About:
One brain, one fact, one conversation.
A network of neurons spins, slows down and stops. Then a fact appears, with its source, a label for how strong the evidence is, and an AI you can ask about it.
Category:
Design concept
Year:
2026
Design framework:
Evidence-graded content design
Tools:
Framer, React, Canvas, Vercel, Gemini
Spin
Stop
Fact
Ask
The problem:
Brain facts spread faster than the studies behind them. They lose their sources and caveats on the way, and some of the most quoted ones are simply myths.
Goal:
Design a calm, one-fact-at-a-time experience where every claim carries its source and evidence status, and an AI helps you go deeper without overstating the science.
My role:
My own concept. Idea, content design of 846 facts, conversation design and AI guardrails, interaction design and the creative coding of the neuron network.
Methods:
•
Concept design
•
Content design
•
Conversation design
•
Interaction design
•
Creative coding
•
Prompt design
Discover
Popular neuroscience travels fast and loses its caveats on the way. I wanted to slow it down.
/01
The idea
This was my idea: a brain that spins, slows down and stops, and then hands you one fact you can trust and question.
Brain facts are everywhere, in feeds, talks and office posters. Most of them arrive without a source, and many of them are myths. I wanted a small ritual instead of another scroll: one fact at a time, and the space to ask about it.
The interesting design problem was not how to show more facts, but how to make one fact carry everything you need to judge it: where it comes from, how strong the evidence is, and where it stops being true.
/02
Myths travel faster than studies
The most quoted brain facts are often the least true. The ten per cent myth and the left and right brain split survive because they are simple, not because they hold up. Even the famous hundred billion neurons was a rounded guess until someone actually counted.
That last one is my favourite example, because it isn't a myth at all, just a number that lost its caveat. When researchers counted cell nuclei in dissolved brain tissue, they arrived at roughly 86 billion. The claim was close. What was missing was the how-do-we-know.
So the product couldn't just debunk myths. It had to show the evidence behind true claims too, in a way that makes the difference visible.
/03
What a fact needs
A fact is not just a sentence. To be useful, each one needs a source, a plain account of how it was found out, what it means and, above all, what to be careful about. That last part is what headlines usually leave out.
I turned this into a fixed anatomy that every one of the 846 facts follows: the claim, a short answer, an evidence status, the source, how it was found out, and what it means together with what to be careful about.
A fixed structure was a deliberate constraint. It makes facts comparable, it makes gaps obvious while writing, and it gives the AI a predictable shape to read from.
Define
Every fact carries its evidence with it, and the AI is held to the same standard.
/01
Evidence, labelled
I wrote and checked 846 facts. Each one has one of three labels, so you can see at a glance how much weight it carries.
529 facts, 63 per cent, are well established. 224, or 26 per cent, are debated in the field. 93, 11 per cent, rest mostly on animal or small-sample evidence.
A third of the deck is openly debated or based on small samples, and I kept it in on purpose: showing uncertainty is part of being honest about science. A deck of only rock-solid facts would look more trustworthy, and be less true.
There are three plain labels and no score. A number like 7 out of 10 would suggest a precision the evidence doesn't have.
/02
Principles
Four principles shaped every decision.
One fact at a time. No feed, no infinite scroll. The calm is part of the product, not decoration.
A source on every claim. Author, year and journal, linked.
Say where it is unsure. The label and the caution are as visible as the claim itself.
Never about your brain. It explains research, it does not diagnose.
/03
Guardrails for the AI
The chat runs on a short list of rules that is rebuilt for every fact, so the AI always knows which fact you are reading, its source and its label. It is designed to say I do not know rather than fill the gap.
There are six rules. Only well-supported findings: author and year only when it is sure, and never invented citations, numbers, journals or links. Say how strong it is: keep apart what is established, what is debated and what was only shown in animals or small samples. Never about your brain: no diagnosis and no treatment advice, and health worries are pointed to a doctor. Stay on the fact: neuroscience and cognitive science only, and off-topic questions are gently steered back. Short and plain: two to five sentences in British English, no markdown and no praise for the question. Keep the talk going: it can end with one short question, but only when it makes sense.
Writing what the AI must not do was as much interface design as anything on screen. A friendly chat that confidently invents a citation would undo everything the labels are trying to do.
Develop
From a spinning network to a conversation, in three moments.
/01
A brain made of neurons
The opening is drawn in code. 520 neurons are scattered inside brain-shaped volumes and linked to their nearest neighbours, with signals running along the links. The brain spins, slows down and stops, and only then does the fact appear.
Each neuron connects to its four nearest neighbours, so it reads as a network rather than a cloud. The spin lasts 6.5 seconds before the brain slows and stops, long enough to feel like a small ritual, short enough not to be a loading screen.
The layout uses a fixed seed, 86, so it is always the same brain, an object you come back to rather than a new random pattern each time.
/02
The fact window
Every fact opens with Did you know that, typed in letter by letter. The short answer, the label and the source follow, and the details fade in last, so the reading order matches the importance.
This is the content hierarchy made visible. If you read only the first line, you still get the claim and its label. If you stay, you get the source, how it was found out and what to be careful about.
/03
Ask anything
Below each fact there is a chat. Three suggested questions are written for that specific fact, so nobody faces an empty box. Answers stream in, can be stopped at any time and stay on the fact.
An empty chat box asks people to know what to ask. The suggested questions do that work for them, and they also show what kind of questions the AI is good at: how was this found out, how sure are we, what does it mean in everyday life.
/04
For everyone
Skip jumps straight to the fact. Reduced motion shows a still brain and the full text at once. Keyboard first. The fact is a real dialog, Escape closes it and focus returns where you were. Show fact brings the window back after you close it.
An animation-led concept is easy to make inaccessible. Every moment of motion has a way around it, and nothing important exists only in the animation.
Deliver
A live site with 846 facts and an AI that knows when to say it is not sure.
/01
Three facts, three labels
The label is the first thing you see after the claim. Its style changes with the strength of the evidence: outlined, filled or dashed, so the difference reads even without reading the words.
Well established is an outlined label. Debated in the field is filled. Mostly animal evidence is dashed. The three styles stay legible in greyscale and don't rely on colour alone.
/02
Under the hood
The page is a single Framer code component. The facts live as a JSON file on Vercel, and the chat goes through a small Vercel function, so the AI key never touches the browser. The function checks the origin, caps the length of the conversation and streams the reply. The model is Gemini Flash-Lite, with a system prompt built for the one fact on screen.
Allowed origins only, a maximum of 24 messages, 1,200 characters per message, and the same fact order in the data and the prompt. These limits keep the chat cheap to run and hard to misuse, and they mean the AI always talks about the fact you are actually looking at.
Outcome
/01
What I learned
Designing for trust is mostly content design. The hardest work was not the animation or the chat, but writing 846 facts so that the caveat is as clear as the claim, and keeping the AI as careful as the text.
/02
Next steps
Test with readers whether the labels change how much they trust a fact. That is the core assumption of the whole concept, and it hasn't been tested yet.
Let readers flag a fact and suggest a better source.
Add a Czech version of the deck.
© More works
(WDX®)
digital design
© help centre
(WDX)
clarification
Let’s talk