How much of my boss's job can AI do?

Six months ago, one Platformer writer built an AI version of themselves called Claudella to see if it could handle their newsletter duties. This time they went bigger: an agent modeled on their actual boss, Casey Newton, built on Claude Fable 5 and nicknamed Claudeasey Newton. The goal wasn't just curiosity. It was a real test of how far AI has come since February, when the same writer first started poking at these questions.

The setup involved feeding the model roughly six years of Platformer archives, a full history of Casey's Google Docs edits on the writer's own pieces, and about a year of internal Discord chats. From that pile, Claude produced a style guide covering things like paragraph length and how Casey refers to colleagues, including references to his former landlord, Kara Swisher. The idea was to recreate three parts of the job: writing columns, chatting and sharing takes, and editing.

The results were mixed but genuinely improved from before. Claudeasey's first column, about Microsoft layoffs, got bogged down parsing the semantics of Microsoft's official denial that AI caused the cuts. But after the writer had Claude critique its own draft against real Platformer columns, the bot sharpened its approach, learning to frame arguments around who made a decision and who pays for it, and to steelman opposing views. A later column, on the White House's undisclosed voluntary AI safety framework, read noticeably more human, closing with a line about officials needing to take risks seriously