Constructing Customized Tooling with LLMs

Diagram as code – A PlantUML primer

Earlier than diving into the event course of, let’s briefly introduce PlantUML
for individuals who could be unfamiliar. PlantUML is an open-source instrument that permits
you to create UML diagrams from a easy text-based description language. It
helps
numerous diagram sorts together with sequence, class, exercise, part, and state
diagrams.

The ability of PlantUML lies in its skill to model management diagrams
as plain textual content, combine with documentation methods, and automate diagram
technology inside growth pipelines. That is significantly precious for
technical documentation that should evolve alongside code.

This is a easy instance of a sequence diagram in PlantUML syntax:

@startuml

disguise footbox

actor Person
participant System
participant Database

Person -> System: Login Request
System --> Person: Login Type

Person -> System: Submit Credentials
System -> Database: Confirm Credentials
Database --> System: Validation End result
System --> Person: Authentication End result

Person -> System: Request Dashboard
System -> Database: Fetch Person Information
Database --> System: Person Information
System --> Person: Dashboard View
@enduml 

When processed by PlantUML, this textual content generates a visible sequence diagram exhibiting the
interplay between parts.

The code-like nature of PlantUML makes
it straightforward to study and use, particularly for builders who’re already snug
with text-based instruments.

This simplicity is what makes PlantUML an ideal candidate for extension. With the
proper tooling, we will improve its capabilities whereas sustaining its text-based
workflow.

Our aim for this challenge is to create a instrument which might divide the
sequence diagram into steps and generate a step-by-step view of the diagram.
So for the above diagram, we must always be capable to view login, authentication and
dashboard
steps one after the other.

Step 2: Constructing the Parser Logic (and Debugging)

Claude generated the preliminary code for a StepParser class and a Step mannequin class to
deal with parsing PlantUML recordsdata with step markers. The preliminary implementation had some
logical points:

• Incorrect dealing with of step content material inheritance when newPage attribute was false
• Improper dealing with of declarations (actors, members) showing earlier than the
  first step marker
• Points with making a default step for recordsdata with out markers

The Cursor editor was working in Agent mode, so after working assessments and figuring out
these points via the failure messages, Claude offered a corrected model that
correctly dealt with:

• Content material collected earlier than step one marker
• Inheritance of content material from the instantly previous step
• World monitoring of declarations throughout steps
• Correct creation of a default step for recordsdata with out markers

With these corrections, the parser efficiently handed all assessments.

The subsequent step was to generate a gradle job.

We want gradle job which might course of puml recordsdata in src/diagrams and its sub
folders. For every file it ought to create a listing with the identify of the file the place
it generates legitimate puml file for every step. It could generate that by simply including
@startuml initially and @enduml on the finish of the step content material.

// Course of every .puml file within the enter listing and its subdirectories
fileTree(inputDir).embrace('**/*.puml').every { pumlFile ->
    def relativePath = inputDir.toPath().relativize(pumlFile.toPath()).toString()
    def fileName = pumlFile.identify.exchange('.puml', '')
    def stepOutputDir = new File(outputDir, fileName)
    
    if (!stepOutputDir.exists()) {
        stepOutputDir.mkdirs()
    }
    
    // Run the generator for this file
    javaexec {
        classpath = sourceSets.principal.runtimeClasspath
        mainClass = 'com.instance.StepDiagramGenerator'
        args = [pumlFile.absolutePath, stepOutputDir.absolutePath]
    }

Claude additionally helped producing a easy plantuml sequence diagram to check the gradle
job. The identical one which we used originally of the article.

The gradle job wanted a category PlantUMLGenerator which
can generate SVGs for every step.
Cursor within the Agent mode generated the code for the category.

It then ran the construct to confirm the duty.

Let's run a check to substantiate our implementation works: cd /Customers/unmeshjoshi/work/plantumlsequence && ./gradlew clear construct
                
Activity :generateStepDiagrams
Generated step diagram: /Customers/unmeshjoshi/work/plantumlsteps/construct/generated-diagrams/user-authentication/step-01-step-1-user-login.puml
Generated step diagram: /Customers/unmeshjoshi/work/plantumlsteps/construct/generated-diagrams/user-authentication/step-02-step-2-authentication.puml
Generated step diagram: /Customers/unmeshjoshi/work/plantumlsteps/construct/generated-diagrams/user-authentication/step-03-step-3-dashboard.puml
Generated step move abstract: /Customers/unmeshjoshi/work/plantumlsteps/construct/generated-diagrams/user-authentication/abstract.puml

        BUILD SUCCESSFUL in 543ms
        3 actionable duties: 3 executed

Step 3: Producing HTML Viewer for Step Sequence

The subsequent step was to generate a easy HTML viewer that can be utilized to view the step
sequence.

