Run prompt evals side by side, in a tree
Send the same prompt to different models on sibling branches and compare their answers. Each branch keeps the conversation that led to it.
See it in action
Two models answer the same refund-policy questions side by side, so you can see which one sticks to the policy.
How it works
Write a prompt, then branch from the same point to ask other models. Each answer sits on its own branch, next to the others.
-
Start with your prompt. Write it as you normally would, such as a system prompt followed by a test question. The first model answers.
-
Branch to another model. Add a sibling with the same prompt and pick a different model, such as Claude, GPT, Gemini or DeepSeek. The answers sit side by side.
-
Follow up on either answer. Ask the next question under whichever answer is more useful, or under both, and keep comparing.
-
Switch models partway through. Every branch carries the conversation from the root down to it. A model you switch to sees everything above it on that branch, and nothing on other branches.
Why a map instead of separate chat tabs
Comparing models in separate chat apps means pasting the same context into each tab. On a map, the answers share the conversation above them, so you can branch at any point and carry on.
You can also compare models over several turns. One model may give the better first answer and then fall apart on follow-ups, and the branches show that.
Example
Say you are choosing a model for code review and ask two models about the same race condition. One spots the bug but suggests a heavy-handed fix, and the other proposes a cleaner fix but misses an edge case. You ask the first how to simplify its fix and ask the second about the edge case. Two follow-ups later, the best answer combines both.