Vision Design Lab
A page that compiles is not proof that the design works.
Coding agents can build websites quickly. They can produce valid components, responsive layouts, and polished interfaces that look like something a modern startup might ship.
They still have very little understanding of what they created.
The Vision Design Lab is testing whether thoughtful web design requires a visual reasoning system that can understand geometry, hierarchy, balance, and flow before an agent begins writing code.
We are asking language models to solve a visual problem.
Most AI web-design systems begin with language.
The user describes a page. The model predicts code. A browser renders the result. The agent may inspect an error log or receive a screenshot, then generate more code.
That workflow can determine whether the page functions technically. It does not mean the system understands the page visually.
The agent may recognize that a button exists without understanding whether it competes with the headline. It may see that two sections are present without understanding that their spacing makes them feel unrelated. It may reproduce the general style of a reference while missing the geometry that gives the design its structure.
The system knows how websites are commonly written. That is different from knowing why a particular arrangement works.
Design is geometry before it is decoration.
Thoughtful design depends on relationships.
How large is one element compared with another? Where does the eye move first? Which objects belong together? Where does the page create tension, rest, momentum, or confusion?
These questions exist before color, gradients, shadows, and animation.
A page can contain attractive components and still fail because the geometry between them is wrong. The headline may be too weak for the space around it. The content width may fight the reading rhythm. A card grid may be technically aligned while creating no meaningful hierarchy.
That is why generated websites often feel strangely hollow.
The page contains the objects associated with good design, but the relationships between those objects have not been understood.
The same patterns keep returning.
More instructions can improve the surface.
You can give the agent a design system, reference sites, component rules, spacing tokens, typography scales, and examples of layouts you like. You can tell it to avoid excessive gradients, crowded cards, giant rounded rectangles, and the same familiar startup aesthetic.
The output may improve.
Then the model slowly returns to what it already knows how to predict.
You get the oversized hero, glowing background, dashboard mockup, three cards, logo strip, and a set of decorative elements that communicate “technology” without helping the page communicate its actual idea.
The problem is not that the agent ignored every instruction.
The problem is that imitation remains easier than visual judgment.
A screenshot does not automatically create sight.
Giving an agent a rendered screenshot sounds like it should solve the problem.
Sometimes it helps the model notice obvious issues. It may identify text that overlaps, a component that is cut off, or a section that differs substantially from a reference.
That is still far from understanding the design.
The model may describe what is visible without correctly reasoning about why the layout feels weak. It can suggest increasing spacing when the real problem is grouping, or changing font size when the underlying problem is hierarchy.
The screenshot becomes another input to interpret through language.
The Vision Design Lab is testing whether visual feedback needs its own architecture rather than being reduced immediately to a written description.
The hypothesis
The central hypothesis is that design judgment and code generation should be separate systems.
A coding model is useful once the desired structure is understood. It can implement components, manage responsive behavior, connect interactions, and revise the underlying code.
It should not be trusted to invent the visual logic and judge its own implementation using the same language-first process.
The Lab is exploring a pipeline in which visual systems first reason about the page as space.
That means understanding where objects are, how they relate, what attracts attention, and whether the rendered result matches the intended hierarchy.
Only then should code generation decide how to implement the correction.
Spatial intelligence may be the missing layer.
Some of the most important work in AI is focused on helping models understand physical and simulated worlds.
Robots need to reason about distance, shape, orientation, movement, obstruction, and the relationships between objects. World models need to preserve geometry and consistency as environments change.
A website is not a three-dimensional world, but many of the underlying questions are related.
The system still has to understand:
- Where elements exist relative to one another.
- Which relationships create hierarchy or grouping.
- How the composition changes across space and screen size.
The Vision Design Lab is exploring whether research built for spatial intelligence can be adapted to interfaces.
The goal is not to turn a website into a robotics problem.
It is to stop pretending that a language model’s familiarity with frontend code gives it visual understanding.
The experimental pipeline
01
Intent
02
Spatial plan
03
Code
04
Render
05
Vision review
06
Correction
Define the purpose
Before designing the page, establish what the person should understand, where their attention should move, and which action the page should support.
Create a spatial plan
Represent the intended hierarchy, grouping, density, proportions, and relationships before committing to specific components.
Generate the implementation
Use coding agents to translate that plan into working components and responsive behavior.
Inspect the rendered result
Evaluate the actual pixels produced by the browser rather than assuming the code created the intended design.
Reason about the geometry
Identify where alignment, hierarchy, balance, density, or visual flow diverged from the plan.
Revise and verify
Change the implementation, render it again, and determine whether the correction produced a measurable improvement.
The page becomes something the system observes and tests, not merely code it generated successfully.
Visual judgment needs structure.
“Make it look better” is not a useful evaluation standard.
It gives the model no stable definition of success and makes comparisons depend on taste that has never been explained.
