Run Prompt Evals Side by Side, in a Tree

How it works

Different models have different strengths. One is better at reasoning, another at code, another at concise answers. The only way to know which fits your task is to ask the same question and compare.

1. Start with your question. Write a prompt like you normally would. The first model responds.

2. Branch to another model. Fork the node and switch models. Same prompt, same context, different perspective. The tree holds both responses side by side.

3. Go deeper on the better answer. Follow up on whichever response is more useful. Or follow up on both and keep comparing. The tree preserves every path.

4. Keep the full context. Every branch inherits the conversation history from its ancestors. When you switch models mid-conversation, the new model sees everything that came before.

Why trees beat split-screen for model comparison

In a linear chat, switching models means starting over or copy-pasting context between tabs. You lose the thread. In a tree, every model’s response lives in the same conversation. You can branch at any point, compare, and continue without losing a single message.

This also means you can evaluate models at depth, not just on a single prompt. Model A might give a better first answer but fall apart on follow-ups. The tree shows you that.

Example

A developer evaluating Claude and GPT for code review asks both the same question about a tricky race condition. Claude spots the bug but suggests an overly cautious fix. GPT proposes a cleaner solution but misses an edge case. The developer branches from Claude’s response to ask about simplifying the fix, and from GPT’s response to probe the edge case. Two follow-ups later, the best answer is a hybrid of both, and the tree shows exactly how they got there.

For related workflows, see decision making for weighing trade-offs, or deep dive for stress-testing an idea.