Spent the last weekend doing a side-by-side. Same codebase (a small Next.js app I'm refactoring), same questions about the same files. Sharing what I found because the "AI for coding" space is becoming impossible to navigate.

Setup:

Same 3 prompts: "find bugs", "suggest refactor", "explain this useEffect"

ChatGPT 5 (Plus), Claude Sonnet 4.5, and a Gemini 2.5/3 with a custom system prompt I built (Kody, on codemasterip.com)

Findings:

ChatGPT was best at one-shot bug detection but verbose

Claude had the cleanest refactors but missed a subtle race condition

The custom-prompted Gemini was middle of the pack on raw output but much better at "stopping to ask" before refactoring — which actually saved me from a wrong direction once

The custom one also kept session context better for follow-ups (which makes sense, it's prompted for that)

Caveat: I built one of the three tools. So I tried to be harsh on my own.…