MENU
RABBIT HOLE
VISIT WEBSITE
YEAR
2025
ROLES
Designer
Developer
PROJECT SCOPE
Product Design
Frontend Development
AI Integration
DESIGN TOOLS
Figma
TECH STACK
Next.js
React Flow
Convex
Claude
OVERVIEW
A map of your own reasoning.
Rabbit Hole began as a build I made on my own for the Lovable competition, where it landed in the top 15. The idea stayed with me long after, so I went back and rebuilt its core into what it is now. The way it works: you talk through a problem you are actually stuck on, and while you talk, a second surface next to the chat draws out the structure of what you are saying. The claims you make, what they rest on, where they came from. What you get is a look at the shape of your own thinking, which turns out to be much harder to do on your own than it sounds.
THE PROBLEM
The hard part was never the ideas.
Coming up with ideas was never really the hard part. The hard part is seeing what the ideas you already have are sitting on. The assumption you never checked. The framework you picked up somewhere and never questioned. The thing you want to be true for reasons you would rather not say out loud. Most AI tools went the other way and raced to generate more of them. Almost none help you look harder at the ones already in your head.
THE GRAMMAR
Five kinds of thought, mapped live.
On the map, your thinking breaks into five kinds of thing. The claims you make. The assumptions sitting under them. And underneath those, the influences: something you lived through, an idea you picked up from somewhere, or a stake you are protecting. None of them are fixed. A reading shows up faint and unsure, gets more solid as you back it up, fades when you say something that cuts against it, and drops away when it stops holding. I wanted the ones that die to stay visible too, because watching a reading you had fall apart is half of what makes it useful.
FALSIFIABILITY
It has to be able to be wrong.
The risk with a tool like this is that it becomes a horoscope. It says something about you that sounds deep, you nod along, and none of it was ever something you could check. So I gave it one rule it cannot break. Every guess the map makes about you has to point back to the exact words you said that it came from, and carry the one question that would settle it either way. And it never states these things as facts. Each one is marked as a reading of what you said, a guess you are free to throw out. That rule is the whole difference between a tool you can trust and one that just flatters you.
THE AGENT
The agent pushes back.
The agent you are talking to can see the map too. Each turn, it looks for the assumption doing the most work on the least support, and finds a way to ask the question that would test it, once, without turning into an interrogation. It also notices when you contradict yourself. In the demo, you say a cofounder would slow you down, then a minute later that you would take one if they brought distribution. It catches that, names the tension, and asks what the no is actually protecting. That is the loop I cared about: the map flags something shaky, the agent asks about it, and your own answer either firms it up or kills it in front of you.
REFLECTION
What I actually wanted from it.
I did not want to build another tool that thinks for you. There are plenty of those. The thing I kept landing on while building this is that the scarce part was never generating ideas, it is seeing the ones you already have clearly enough to keep them or let them go. The most useful thing the AI does here is not answer your question. It holds up an honest picture of how you are already thinking, and leaves the decision where it should be, with you.