The Lab is developing ways to examine a design across clearer dimensions.
Hierarchy
Can a person tell what matters first, second, and third without having to inspect every element?
Geometry
Do size, spacing, alignment, containment, and proportion create the intended relationships?
Flow
Does the composition help the person move through the information and understand what to do next?
Aesthetic quality still matters, but it has to sit on top of structure.
A beautifully styled page with broken hierarchy is still a broken design.
Taste cannot be reduced to a style prompt.
A prompt can describe an aesthetic.
It can ask for editorial restraint, industrial simplicity, playful illustration, or the visual language of a particular era.
That gives the model a direction. It does not give the system taste.
Taste includes knowing which idea deserves emphasis, when an element adds meaning, and when the composition has already said enough. It requires restraint as often as invention.
A model may reproduce the visible habits of a style while missing the judgment beneath them.
The Vision Design Lab is not trying to produce one universal definition of beautiful design. It is testing whether a system can understand the purpose and internal logic of a visual direction well enough to apply it consistently.
References should teach relationships, not just appearance.
Designers use references to understand more than color and typography.
They study how a page controls scale, where it creates contrast, how it groups ideas, and why a particular composition feels clear or memorable.
Generated systems often use references as skins.
They borrow the dark background, typeface, border treatment, or animation style while rebuilding the same underlying template they always produce.
The Lab is exploring how to extract the spatial principles from a reference without copying its surface.
What makes the hero feel confident?
Why does one layout feel dense without becoming cluttered?
How does the composition direct attention before the reader understands the words?
The useful part of the reference is the judgment encoded in those relationships.
Responsive design is part of the reasoning problem.
A desktop layout is not a fixed composition that can simply be squeezed onto a smaller screen.
Relationships change.
A balanced horizontal section may need to become a deliberate sequence. Supporting information may need to move closer to the idea it explains. An element that anchors the desktop design may overwhelm the mobile version.
Coding agents can apply breakpoints and stack columns.
The harder question is whether the hierarchy still survives.
The Vision Design Lab will judge responsive states as distinct compositions connected by the same purpose, not as technical variations of one desktop layout.
What would prove the hypothesis wrong?
The Lab should not assume that a new visual architecture is necessary merely because current outputs are frustrating.
A well-designed language-first agent may eventually learn enough from rendered feedback, stronger multimodal models, and better evaluation to solve the problem without a separate spatial system.
The hypothesis should weaken if the vision-first pipeline fails to produce consistent improvements, if human reviewers cannot distinguish its output, or if the additional complexity only improves a narrow set of designs.
It should also weaken if the system becomes excellent at satisfying geometric metrics while producing lifeless, mechanical work.
A design can be aligned, balanced, and completely forgettable.
The research has to improve the whole result, not only what is easiest to measure.
What the Lab will publish
The research should make the difference visible.
That means publishing rendered comparisons, spatial analyses, evaluation criteria, failed layouts, correction attempts, and examples where the vision system misunderstood the design.
The Lab should also show where rules and skills stop helping.
If a design system, screenshot, or detailed prompt improves the result only slightly, the comparison should make that limit clear.
The claim has to survive inspection at the pixel level.
The connection to The Adam Repo and Arthor
The Adam Repo can help agents plan, build, test, and ship software.
The Vision Design Lab extends the meaning of verification.
A website should not be considered complete merely because the tests pass and the page renders without errors. The system also needs evidence that the interface is readable, coherent, and visually faithful to its purpose.
Arthor has the same need from a different direction.
A living website may understand the business and know what a page needs to communicate. That knowledge still has to become a design people can understand and trust.
The Vision Design Lab is where the visual judgment required by both systems can be tested before it becomes a product claim.
FAQ
Frequently asked questions
Is the Vision Design Lab an AI website builder?
No.
The Lab is researching how AI systems can understand and evaluate design. Website generation is one place that research can be applied.
Why not just use a better design system?
A design system provides useful constraints, components, and patterns.
It does not determine the correct hierarchy, composition, or visual relationship for every page.
Can current multimodal models see websites?
They can interpret screenshots and identify many visible elements.
The research question is whether they can reason about the geometry and purpose of the design well enough to improve it consistently.
Is design quality objective?
Some parts are subjective.
Other parts, including alignment, grouping, contrast, legibility, hierarchy, and responsive integrity, can be evaluated with greater structure. The Lab is studying both measurable properties and informed human judgment.
Will the system generate code?
Code generation is part of the pipeline.
The Lab’s primary focus is the visual reasoning that should guide and evaluate that code.
AI will not become a good designer by producing more pages.
It has to understand the relationships that make a page work.
That means treating layout as geometry, design as judgment, and the rendered interface as something the system must see and evaluate rather than merely generate.
The Vision Design Lab is testing whether that missing layer can be built.
Follow the research.
Get updates when experiments ship, fail, or change what we believe about the architecture.