[14sea]

I'm glad to have your attention. To clarify, while Claude Code completed most of the work, hardware verification required my manual intervention.

Furthermore, I used Gemini in my work. After Claude completed a small step, I would ask him, "What's the next step?" Then I would give Gemini the summary of the current progress and Claude's next steps, along with, "This is my current progress; do you have any suggestions?" I would then pass Gemini's feedback back to Claude, adding, "This is a suggestion from another LLM; do you think it's helpful?" (This step is crucial; Claude doesn't dare contradict the user, but his evaluation of another LLM is extremely sharp (even a bit harsh)). Through this process, Claude absorbs good suggestions, broadens his thinking, and completes the work better.

Claude gets stuck in locally optimal solutions on his own (a bit obsessive), while Gemini's work is too scattered (he can't work independently at all). Combining them works well.

(I can't wait to see everyone use the same methodology to analyze more previously closed systems!)

(By the way, you can find the skill and full documentation in my project directory.)

[fleeceshepherd]

This is great! Definitely going to give this a try. But maybe thinking of using qwen cos Claude code credits are ex T-T. Do you think there would be big differences?

[14sea]

Thank you for your reply. As long as the hallucination model can be controlled, I believe these tasks can be accomplished.