This project started with a conversation with a friend who is passionate about knowledge building. She had been talking with her friend about how AI conversations might help people develop their understanding, and shared the idea with me.
The idea resonated with me. I had noticed similar difficulties in my own AI conversations and had been considering a small weekend prototype. In a long response, I often have a question about, say a small portion of the paragraph: a term is unfamiliar, an example deserves more attention, or I want to check the reasoning behind a claim.
I can ask about it in the same chat, but the original question gradually moves further up the page. Opening another chat means bringing the context over myself. Sometimes I leave the question alone and lose track of it later.
I wanted to try a small interaction around this: follow a side question while keeping the main conversation accessible.
That became Unfold, a preliminary v0 prototype. The name reflects the idea of letting a thought unfold into new questions while keeping its original context. I built it to make the idea concrete and see how it feels to use. Whether it helps people develop a better understanding is still an open question.
A useful analogy is following a link while reading Wikipedia: a term opens up another question. I want that branch to keep its connection to the original conversation.
A question beside the conversation
The first interaction is simple. Select a passage, choose Ask separately, and ask a question about it.
A card opens in a side panel called Thoughts. It contains the selected passage, the question, and the model's response. I can continue asking there, return to the original passage, or quote something from the exploration back into the main chat.
The side question retains the relevant conversation leading up to its source. I can also attach more than one passage to an exploration when I want to ask about their relationship.
The basic loop is: notice something, ask about it, and bring what seems useful back into the larger inquiry.
Shaping the surface
I started with a main chat and a small side panel. I wanted organization to come from the questions I was already asking. Selecting a passage creates a Thought, so the questions begin to give the collection its structure.
Keeping the main chat in place became an important detail. Opening Thoughts does not shift the reading column on desktop. The cards scroll independently, and only one expands at a time. That gives a side question some room without making every branch compete for attention.
The connection to the source also needed to stay visible. Selected passages have a light yellow highlight, even when their Thought is collapsed. Clicking a highlight opens the corresponding card. From the card, Go to source takes me back to the passage.
Underneath that interaction, each reference identifies a specific message and an exact passage. Matching the words alone would be ambiguous when the same phrase appears in several replies, or appears both in Chat and in a Thought.
Finally, I wanted a way back into the conversation. Quote in chat attaches an exploration reply to the main composer and preserves the question I am already writing. The quote keeps its connection to the original exploration.
What v0 includes
The current version has saved conversations, linked Thought cards, follow-up questions, streaming replies, and web citations. It also supports Markdown and mathematical notation. On a smaller screen, Chat and Thoughts become separate views.
This is an early surface to experiment with. Shared workspaces, richer notes, and cross-conversation connections are not implemented. I also have not tested whether these interactions improve learning.
Looking back on the inquiry
One thing I want to explore next is a wrap-up or reflection action.
After following a few questions, I might want to look back: What was I trying to understand? Which explanation changed my thinking? What remains unclear? What would I investigate next?
I'm thinking about a Reflect action alongside Thoughts. It would produce a draft connected to the relevant messages, with room for the user to describe what changed in their own understanding. It could be used midway through a conversation as well as at the end. A model can summarize a conversation, but the user still needs to judge what they understand and what they want to do with it.
This is still a preliminary v0, but I’d love to see where it goes. It could become part of a research project, grow into a tool people find useful, or develop through both. For now, I’m interested in using it, sharing the experiment, and seeing what we learn along the way.