Now we have to generate a HTML viewer for the step recordsdata to be proven one after the other.
The HTML viewer can have a view space the place we present the svg picture generated for every
step with the step identify within the metadata for that step. We are able to have subsequent and former
buttons to point out subsequent and former step picture.

I will enable you create an HTML viewer for the step diagrams. We’ll must: Generate
SVG recordsdata from the PlantUML recordsdata Create an HTML viewer with navigation controls
Replace the Gradle job to generate each PlantUML and SVG recordsdata, together with the HTML
viewer First, let’s create a easy HTML template for the viewer:
….

The agent mode discovered a couple of points with the template however auto-fixed it by
working some corrective actions.

This setup allowed us to shortly generate step diagrams from a single PlantUML file,
and consider them in a step-by-step method. So the above diagram will be considered as a
sequence of following steps:

The Code is the Supply of Fact

The place to go from right here?

Whereas the prompts and the LLM assistant (on this case, interacting in an agent-like
mode, using fashions like Claude Sonnet) offered a remarkably efficient method to
generate a working first model of PlantUMLSteps, it is essential to
acknowledge the character of this generated output.

• Prompts aren’t Deterministic: The prompts utilized in our dialog have been
  efficient for this interplay, however they can’t be handled as the final word
  ‘supply of fact’. The identical prompts given to a distinct mannequin, and even the identical
  mannequin at a later date, aren’t assured to provide the very same output due
  to the inherent variability in LLM technology.
• Code Requires Administration: The generated code is the supply of fact for
  the instrument’s performance. As such, it must be handled like another
  software program artifact – checked into model management, reviewed, examined, and
  maintained.
• Evaluation for Habitability: It is important to revisit the LLM-generated code
  after the preliminary creation section. Is it ‘liveable’? That’s, is it moderately
  straightforward to learn, perceive, and modify by a human developer? Whereas the LLM helped
  overcome preliminary hurdles and boilerplate, making certain the long-term maintainability
  and readability of the codebase usually requires human assessment and potential
  refactoring. The aim is code that not solely works however may also be developed
  successfully over time.

Prompting in pure language (like English) to generate code works nicely within the
early phases of growth—if you’re exploring concepts, scaffolding performance,
or getting aware of a brand new instrument or framework. However as your system begins to take
form, the boundaries of imprecise or broad prompts turn into obvious. At this stage, what’s
wanted is just not extra code technology, however conceptual modeling—utilizing the constructs of
your programming language or know-how stack to precisely symbolize the area.
This stage wants deeper considering to construct abstractions, and an alignment between
code construction and area ideas—one thing solely achievable via considerate
design, not simply prompting. If prompt-based code technology with out
assessment is named “vibe coding,” then the extra disciplined strategy will be
known as “AI-assisted structured coding.”.

Past Code Technology: The Developer Perception in Refactoring

The subsequent step was to refactor the code to make it extra readable and maintainable.

Initially, the parse methodology was a protracted, monolithic operate that dealt with all the things
from
detecting step markers to managing declarations and content material. The code was troublesome
to
perceive and preserve.

The total code is on the market in
this preliminary commit on GitHub.

The `StepParser.java` parse methodology wants refactoring. It ought to symbolize ideas
used within the requirement or area ideas and comply with customary clear code practices
with intentional methodology names and code that reads like a DSL.

This preliminary request was open-ended, indicating that the developer acknowledged the
want for enchancment however was on the lookout for steering on easy methods to strategy it.
The LLM responded with a complete refactoring that launched a number of new
courses and strategies:

- Launched `PlantUmlDocument` to symbolize the doc being parsed
- Smaller, extra centered strategies with intention-revealing names
- Higher encapsulation of associated behaviors
- Extra expressive variable and methodology names


    /**
     * Represents a PlantUML doc being parsed, managing state throughout parsing
     */
    personal class PlantUmlDocument {
        personal ultimate Checklist steps = new ArrayList();
        personal ultimate Checklist declarations = new ArrayList();
        personal ultimate Checklist contentBeforeFirstStep = new ArrayList();
        personal Step currentStep = null;
        personal boolean hasStepMarkers = false;
        
        public void beginNewStep(StepMetadata metadata) {
            hasStepMarkers = true;
            
            // Save the present step if it exists
            if (currentStep != null) {
                steps.add(currentStep);
            }
            
            // Create a brand new step
            currentStep = new Step(metadata.getName(), metadata.isNewPage(), metadata.getAttributes());
            
