Every, the AI-focused publisher and product studio run by Dan Shipper, has built an internal agent trained on 30,000 historical edits from its editor in chief Kate Lee, and now uses it to copy-edit essays, launch emails and landing pages across the company. Shipper disclosed the system in an interview with Casey Newton, framing it as a test case for how small companies can bottle the taste of a single expert employee. Every doubled from about 15 people to around 30 over the past year while running six software products and a daily newsletter — a workload Shipper called "insane" for a company that has raised little outside capital.
The Kate Lee agent works by pairing a hand-tuned prompt with the 30,000-edit dataset, then hill-climbing against her prior documents until the model's output matches her judgment closely enough to ship. Staff invoke it inside Google Docs by tagging "the Every Agent" and asking for a Kate copy edit, and the model uses browser-based computer use to leave suggested changes. Shipper said the workflow only clicked recently, after instruction-following and computer-use capabilities crossed a usable threshold.
Shipper has been trying to automate the role since GPT-3, without success until this cycle. The unlock, in his telling, is that current frontier models can now follow a long, specific prompt and operate a document editor reliably enough to intervene at the sentence level.
“We have an editor in chief, Kate Lee, who's fantastic, who I've been trying to automate out of a job for years in an extremely benevolent way.”— Dan Shipper, CEO of Every
Key facts
- 01Every collected 30,000 historical edits from editor in chief Kate Lee to train an internal copy-editing agent that back-tests against her past work.
- 02The company doubled from about 15 people to around 30 over the past year while running six software products and a daily newsletter.
- 03Every bundles Cora, Sparkle, Spiral, Monologue and its journalism into a single $20-a-month subscription.
- 04CEO Dan Shipper says AI now writes essentially all of Every's code, and product cycles reset every three to six months.
- 05Every launched in 2020 with Nathan Baschez and has not raised significant venture funding, per Shipper.
Every launched in 2020 with Nathan Baschez as a bundle of business newsletters and has since become a hybrid publication and product lab. Its consumer software today includes Cora for email, Sparkle for file organization, Spiral for writing, and Monologue for dictation — all packaged with the journalism into a $20-a-month subscription. Shipper says AI now writes essentially all of Every's code, while humans still write most of the essays.
The lean structure only became viable because a single engineer can now run an entire software product end to end, Shipper said. Every eventually staffs more than one person per product, but the ability to reach real customers with a solo owner is what enabled the six-product lineup on a small team.
Shipper compared the trajectory to The New York Times, which took 150 years and enormous scale before it could bolt Cooking, Games and The Athletic onto the core newsroom. Every, he argued, can attempt a similar bundle much earlier and with less capital because model capability has collapsed the cost of shipping software.
The editorial side runs "vibe checks" on new frontier models before general release, and Shipper acknowledged the awkwardness of publishing critical reviews of labs the company depends on. Every recently panned Anthropic's Sonnet 5 as "a model pitched for everyone impresses no one," and Shipper said the labs typically ask for feedback before launch precisely because a negative Every review predicts broader user reaction. He argued Every's value as an independent arbiter is one asset the model companies cannot replicate.
That independence sits inside an uncomfortable dependency. A June Every piece titled "Built on Moving Ground" laid out the vertigo of building on models the company does not control, and Shipper concedes he has no clean answer for the risk that Anthropic or OpenAI ships a native feature that vaporizes a product Every spent a year building. His working hypothesis is that model makers build ovens; someone else has to build the soufflé, and product cycles at Every now reset roughly every three to six months.
“I liken them a little bit to oven makers. You can make the oven, but it doesn't mean you know how to make a soufflé.”— Dan Shipper, CEO of Every
Shipper made a broader claim about the profession that is likely to draw pushback: that adoption of AI writing tools among working writers is far higher than the public conversation suggests. He said most writers using AI are not disclosing it, framing that gap as the industry's dirty secret rather than an argument against the practice.
The counterweight sits in what the models still cannot do. Shipper said AI is "trained on the residue of human expertise" and cannot see past it, which is his stated reason for continuing to hire even as automation eats internal workflows. That framing is convenient for a CEO doubling headcount, and it papers over the harder question of what happens when the residue gets deep enough that new hires stop adding signal. The Kate Lee agent itself is an existence proof that a single senior employee's judgment can be extracted, distilled and distributed — the near-term productivity story, but also the medium-term automation story.
For the broader AI market, Every is worth watching because it is one of the cleanest public examples of a small company using frontier models to punch far above its headcount in two very different disciplines at once. If a 30-person team can credibly run six software products, a daily publication, and an internal agent that mimics its best editor, the operating leverage available to any information-services business is much larger than most incumbents have priced in. The uncomfortable part for the labs is that the same shop reviewing their models is also the sharpest demonstration of what happens when someone else figures out how to cook with them.
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.
