Turns out write drunk edit sober works for models too. Have been having a great experience with big projects to ask models (GPT 5.6 Sol x-high) to just write drunk (without overly testing, compiling, etc), reviewing drunk and than editing sober to fix mistakes. Not sure why, but
