A journalist at Platformer built Claude Fable 5 into an agent called Claudeasey Newton, trained on six years of Platformer archives, every edit Casey Newton ever made on a colleague's drafts, and nearly a year of private team Discord logs. The goal: find out how much of an editor's job the smartest publicly available model can now replicate.
The methodology is worth reading in detail. The author had Claude generate its own style guide from the archive, then critique its own drafts against real Platformer columns, then rewrite its own instructions. That feedback loop produced measurably better output: more concrete arguments, sharper editorial judgment, and a pitch that the real Casey Newton accepted. Six months ago, a similar experiment with an earlier model fell short. This time, the gap closed.
The failure mode is still the same one from February: the agent broke mid-story during a live demo. But the author's conclusion isn't reassuring. Any part of the job a model can't do today, it may be able to do soon. The piece ends not with a technical verdict but a personal one, which is exactly why you should read it in full.
[READ ORIGINAL →]