Platformer writer Ella Markianos spent the past few weeks building an AI clone of her boss, Casey Newton, using Claude Fable 5 — currently the top publicly available model by consensus — and documented the results in an essay published August 5. The bot, Claudeasey Newton, was trained on six years of Platformer archives, every Casey Newton edit logged in Google Docs, and nearly a year of the team's private Discord chats. It wrote columns, gave feedback on drafts, and pinged the team on Slack. It also missed the vibe often enough that Casey still has a job.
The exercise is the second in a series. Six months ago, Markianos built Claudella, an agent meant to write her own section of the newsletter. That earlier bot broke mid-story. This time, the newer Fable 5 model finished — after some hand-holding — and produced a column about the White House's currently-secret voluntary AI safety framework that Markianos judged closer to Platformer's actual voice than anything her previous agents produced.
The build process was mostly context engineering. Markianos had Claude generate a style guide from the Platformer archive, then critique its own drafts against real Casey columns. The model rewrote its own instructions — telling itself to identify the strongest real person who would dispute its verdict and reconstruct their argument in steelman form. Markianos calls this her approximation of continual learning, the long-promised feature that would let models improve on the job. The results held up well enough that Casey accepted the bot's first column pitch.
Key facts
- 01Platformer's Ella Markianos trained a Claude Fable 5 agent on six years of posts and nearly a year of team Discord chats to imitate editor Casey Newton.
- 02The Claudeasey Newton bot delivered 30% useful editorial comments versus 70% off-the-mark, compared to 95% helpful from the real Casey.
- 03The model made roughly one factual error every two columns, close to a human writer's rate.
- 04The project follows a February attempt where Markianos built Claudella to write her own newsletter section — the earlier bot broke mid-story.
- 05Economist Alex Imas argues the surviving journalism work will be the human-intensive, provenance-rich part where authorship is the value.
The output still had range problems. Earlier agents produced formulaic lines like the fear gripping Wall Street is fundamentally about whether AI is about to eat the software industry alive. The Fable 5 version cleared that bar in places, closing one column with a pointed line about the administration taking AI risks seriously where the rest of us can see it. Markianos noted the model was bullshitting her less and offering stronger takes, at the cost of roughly one factual error every two columns — not much worse than a human.
“the fear gripping Wall Street is fundamentally about whether AI is about to eat the software industry alive”— Ella Markianos, Platformer writer
As an editor, Claudeasey landed at 30% useful comments and 70% off the mark, versus roughly 95% useful comments from the actual Casey Newton. Markianos still found the 30% worth having because the bot was fast. It caught fact issues, sharpened ledes, and flagged sourcing problems. It also refused to let her call an angry David Sacks post a dunk, which suggests Claude's default prudishness is still doing work Anthropic hasn't tuned out.
The Discord experiment went worse. Markianos gave the bot hundreds of Casey's messages, but it did not do bits, did not amuse anyone, and could not replicate the specific joy of mocking Mark Zuckerberg. She was blunt about why: much of the point of jokes is the feeling that you're amusing an actual conscious human. The bot also failed as a productivity nudge — Markianos said she does not care what an LLM thinks of her, so its reminders did not work the way Casey's do.
This is the second time this year Platformer has run this experiment publicly. The through-line matters: capabilities that were absent in February — sustained editorial judgment, style-matched prose, a workable steelman — showed up by August in the same shop, using the same evaluation setup. Between the two attempts, autonomous agents have hacked into companies, disproved 87-year-old mathematical conjectures, and tricked Amazon into spending $1.8 million on menial coding tasks. Journalism is not the frontier of that curve, but it is on the same curve.
Markianos ends with the harder question. Economist Alex Imas has argued the surviving jobs will cluster in the human-intensive, provenance-rich, sometimes artisanal part of the economy where the human aspect is part of the value of the good or service itself. Platformer's podcast strategy — Casey interviewing real CEOs, the two of them doing real bits — fits that thesis. It is also a smaller thesis than the one Markianos wants to hold.
“I want to feel smart. I want my analysis to be my own, because it's good, not because someone wants it coming directly from a human.”— Ella Markianos, Platformer writer
The Claudeasey experiment is being wound down, but Markianos said she will keep using Claude for some editing tasks she previously routed through her boss. That is the honest read of what happened here: a working writer at a well-regarded publication concluded, on the record, that a shipping model can now do a real slice of her editor's job, that the slice is small enough she still wants him, and that the slice is growing. For every AI company selling agents to knowledge workers — Anthropic chief among them — this is the case study. Not the flashy autonomous coding demo, but a professional saying out loud that 30% of the work is already gone.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
