A tiny tool for thinking and reasoning in color — and a first experiment toward VRGB.
RGB-LLM is a minimal interface for blending three poles of a problem into a single RGB vector. Slide → blend → interpret the resulting color → export a JSON signal.
It's not a framework or a system. It's a calibration layer — a lightweight tool for reasoning about tradeoffs, direction, and emphasis.
Use it for:
- creative prompts
- lightweight state cues
- multi-step workflows
- decision framing
- metadata for demos (e.g., Magic Fridge)
VRGB is the conceptual layer behind this repo —
a low-dimensional vector space (color) used as a small shared signal surface for AI workflows.
RGB-LLM is the first UI for experimenting with that idea.
The motivating question:
Can a tiny, bounded vector — literally R/G/B — carry useful micro-intent between AI modules without schema overhead or complex preference tables?
This repo exists so I can explore that publicly.
In multi-step AI workflows, the missing layer is often the simplest one: the small signals that shape behavior between tools.
Signals like:
- "be more strict"
- "lean creative"
- "reduce detail"
- "keep tone consistent"
VRGB proposes a weirdly simple approach: a three-number vector that encodes these shifts continuously.
RGB-LLM is a tiny interface for generating that vector.
High-dimensional embeddings are great, but:
- expensive to serialize
- hard to inspect
- brittle across modules
- difficult to align
- not human-interpretable
VRGB flips the model:
- bounded (0–1 floats)
- continuous gradients
- cheap to transmit
- easy to visualize
- vector-operable
- stable under drift
It won't replace embeddings — it complements them for the small, continuous cues that keep tools coherent.
This is a "what if?" experiment:
What if 3D color space is the right dimensionality for micro-intent?
RGB-LLM is how I explore that.
[ Your Poles ]
R G B
↓ ↓ ↓
[ RGB-LLM UI ] ← adjust sliders
↓
Color Swatch ← intuitive feedback
↓
{ "r": 0.42, "g": 0.79, "b": 0.12 } ← VRGB vector
↓
Downstream prompt / model / workflow
A tiny vector carrying directional meaning.
Just open index.html in a browser. No build step, no dependencies.
Adjust sliders, edit pole labels, and copy the resulting JSON.
Two modes:
Create three custom axes to map your thought space. Each axis captures a dimension with two opposing poles and a slider to position yourself along that spectrum.
Click any pole label to customize it. The JSON output captures your pole definitions and current positions.
Build structured prompts with fill-in-the-blank templates plus adjustable style axes. Great for creative writing or prompt engineering.
RGB-LLM is released under the MIT License. You are free to use, fork, modify, and build on this work.
If you explore the VRGB concept or build something interesting with this repo, I'd love to hear what you discover.
Not for credit or control — just genuine curiosity about how others think about low-dimensional reasoning.
Feel free to open an issue, share a link, or reach out anytime.
GitHub Issues • LinkedIn Happy to collaborate, learn, and exchange ideas.