I used "manual reasoning" with Claude from @anthropic.com: Forced the model to write an XML with its thought process, then parsed this to only extract the final response.

I switched to the built-in reasoning, tried both the previous prompt and a refined one, result: