Fifteen People, Four Products, Zero Handwritten Code: Inside Every's AI-First Company

Shipper opens with what he calls his hottest take and the one he has the least evidence for: AI could be one of the biggest forces for reshoring American jobs. Most commentary focuses on unemployment, and he accepts that the skills needed for many jobs will change. But he thinks the debate is missing a second effect.

The first mechanism is demand. Services that only rich people and large companies could afford, such as in-house counsel or a call center, become affordable for small companies and individuals once intelligence is cheap. The second is leverage: a handful of people in the Midwest could serve hundreds of thousands of customers instead of working the phones all day.

That changes the economics of hiring offshore, he argues. If a US-based worker armed with AI tools can do the job of a far larger team, it becomes more cost-effective to keep the work at home. He adds that the model companies are in the US too, and that people here are likely to get good at using the tools, though he leaves it to listeners to decide whether that is a good thing.

Shipper's second take is that people underrate how good Claude Code, and similar tools like Google's Gemini CLI, are for non-coders. He explains it is a command-line program with access to your file system and the web that can run autonomously for 20 or 30 minutes, even spawning sub-agents in parallel. Everyone at Every uses it all day, and he says they are all "agent-pilled".

The barrier for non-programmers is only that the terminal feels intimidating. Once past it, you can drop a folder of meeting notes in and ask for every moment you subtly avoided conflict. The agent writes itself a to-do list, reads each file in turn and returns a summary, rather than stuffing everything into one context window like a regular chat would.

Shipper also used it for his writing. He downloaded the public-domain text of War and Peace, had Claude extract Tolstoy's subtle character descriptions from the early chapters and build itself a style guide, then compare the English and Russian versions to spot what translation might have lost. The same approach, he says, works for piles of customer interviews or other large datasets.

He sees this pointing to a different kind of interface. The first AI apps pasted a chat box into an existing UI, as Copilot and Cursor did. With Claude Code you never look at the code: you delegate a task and it gets done. Shipper expects that for most applications AI will soon be good enough that we stop being interleaved with its execution, and he notes that no one at Every codes manually anymore.

Shipper offers his own working definition of artificial general intelligence. You can tell how much better AI is getting, he says, by how long a leash you can give it before you have to intervene. Copilot offered tab completion, ChatGPT answered a question, and now models like Claude Opus 4 and deep research tools can work alone for 20 or 30 minutes.

He borrows an analogy from the child psychologist Winnicott, who described growing up as being gradually let down in moments you can handle, learning you are separate from your caregiver. Raising a child, Shipper says, is knowing when they are ready to be left alone a little longer. AI, in his view, is somewhere a bit past toddler stage.

His definition of AGI follows from that: it is the point at which it becomes economically profitable to run agents indefinitely, with a Claude Code that never turns off and is never waiting for your next instruction. Lenny notes that the profit includes the cost of running it. Shipper agrees it is partly cost and partly value, and that the profitability bar keeps it honest, since the agent must actually do something useful.

Shipper says he hates headlines claiming AI will unemploy two thirds of the workforce, that using ChatGPT switches your brain off, or that AI alone beats doctors with AI. Using AI is a skill, he argues, so a study of doctors with little experience tells you little about the future. Medicine involves too many contexts for a single study to settle, and he guesses the picture will look very different in five or ten years.

On the brain-connectivity studies he points to Plato, who distrusted writing because it would damage memory. It did, Shipper says, and the trade was still worth it. Every technology costs some skills and unlocks others, and no one would go back to a world without literacy. Lenny adds that studies from places like Nigeria show students progressing quickly with AI.

He also says it is easy to underestimate how complex specific human contexts are, using the tailor near his Brooklyn home who still does not take credit cards as an example. To make the point he built a "CEO benchmark" that tests whether models can predict what he will say in a meeting, using his Granola transcripts. The frontier models were pretty bad at it, which he attributes to context rather than intelligence.

That leads to context engineering, getting the right context to the model at the right time, which he says is at least half of performance. He had been writing about it for three years under the name knowledge orchestration. As a CEO, he says, so much of the job is repeating yourself, so last quarter's goal was to never say the same thing twice. His team has codified his headline and intro feedback into prompts so writers talk to a simulation of him first.

Shipper describes Every as a company that does ideas and apps at the edge of AI. The core is a daily newsletter, about five years old with roughly 100,000 subscribers, that gets early access to new models from OpenAI and Anthropic and publishes what it calls vibe checks. Those capture how a model feels in real work, which he says standard benchmarks cannot. Around that sit a suite of apps, bundled under one subscription, and a consulting arm.

