Every trains an AI agent on 30,000 edits from its writer

⚡

Key Takeaways

1An internal agent from Every, trained on 30,000 edits by Kate Lee, suggests corrections in Google Docs and improves over time with reviews.

2The workforce has increased from about 15 to 30 people in a year despite automation, according to Dan Shipper.

3Every has published a critique of Sonnet 5 and claims a role as an arbiter among model laboratories.

💡Why it matters — This demonstrates how Every combines intensive internal use of AI with editorial activity and a critical stance towards model providers, all while continuing to hire.

Every has developed an internal agent capable of applying the edits of its editor-in-chief in Google Docs, based on 30,000 historical corrections. Dan Shipper claims that the AI writes almost all of the in-house code, allowing the management of multiple products with small teams while continuing to hire. He asserts a role as an arbiter between labs, despite published criticisms of certain models, and insists that far more authors use AI than they admit.

An Internal Agent Reproduces Kate Lee's Edits

Dan Shipper explains that the models are now reliable enough to follow complex instructions and even use a browser or computer to insert suggestions into Google Docs. According to him, a task that has long been resistant, such as copy editing, is starting to work, following trials dating back to the era of GPT-3.

To achieve this, he says he gathered 30,000 historical edits from Kate Lee, then designed and refined a prompt tested on previous documents. Every now uses an agent that teams refer to as "applying a Kate-style correction" on documents and landing pages; the tool learns from subsequent reviews by Kate and highlights what it missed. It is not perfect, Shipper acknowledges, but it would advantageously replace some manual interventions.

The stated goal is to capture the expertise of a key individual and disseminate it more widely within the organization, a process that Shipper presents as indicative of future relationships between AI and employees. He specifies that Kate Lee also undertakes broader editorial tasks and that the team describes this system as "composed," or composable.

Increased Productivity Without Giving Up Hiring

Dan Shipper asserts that Every would not have been able to accomplish nearly everything it does without AI. At one point, the company may have had 12 or 15 people and was already managing six software products as well as a daily newsletter, a task he finds challenging, with little money raised and a decisive contribution from AI to multiply the effectiveness of a single engineer.

He states that AI now writes almost all of the company's code, while humans focus mainly on testing. According to him, one person can now gather user signals and serve a real customer base with a single product, a level of scale that was once out of reach; some product teams now consist of several people.

Over the past year, the number of employees at Every has increased from about 15 to 30, alongside a "noisy" automation of internal processes. Shipper attributes this evolution to his conception of AI that relies on what remains of human expertise, without the ability to exceed that framework, and the necessity to be able to start from scratch every three to six months based on progress made, a method he considers demanding but achievable.

In contrast, Shipper cites The New York Times, which only launched its gaming, Cooking, and The Athletic ventures after 150 years of existence and on a large scale; he believes that Every can initiate this type of diversification more quickly and earlier, with fewer financial resources.

Public Criticisms of Models and Relations with Labs

Every has published a harsh critique of Sonnet 5, described as "designed for everyone but impressing no one." Dan Shipper claims to know many people at OpenAI and Anthropic and states that he does not want to be "mean" to friends, while explaining that these labs seek their feedback in advance and prefer early feedback to massive user reports.

He argues that the labs would like to avoid papers stating that a model is bad, but presents Every’s goal as helping to bring forth better AIs, in partnership, despite tensions deemed exceptional. Shipper claims a lasting value of arbitration: stating which models are good, a task he finds difficult to entrust to the manufacturers themselves.

To illustrate, he compares model companies to oven manufacturers, who do not necessarily know how to make a soufflé. He also recalls an article published in June, Built on Moving Ground, regarding the risk of building products on uncontrolled models, as labs continuously improve their systems and also operate application layers that can support and compete with players like Every. He describes a cross-play where Every proposes uses, labs improve models, and sometimes develop similar features themselves.

Product Offering and $20 Monthly Subscription

Every combines publishing and product studio, with about 30 employees. Launched in 2020 by Dan Shipper and Nathan Baschez, the company markets Cora, Sparkle, Spiral, and Monologue, bundled in a $20 monthly subscription.

Editorially, Every relies on the Chain of Thought column, the AI & I podcast, and "vibe checks" that test cutting-edge models before their public launch.

"Almost All Writers Use AI," According to Shipper

Dan Shipper asserts that far more authors use AI than they publicly acknowledge. He goes so far as to estimate that almost all writers use it, while noting that most do not say so.