AI Integration with Next Design MCP

Next Design MCP allows AI agents to view and edit Next Design models and profiles.

This page explains how to prevent unintended operations and configuration omissions when requesting AI agent actions.

For information on setting up AI agents and DC Agentiqs, please refer to here.

Basic Procedure

When requesting viewing, verification, or modification using Next Design MCP, consider the following points depending on the purpose and operation:

• (A) Confirm the state (context) before execution
• (B) Specify the purpose and operation
• (C) Make changes in small units and verify

Do not judge solely on the AI ​​agent's completion report

Even if the AI ​​agent responds with "completed," the actual model field values, relationships, and generated count may not be as expected. Always check the results in Next Design and, if necessary, carefully examine the processing results for each target.

(A) Confirm the state (context) before execution

Next Design MCP recognizes the following states in Next Design as context:

• The state of the currently displayed view

• The model selected in the view

When making a request to Next Design MCP, confirming the above states before giving instructions will clarify the target of the operation.

(B) Specify the purpose and operation details

To specify the purpose and operation details of the task you are requesting from the AI ​​agent, we recommend including the following information in your instructions:

When targeting a design model

Information to include in the instructions	Description	Example
Purpose	What you want the AI ​​agent to achieve	Verify the trace (derivation relationships) of requirements and components
Target	Information that identifies the model to operate on and verify	Requirements model under the package "Vehicle Control"
Operation details	Operations such as viewing, creating, and updating	Refer to derivation relationships and list unconfigured requirements
Relationship type	Type of relationship to operate on (*1)	Derivation relationship between requirements and components
Search conditions	Direction, hierarchy, and depth of search to traverse the relationship	Child models of requirements. Child models are limited to one level
NotApplicable	Packages, models, and relationships not to be manipulated	Reference relationships, parent models, and models other than the "Vehicle Control" package
Completion conditions	Expected outcome state	Required fields, necessary trace information, and view display state

Hint (*1)

Next Design distinguishes between ownership relationships, reference relationships, and derived relationships. When relationships are complex, requests like "Check relationships" may leave it unclear whether the AI ​​agent will handle the relationships as expected. In such cases, it is recommended to clearly specify which relationship (ownership, reference, or derived) to handle, and the direction to trace the relationships.

info

Even with ambiguous instructions and insufficient information, the AI ​​agent may proceed by inferring and supplementing the information.

By instructing the agent to present missing information as unconfirmed items or items requiring user judgment, rather than relying on inference, you can proceed while clarifying uncertainties.

When targeting profiles

When creating a class diagram for a profile, specify the scope and division unit for each process or responsibility. If not specified, only one large class diagram including the entire profile may be created.

(C) Change and verify in small units

Changing many models at once may result in manipulating unintended models or missing some settings. Divide the target into small units, verify the results with a small number of models, and then expand the scope.

When processing multiple models together, it is recommended to repeat the following steps for each target unit:

1. Obtain the target and confirm the name, ID, and number.
2. Confirm the necessary processing content for each target and request it.
3. Output a list of processed and unprocessed targets.
4. Check the processing results and compare them with the list.
5. Save the project at the end of each processing stage so that it can be restored.

Batch processing may result in missing required fields, related settings, or duplicate targets. Record the IDs and names of the items to be processed, and compare the number of processed items, unprocessed items, and duplicates with the total number of items.

Useful methods for verifying processing results include checking the model differences before and after changes, and checking the differences with the most recent commit.

Do not treat the AI ​​agent's responses or completion reports as direct evidence of the actual execution results.

Even results such as "0 items," "empty," or "not set" should not be judged solely on the response; verify the actual data, relationships, and number of items in the target model using other methods.

It is recommended to ask the AI ​​agent to include inferred information from names and context, and actually confirmed facts, in its responses.

When dealing with long conversations or large amounts of context, consider the possibility that previous instructions may not be sufficiently referenced, and clearly specify the necessary information such as the target, purpose, conditions, and expected results.

Creating and Adding Skills According to the Target Domain

Here, the target domain refers to the scope (domain) that shares common design targets, terminology, and design/operation rules.

When projects are large in scale, or when profiles define numerous classes and fields, Next Design MCP's general-purpose skills alone may not be sufficient to efficiently search for the desired models and classes.

