Independent R&D
What if users don’t have to choose colors?
An exploration of how AI can turn human visual intent into an adaptive color system - without exposing the complexity behind it.
The problem
Traditional design systems put the complexity on designers.
In a traditional system, designers manually choose palettes, build scales, define token groups, map semantic roles, define component states, and check contrast - for every brand, every theme.
01Choose palette
→
02Build scales
→
03Define tokens
→
04Map roles
→
05Define states
→
06Check contrast
→
07Apply to components
Users shouldn’t have to think like designers.
Brand A
Purple · light environment
Brand B
Green · green environment
Brand C
Orange · dark environment
Brand D
Enhanced accessibility
How can we expose a few decisions users actually understand, while hiding hundreds of decisions the system can calculate?
The experiment
Start with intent, not tokens.
These aren’t tokens the user is setting. They’re constraints - a small amount of intent the engine has to satisfy.
Illustrative - full working version further down
#6C72E0
#F4F3F7
Contrast
Accessibility
Visual temperature
User intent
→
Color engine
→
Color structure
Understanding color
The dimensions the engine actually needs.
H
Hue
Where the color sits on the color wheel.
L
Lightness
How light or dark it appears.
C
Chroma
How colorful or muted it is.
T
Temperature
A contextual warm/cool read, not a primary dimension.
⇄
Contrast relationship
Contrast belongs to the relationship between two colors - not to either one alone.
Interactive - try it
HEX #6C72E0
The color model
Not accent + neutral + semantic. Something broader.
That model is too restrictive. The system is built from two kinds of roles instead.
Dynamic color roles
Adapt strongly to user intent - expression changes.
Accent
Background
Surface
Content
Border
Overlay
Interaction
Semantic color roles
Adapt their values and shades - meaning stays stable.
Success → green family
Warning → amber family
Error → red family
Info → blue family
Dynamic colors can change their expression.
Semantic colors can only change their resolution.
Semantic colors can only change their resolution.
Accent normalization
Any input → a stable color anchor.
A user enters #5E6AD2. The engine doesn’t assume that’s Accent/500 - it treats the input as intent, analyzes its perceptual properties, and decides where it should sit within a generated scale.
User color
→
Analyze
→
Normalize
→
Generate scale
The user chooses the color. The system decides how it behaves.
Environment detection
Background is more than a color.
The engine first determines whether the chosen environment reads as light or dark - from perceptual lightness, not a checkbox.
Background
→
Perceptual analysisLightness · contrast range · color character
→
Light / dark environment
Environment doesn’t have to be neutral. It can follow the user’s visual intent instead of forcing every product into the same predefined neutral palette.
Computed example - green environment
The values adapt. The hierarchy doesn’t.
Generating color roles
One intent → a network of relationships.
Accent→Attention + interaction
Surface→Depth + layering
Content→Readability + hierarchy
Border→Separation + structure
Semantic→Meaning + feedback
Color relationships
The engine doesn’t ask: is this color good?
It asks whether a color works in its context. Hover a group below - colors are calculated as relationships, not isolated values.
Contrast
Same colors. Different character.
The same intent, expressed with different hierarchy strength, surface separation, accent dominance and density. Accessibility is a constraint underneath, not a style choice.
Project overview
A live comparison of the same UI intent under three contrast settings.
User preference
→
Generate candidates
→
Evaluate contrast
→
Accessibility constraints
→
Reject invalid
→
Optimize remaining
Visual preference is flexible. Minimum usability is not.
Visual harmony
AI optimizes the system, not individual colors.
There is no single correct color. There is a better solution within a defined set of constraints - never a magic beauty score.
Best-fit system
Resolved from the candidate field above, within the defined constraints.
Semantic resolution
Meaning stays stable. Values adapt.
Success is never red - it always resolves within the green family. What changes across environments is the shade, not the meaning.
Light environment
Dark environment
From values to rules
AI needs more than hex values.
Accent / 500
#5E6AD2
Role
Primary interaction
Used for
Buttons, links, active states
Context
Interactive elements
Constraints
Maintain sufficient contrast
Do not use
Destructive feedback
Large surfaces
Value+
Meaning+
Context+
Usage+
Constraints
=
AI-readable system
The solution · Live demo
Give users intent-level controls.
The output isn’t a palette - it’s a system. Change any input below and watch the whole structure recalculate, live, in your browser.
Generated & resolved
- environment: light
#5E6AD2
#F7F6FA
Contrast
Accessibility
✓Analyzing accent
✓Detecting environment
✓Generating relationships
✓Checking contrast
✓Resolving semantic colors
✓Optimizing harmony
Generated scale - accent
Dynamic roles
Semantic roles
Applied to UI
Time to hire
18.4 days avg
Offer acceptance
72% this quarter
Success
Warning
| Role | Candidates | Status |
|---|---|---|
| Design Architect | 24 | In review |
| Staff Engineer | 16 | Blocked |
| Product Manager | 31 | Offer sent |
Change the intent. The system recalculates the experience.
Impact
What changes when color becomes computed.
Traditional
User
↓ choose brand color
↓ fixed palette
↓ manual mapping
Adaptive
User
↓ define intent
↓ AI understands context
↓ AI calculates relationships
↓ system generates structure
↓ UI adapts
For users
- Customize without design expertise
- Express brand identity
- Support visual preferences
- Better accessibility
For designers
- Define rules instead of repetitive values
- Encode design knowledge
- Reduce manual theming
- Create adaptive systems
For AI
- Understand roles
- Understand relationships
- Reason within constraints
- Generate valid UI
Key learnings
What this research actually taught me.
01
A color system is not a palette.
A palette describes what colors exist. A system describes why they exist, how they relate, and where they should be used.
02
Users should configure intent, not tokens.
Designers think in tokens, roles, relationships, constraints. Users think in brand, environment, visual preference, accessibility. AI bridges the two.
03
Environment doesn’t have to be neutral.
It can be neutral, green-tinted, blue-tinted, warm, purple, dark, or light. What matters is coherence.
04
Visual harmony is a constraint problem.
Intent, plus perceptual properties, plus relationships, plus context, plus constraints - resolves to a better solution.
Transition
Color is only the first layer.
Part 1 explores how AI can make visual decisions from human intent. But an AI-native product needs to solve a larger problem: what happens when AI doesn’t just choose a color - but decides what UI to build?
>_
Create a hiring report showing time-to-hire, offer acceptance rate and monthly hiring trends.
Prompt
→
Product knowledge
→
Components
→
Patterns
→
Business logic
→
Permissions
→
Consistent experience
Part 02
Adaptive Product UI
From adaptive color to adaptive product decisions.