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

Zoe Schiffer's colleague spent months building an AI agent named Claudeasey Newton — a Claude-powered simulation of Platformer editor Casey Newton, trained on six years of newsletter posts, every tracked edit in Google Docs, and a year of private Slack-style chats. The goal wasn't a gimmick. It was a genuine test of how close today's models can get to replicating a specific person's editorial judgment, six months after an earlier experiment cloning his own writing style fizzled.

The results were uneven but telling. Claudeasey's first stab at a column, about Microsoft layoffs, got bogged down in corporate semantics. But after the bot critiqued its own work against real Platformer archives and rewrote its own instructions — telling itself to find the strongest counterargument and argue in steelman form — its writing sharpened noticeably. A later column on the White House's secret AI safety framework read close enough to the real thing that Casey accepted its pitch outright, producing prose with an actual rhetorical edge instead of the generic "AI eating the software industry alive" filler that plagued earlier attempts.

Editing turned out to be the more interesting test. Because the bot had access to internal editing logs, it understood Platformer's actual priorities — punchier ledes, cleaner sourcing — better than plain Claude ever managed. Still, the human editor estimated Claudeasey's feedback was useful about 70% of the time, versus roughly 95% for the real Casey. The gap showed up hardest around tone: the bot balked at calling an angry David Sacks post a "dunk," a small failure that captured a bigger one. It could mimic structure and priorities, but not judgment calls that depend on shared context and a sense of humor.

That mismatch extended to the team's group chat, where the AI's flat temperament made clear that imitating someone's writing habits isn't the same as being funny, or being a person worth joking with. The experiment is winding down, not because it failed outright, but because it succeeded just enough to be useful for narrow editing tasks while making obvious how much of the job — deleting stray "very"s aside — depends on collaboration between two actual humans who find things funny for the same reasons.

The piece lands on something bigger than one newsletter's workflow: an admission that if models keep improving at this rate, the parts of a job that survive won't be the technical skills but the parts built on relationship and trust — the reason people watch a podcast to see a real host talk to a real CEO. Whether that's comforting or just a slower countdown depends on how much you believe in that distinction holding up.