Furthermore, the names of models and classes alone may not be sufficient to determine the design intent or operational rules specific to the target domain.

In such cases, we recommend creating and adding skills tailored to the target domain. Including package structure, model classification rules, and rules regarding model creation and modification within the target domain in the skills will make it easier for the AI ​​agent to search, reference, and edit models and classes in accordance with the design and operation of the target domain.

Creating Skills

When creating a skill, include the following information about the target domain:

• Overview of the Target Domain

• Overview of the target domain and design target

• Terminology, abbreviations, and naming conventions specific to the target domain

• Information for Manipulating Models

• Package structure and model classification rules within the target domain

• Purpose and design intent of classes and fields

• Input rules for fields and meaning of settings

• Starting point, search conditions, and search direction when searching for models

• Rules for setting up relationships and traces between models

• Methods for tracking requirements, functions, components, etc.

• Models that can be created and modified, and read-only models

• Models to be excluded and items that should not be modified

• Required items when creating or modifying models

• Verification Method After Work

• Fields and relationships to be verified

• Number of models created/modified and processing results

The content of the Skill should be organized as design and operational rules common to the target domain, and should be confirmed to be consistent with the profile and design model structure of the target project. If the design or operation of the target domain changes, the content of the Skill should also be updated.

Furthermore, it is recommended to include rules in the Skill that prevent the AI ​​agent from making inferences about matters not explicitly stated in the Skill or instructions, and instead present unclear points or matters requiring user judgment to the user.

Dividing Skills

When creating Skills, do not combine information for the entire target area into a single Skill. Instead, divide it by design target or work unit.

For example, divide it as follows:

• Requirements and Function Tracing
• Vehicle Control Component Configuration
• Safety Analysis Model Editing

Dividing the scope narrows the information the AI ​​agent references, making it easier to find the desired model or class.

Checking Skill Usage Results

Even when using Skills, do not rely solely on the AI ​​agent's completion report; verify the actual results on Next Design.

• Is only the models within the scope specified in the Skill being processed?
• Are the specified class, field, and association rules being applied?
• Are there any missing required fields or association settings?
• Have the models excluded from the Skill been changed?
• Are the number of processed, unprocessed, and duplicate items as expected?

How to Review When You're Having Trouble

We will explain basic troubleshooting methods and situation-specific troubleshooting methods.

Basic Troubleshooting Methods

• If unintended changes occur, use Undo (Ctrl+Z) to revert to the previous state.

• If there are any omissions, request additional information to complete them.

Response Methods Depending on the Situation

Situation	Points to Review
Unexpected models were processed	Add the target ID or path, parent model, excluded range, and search direction/depth
Some required fields or relationships are not set	Provide a checklist of required items and confirm the setting results and number for each target
The diagram structure or display differs from expectations	Clearly specify the classes/fields to be used, the diagram's range, the information to be included in the elements, and the display method
The result is empty or the number of results differs from expectations	Confirm the target range and relationship types, and reconfirm with detailed model information and individual targets
The AI ​​agent's report does not match the actual model	Record the AI ​​agent's interpretation and the field values, relationships, and numbers confirmed in Next Design separately
Necessary conditions are missing	Do not fill in with assumptions; organize unconfirmed items and items requiring user judgment
Unable to process a large number of targets	Divide the request into target units, organize the processing status of each target, and then request the next processing

Checklist

The following are points to check before and after using Next Design MCP.

• Have any models or relationships outside the target scope been processed?
• Have you checked the number and records of processed, unprocessed, and duplicate items within the target scope?
• Are the field values, relationships, and generated number of items before and after the change as expected?
• Are the class and field values ​​of the created/modified model as expected?
• Are the relationship type, direction, and search scope as expected?
• If you operated the screen, are the elements and display state on the view as expected?
• Have you assumed that everything has been processed based solely on the AI ​​agent's completion report or results such as "0 items" or "empty"?
• If there are unknown conditions, have you instructed the AI ​​agent not to fill in the gaps with inferences, but to present them as unconfirmed items or items requiring user judgment?
• If you generated a class diagram for the profile, are the scope, division unit, and number of elements as expected?
• Have you considered whether creating/adding Skills according to the target area is useful?