The apps include Cora, an AI chief of staff for email that launched publicly the day of the recording, Sparkle, an AI file cleaner, and Spiral, which automates content. Every also incubated the AI document writer Lex, which was spun out into its own company run by his co-founder. The whole company is 15 people.

One practice he recommends to everyone is having a head of AI operations. He meets her weekly, they list every repetitive thing he does, and she builds prompts and workflows to automate it so the people doing the work do not have to stop and tinker. The trick, he says, is making sure the tools actually get used, which is partly a behavioral change.

That person is Katie Parrott, who began as a ghostwriter and previously worked at the content marketing agency Animalz. Shipper says the ideal profile is someone who loves to tinker, ideally with a process mindset and an understanding of the craft. The work now spreads beyond editorial: an Every engineer built a Claude Code command that applies the copy-edit prompt across a codebase and sends Katie a pull request to review.

Shipper says the first thing he opens is o3, which he calls high quality for writing and coding and which has one thing Claude lacks, memory. After spending so long telling ChatGPT he wants punchy, concise writing, it now just knows. He also uses it for self-reflection, sending it meeting transcripts and asking how he did.

For building, everyone at Every uses Claude Code. Gemini is the model they use most inside their apps because it is powerful and cheap, and they use OpenAI's Codex for one-off, self-contained features. Outside the coding tools he uses Granola, an internal speech-to-text tool called Monologue that is due to ship soon, and Notion's meeting recording.

He singles out one capability of Claude Opus 4: it can tell whether writing is any good. Earlier models would grade a draft B+, then A-, then A as you revised it, because they were guessing what you wanted to hear. Opus 4 has a real gut sense, which lets a model act as a judge. The team spent three months building a complex system to judge writing for Spiral, and Opus 4 one-shotted it.

The Cora team is two engineers, Kieran and Nityesh, plus what Shipper jokes is 15 Claude Code instances. They invented what he calls compounding engineering: for every unit of work, make the next unit easier. If your job in a Claude Code world becomes writing product requirement documents, you can keep writing each one by hand, or write a prompt that turns rambling thoughts into a good PRD.

The speed-ups are mostly prompts and automations, stored in a shared GitHub repository, and Claude Code slash commands are a big part of it. Cora was in private beta with 2,500 active users and millions of emails flowing through it, which Shipper calls kind of crazy for a 15-person company. He says all-in the product cost about 300,000 dollars, salaries included.

The engineers also run several agents at once, including Friday and Charlie, the latter living in GitHub so you can ask it to check a pull request. Shipper says the agents have different personalities, from terse and professional to something else, and that Kieran, a serious Rails fan, has a feel for which suits what. Lenny likens it to hiring a team of Avengers, each strong at something, and suggests there is a bigger market for varied agents than people think. Shipper agrees there will not be one agent to rule them all.

Shipper values generalists with unusual backgrounds. Kieran was a startup VP of engineering, before that a professional composer, and before that a baker who taught the team to make croissants on a retreat in France. Nityesh only started coding once ChatGPT arrived, and has progressed faster than any engineer Shipper knew in the pre-AI era.

That is why he rejects the fear that AI removes entry-level jobs. Whenever he sees a kid with ChatGPT, he thinks they will grow faster than anyone he has worked with. He cites Alex Duffy, who joined as a promising idea person but a weak writer. Every time Shipper explained how to tell a story or write a headline, Alex recorded it, put it in a prompt, and never made the same mistake twice, making a year of progress in two months.

He ties this to his allocation economy thesis: new workers learn to manage AI and do the work at once, effectively starting one level above entry. Management skills are expensive and rare today, with about 8 percent of the workforce as managers, but will spread as managing gets cheaper.

On whether the product team still looks at code, Shipper says they do review it, and Danny, who runs Spiral, spent a couple of days studying a third-party library before returning to directing Claude Code. He insists coding skill still matters: scripting languages sat on top of C, which sat on top of assembly, and English is the next layer. Dropping a level gets rarer over time, but it took 10 to 20 years before it mattered less, and he says we are not close yet.

Shipper's product playbook comes from his earlier point about expensive services. When you run an AI-first company you keep hitting things you wish you had, like a ghostwriter, a lawyer who would cost 25,000 dollars, or a chief of staff for email. Every tries ChatGPT or Claude first, and if it works they unbundle it into a standalone app.

