I've been putting off a big refactor for a while. I know roughly what I want to do, but the scope is large enough that I haven't been able to get my thoughts into a shape I can act on. Today I tried something new to break through that.
I started by jotting down my ideas in a rough notes document — just stream of consciousness, no structure. Then I asked Claude to look over the codebase and use those notes to build a brainstorm.md with three sections: Current State, New Ideas, and Open Questions. The goal was to have something concrete I could share and get feedback on.
Then I uploaded that document to Gemini, ChatGPT, and Grok — one at a time — and asked each for their thoughts. After each response, I pasted the feedback back into the document before moving to the next model. By the end, brainstorm.md had accumulated a layer of commentary from three different sources.
Finally, I brought the whole thing back to Claude Opus and asked it to distill everything into a proper feature proposal.
The most interesting part was how different each model's feedback was. Gemini was surprisingly terse — less useful than I expected given it's usually pretty strong at this kind of analysis. ChatGPT was verbose, but the substance was there if you dug through it. Grok was the standout — good, direct feedback and it came back with questions, which pushed me to think harder about the parts I'd glossed over.
The workflow is rough. The prompt I used at each stage was pretty generic:
"I have uploaded a document which has brainstorming ideas for a refactor and feedback from other models. Can you offer your thoughts and critiques on the brainstorming ideas in this document."
That's fine as a starting point but leaves a lot on the table. A better prompt would probably specify what kind of feedback you're after — are you looking for holes in the reasoning? Alternative approaches? Risk assessment? Giving each model a more specific angle would likely get more useful output.
Still, even with the rough edges, this produced something I couldn't have written on my own in an afternoon. It's worth being honest about what this actually is: a hacky workaround for getting multi-model feedback without paying for API access across everything. Copy-paste isn't a pipeline. But the combination of perspectives forced the proposal to be more complete than if I'd just asked one model and called it done. Worth refining into something cleaner.