I cannot definitively prove this, but if you demand professionalism in its communication, it appears that its thinking process itself altered. The code it writes has a different quality of modularity/composability. We have a usgc-handbook plugin for internal use, I cannot describe the utter pleasure of working with Claude that speaks proper english and responds in a console style that's professional in true sense of the term. I see a bunch of memes where the midwit gives a massive md file of instructions and the grug/elite just say "Fix it". I am not sure if this is a good idea. "Write a to-do list app" -> it's output isn't objective and falsifiable unless you are also accepting a wide range of outputs. From talking to Claude why this may be true: > Attention heads activate on register features (terseness, formality, vocabulary) and those features are correlated in pretraining with authorship traits — senior engineers, documentation, well-maintained repos. Conditioning on the register moves the model toward that region of the data manifold, and code quality is a property of that region. I am not an expert at LLM in the slightest, but what I can say is that my observation aligns with it. YMMV.