value), to (applied value). This opt-in table is the one context where raw ConfigSet parameter identifiers may appear in user-facing output; the description column keeps every row engineering-substantive so the table remains readable. Everywhere else — defaults, confirmations, follow-ups, error paraphrasing — the "prefer engineering-concept language over raw parameter identifiers" guidance above still governs.

• Warn before long operations. When the user mentions a model hierarchy or sets Build=true, inform them that configuration may take significant time for complex/large hierarchies before invoking the script.

• Never read script source at runtime — use this interface documentation as the invocation contract.

Script Interface

configure_for_codegen(model, Name=Value)

Atomic Embedded Coder configuration via ConfigSetManager. Configures the model (and all referenced sub-models) or reverts entirely on failure.

Positional input:

Name-value arguments:

Output: JSON string with fields: success, modelsConfigured, artifactsGenerated, reportPath, errors, parameterChanges, pendingSaves. parameterChanges is an array of {parameter, description, from, to} — one entry per ConfigSet parameter the script touched. Do not itemize this by default; render only when the user explicitly asks (see the "Only itemize parameterChanges" rule above). pendingSaves is an array of {kind, name, path} — kind is "model" or "dictionary"; use it to name files when offering to save and to pick the save mechanism (save_system vs Simulink.data.dictionary.open(...).saveChanges()).

Example call (invoked via evaluate_matlab_code with project_path set to this skill's scripts/ directory):

configure_for_codegen('mbasic', Target="ert", Language="C", Hardware="Intel->x86-64 (Linux 64)", Objective="Speed")

Copyright 2026 The MathWorks, Inc.