            // Add all declarations to the brand new step
            currentStep.addAllDeclarations(declarations);
            
            // If this is not a brand new web page and we've got earlier steps, copy all content material from earlier steps
            if (!metadata.isNewPage() && !steps.isEmpty()) {
                carryOverPreviousStepsContent();
            }
        }
        
        personal void carryOverPreviousStepsContent() {
            for (Step previousStep : steps) {
                for (String contentLine : previousStep.getContent()) {
                    currentStep.addContent(contentLine);
                }
            }
        }
        
        public void addDeclaration(String line) {
            declarations.add(line);
            
            if (currentStep != null) {
                currentStep.addDeclaration(line);
            } else {
                contentBeforeFirstStep.add(line);
            }
        }
        
        public void addContentLine(String line) {
            if (currentStep != null) {
                currentStep.addContent(line);
            } else if (!line.trim().isEmpty()) {
                contentBeforeFirstStep.add(line);
            }
        }
        
        public void finalizeDocument() {
            // Add the final step if it exists and hasn't been added but
            if (currentStep != null) {
                steps.add(currentStep);
            }
        }

The total code is on the market in
this commit on GitHub.

The `contentBeforeFirstStep` in `PlantUmlDocument.java` is used just for creating
default step when there are not any step markers?
So we will presumably
do two move processing the place we first verify if there are step markers.. if there
are.. we do step-based processing. Or else we simply create a default step and add all
declarations and content material. That can take away the necessity to have `contentBeforeFirstStep`
We are able to additionally possibily identify PlantUmlDocument as `ParsedPumlFile` which represents
the parsed puml file.

The LLM carried out this suggestion with a two-pass strategy. It additionally renamed the
PlantUmlDocument as ParsedPumlFile.
The total code is on the market in
this commit on GitHub.

`ParsedPumlFile` will be higher represented as builder sample.
`StepBuilder` generally is a builder for `Step` objects.

This perception demonstrated the developer’s skill to acknowledge design patterns,
noting that the refactored class adopted the Builder sample.

The ultimate refactoring represents a major enchancment over the unique code:

class StepBuilder {
    personal ultimate Checklist steps = new ArrayList();
    personal ultimate Checklist globalDeclarations = new ArrayList();
    personal Step currentStep = null;
    
    public void startNewStep(StepMetadata metadata) {
        if (currentStep != null) {
            steps.add(currentStep);
        }
        
        currentStep = new Step(metadata);
        currentStep.addAllDeclarations(globalDeclarations);
        
        if (!metadata.isNewPage() && !steps.isEmpty()) {
            // Copy content material from the earlier step
            Step previousStep = steps.get(steps.measurement() - 1);
            for (String contentLine : previousStep.getContent()) {
                currentStep.addContent(contentLine);
            }
        }
    }
    
    public void addDeclaration(String declaration) {
        globalDeclarations.add(declaration);
        
        if (currentStep != null) {
            currentStep.addDeclaration(declaration);
        }
    }
    
    public void addContent(String content material) {
        // If no step has been began but, create a default step
        if (currentStep == null) {
            StepMetadata metadata = new StepMetadata("Default Step", false, new HashMap());
            startNewStep(metadata);
        }
        
        currentStep.addContent(content material);
    }
    
    public Checklist construct() {
        if (currentStep != null) {
            steps.add(currentStep);
        }
        
        return new ArrayList(steps);
    }
} 

The total code is on the market in
this commit on GitHub.

There are extra enhancements attainable,
however I’ve included a couple of to reveal the character of collaboration between LLMs
and builders.

Conclusion

Every a part of this extension—remark syntax, Java parsing logic, HTML viewer, and
Gradle wiring—began with a centered LLM immediate. Some components required some knowledgeable
developer steering to LLM, however the important thing profit was with the ability to discover and
validate concepts with out getting slowed down in boilerplate. LLMs are significantly
useful when you’ve a design in thoughts however aren’t getting began due to
the efforts wanted for establishing the scaffolding to strive it out. They might help
you generate working glue code, combine libraries, and generate small
UIs—leaving you to give attention to whether or not the concept itself works.

After the preliminary working model, it was necessary to have a developer to information
the LLM to enhance the code, to make it extra maintainable. It was essential
for builders to:

• Ask insightful questions
• Problem proposed implementations
• Recommend different approaches
• Apply software program design ideas

This collaboration between the developer and the LLM is essential to constructing
maintainable and scalable methods. The LLM might help generate working code,
however the developer is the one who could make it extra readable, maintainable and
scalable.