Coding showed the pattern first: let AI handle the maintenance so the person can spend time on the work that requires judgment, trust and a name.
By Ryan Merket · Published
Primary source: wired.com
Why it matters #
AI newsroom economics become more defensible if automation funds human sourcing instead of replacing it. RuntimeWire is testing that division inside a live publication.
The picture most people have of AI and news is super gloomy, it might as well be an episode of Handmaid's Tale. A model writes the story. An editor hides that fact in a note that does not say what happened. The craft dies. Nobody gets paid.
That is not the only available picture.
The one I am arguing for is smaller and more ordinary: a beat with a sidekick. The reporter still has a name and a phone. The sidekick already did the embargo calendar, the transcript, the cap-table scrape, and the sweep of who shipped overnight. The human spends the day on the conversations that required a human.
I was in Big Bend with almost no internet except the phone. I ran the site over iMessage. More than eighty stories went out that week. WIRED wrote that part down. It sounds like a stunt. It was a proof of the split. The queue can move when you are offline because the daily layer does not need you in the chair. The work that needs you never lived in that queue.
The prompts that drive RuntimeWire's sidekick will soon be public on GitHub. Anyone will be able to see how we draft a round, kill a duplicate, and insist on a primary source. The prompts are not the scarce part. The scarce parts are the same ones launch posts and embargo mornings have always consumed: relationships, taste, and time. The machine handles the speed so a person can invest in the first two.
A Tuesday
Two columns. Same clock.
| Time | You | The sidekick |
|---|---|---|
| 6:15 | Asleep, or driving. | Overnight filings, launches, embargo folder. Three briefs in house style. One killed as a duplicate. Two ready for a human yes/no. |
| 7:40 | Kill or ship the housekeeping in minutes. Do not rewrite a round from a press release. | Already cited the Form D, the blog post, the tweet. A checklist of who did what, attached. Emailed an RFC (request for comments) on press releases and embargoes. |
| 9:00 | One story survived the morning. A founder who is not on an embargo list. You text. | Document dump, last three things they shipped, claims checked against a filing. Source list. Not on the call. |
| 11:30 | The call. | Transcript only. |
| 2:00 | You write. The sentences are yours. | Fetch, don't voice. |
| 4:30 | The day's launches already went out under the house byline. | Volume work is done. |
| 6:10 | One story publishes under a name. | It did not draft this one. |
The gloomy version assumed the job was the sentences, so a machine that writes sentences is the end. A beat with a sidekick assumes the job was always who you can get on the phone and what you decide is worth a day.
Does the daily layer hold
The Tuesday above only works if the housekeeping material by the sidekick is actually good enough to publish. If the briefs are slop, you did not free anyone. You just shipped fast slop.
So we checked the boring layer the way a reader meets it. In early August we took 120 stories that RuntimeWire and another tech outlet both covered, stripped the logos, and compared complete text to complete text. We won 98, lost 21, tied 1. VentureBeat still beat us 17–4–1 when it was their story against ours. That is not “we write better than humans.” It is narrower: on launches, rounds, and the other items ten sites already have, the sidekick-plus-edit version is on par often enough that a reporter should not spend the morning rewriting a press release.
Readers can say so themselves. On each story there is a slider from noise to worth-my-time. About 1,500 of those ratings now sit on 462 pieces. The week averages around 76, the longer window around 70. That number does not mean they thought a person typed every sentence. It means a lot of the builder wire — shipped products, odd hardware, a real funding — did not feel like a waste of a tab. The hidden-in-the-client scoops score colder. Useful, not a trophy.
If those two checks had come back ugly, the sidekick would be a cost. They came back good enough that the scarce hours can go to the call at 11:30. Software already ran this movie
In May 2025, David Heinemeier Hansson — Rails, 37signals, the last person anyone expected to cheer a prompt — posted: I have no interest in giving up writing code. That’s not the unpleasant part that I want AI to take off my hands. Just so I can — what? — become a project manager for a murder of AI crows?
That summer, on Lex Fridman, he called vibe coding “empty calories.” If you can only vibe it, he said, you are not a programmer.
Then the tools changed.
On January 3rd, 2026 he posted: You can’t let the slop and cringe deny you the wonder of AI… If you spent 2025 being pessimistic or skeptical on AI, why not give the start of 2026 a try with optimism and curiosity?
By July he was explaining the flip in a reply: That was before agents and Opus 4.5. I had no interest in the ergonomics of autocomplete, but thankfully AI evolved, and now virtually everything starts with a prompt.
By August, on Lex again, he said he had not handwritten any of the code that shipped in Omarchy Quattro. He reviewed the shape of it, and the critical lines. Agents did the rest. He still owns the repo.
Linus Torvalds, the founder of Linux, took the same walk, with less poetry. In July, on the kernel list, he put his foot down: “Linux is not one of those anti-AI projects, and if somebody has issues with that, they can do the open-source thing and fork it. Or just walk away.” AI, he wrote, “is a tool, just like other tools we use. And it’s clearly a useful one. It may not have been that ‘clearly’ even just a year ago, but it’s no longer in question today.” He will not force anyone to use it. He will “very loudly ignore” people who try to ban everyone else from using it.
A month later he used one on a “debug session from hell” in an Intel graphics driver. The model kept calling the bug impossible and suggesting they just file a report. He pushed. Twenty-four debug patches, eighteen boots, one-line fix: round_up()
to round_down()
. He let the model write the commit message. He still signed the tree.
That is the pattern. The gloom was real. The tools got good at grunt work. The people whose names are on the work did not disappear. They stopped pretending the grunt was the craft.
Journalism is harder than kernel work for a reason that has nothing to do with virtue: the sentences are the product.
A patch compiles into something nobody reads as writing, so a model can do the grunt and a maintainer can still own the tree without anyone having to enjoy the intermediate prose. A story is consumed as prose. The reader meets the labor in the same object they came for, which is why a cowardly note at the bottom — “the key points were written by a journalist” — fails so completely. It will not say who drafted, who reported, who cut, and it leaves the only thing the reader can test, the sentences, in a fog. House copy has to say who did what. Reported copy has to carry a name. The sidekick can take the embargo calendar, the transcript, the overnight sweep. It does not get the byline, because the byline is the part that cannot compile away.
The sidekick is not a co-author. It is the intern who never sleeps and does not get a byline. You can clone the intern. You cannot clone who takes your call, or who knows the embargo is theater.
I built this so more founders get covered than a shrinking press can touch, and so the people who can still do the other job are not also filing three launch posts before lunch. The wire is the calendar. The named work is why the beat still matters.