He says the game board has been reset: five years ago it meant another notes app or B2B SaaS tool, while now everything is new territory, like the early days of spreadsheets. Employees at Every are the first users, and since they came to Every because they read it, readers with the same vibe become the next. He sees it as a pipeline for building apps and believes that what Every does at the edge, everyone else will be doing in three years. Lenny concludes that GPT wrappers are a good idea, and Shipper agrees emphatically.

On funding, Shipper wants Every to be a lasting institution that is also a creative playground, and he has avoided raising a lot of money because it locks you into being serious. They raised a 700,000 dollar pre-seed with a modified safe giving investors the option to convert in three years, warning them it was probably not a venture business. The recent round is what he calls a sip seed: up to 2 million dollars committed by Reid Hoffman and Starting Line, drawn down at a set cap whenever they need it.

He says this lets him take more risk without staring at a huge bank balance that begs to be burned, and that Hoffman cares more about the philosophy than the size. Shipper thinks very small amounts can go a long way now, since two engineers can do what once took 20 people, and he suspects venture capital has not caught up. He admits he might be back in two years regretting not raising more.

Every's consulting arm grew out of audience demand: companies asked to be taught how to work this way. Shipper says it did around a million dollars last year and may double. The process starts with researching the organization, interviewing teams and delivering a report plus a chatbot over the interviews and a dashboard of which teams are keen. He coded that app himself over a weekend with Devin.

Training follows, customized to each team. Shipper says roughly 10 percent of employees are curious, 10 percent will never touch it, and 80 percent will do it if you show them how it applies to their job. Teams get four weeks at an hour a week with their exact prompts and situations, often followed by automation work. Clients include large hedge funds, private equity firms and big companies.

His number one predictor of success is whether the CEO uses ChatGPT, or any chatbot, all the time. CEOs who do not either breed negativity or hold unrealistic expectations because they lack intuition for what is possible. He praises a roughly 10 billion dollar hedge fund client, Walleye, whose founder's AI-first memo said he wrote it with ChatGPT and that employees should too.

Strong companies also hold weekly meetings where people share prompts and use cases, and send weekly emails with usage stats and credit for contributors. The 10 percent of early adopters are the ones to find and reward, because they do the figuring out and the rest of the organization can adopt what they learn. The impact so far is more output with the same headcount, not layoffs, and Shipper dismisses the Klarna CEO story as nonsense.

Shipper explains the allocation economy, an idea from an article he wrote two and a half years ago. Using GPT-3 and GPT-4, he noticed he was spending time on how to communicate a problem, gather the right information, pick the right model, split the task, give feedback and hold a vision with criteria for good. All of that, he realized, is just managing.

That reframes the common complaint that you cannot trust AI to do it well so you do it yourself. He says it is exactly what first-time managers say: if you delegate it is not done your way, and if you do it yourself you get no leverage. Learning when to micromanage and when to delegate is the skill. Lenny adds that the article names evaluating talent, vision, taste and knowing when to dive in as more valuable.

Shipper also argues generalists will gain. Specialization let civilization coordinate better, he says, and he points to Athens, whose citizens were fighters, judges and jurors until it became an empire and needed skilled specialists. AI is like 10,000 PhDs in your pocket, so a person can jump between domains and a team can stay small. He hopes this leads to many smaller organizations of generalists rather than giant corporations where each person turns one button.

In the closing lightning round Shipper recommends War and Peace, The Death of Ivan Ilyich, George Saunders's A Swim in a Pond in the Rain and The Master and His Emissary, and says he loves Deadwood for how order forms out of chaos, a parallel to technology frontiers. His product pick is Granola. For a life motto he asked ChatGPT, whose answer inspired his real one, a line he attributes to Pliny the Younger: do things worth writing about, and write things worth reading.

He then describes a period when he stopped writing to run Every like a tech founder. The business flattened and he was miserable, because media businesses do not follow the startup pattern of hiring others once the product exists. When he asked ChatGPT for examples of writers who built companies, it named people like Joel Spolsky, Jason Fried, Sam Harris and Bill Simmons.

So he flipped his identity to put writing at the center, and says that has been good for both him and the business. Every time he has leaned into what feels like a secret desire, it has worked better, he says, because there is a huge tax in doing something daily that you do not quite fit. Finding a unique shape for your company, instead of copying what a company should look like, is both a better way to succeed and a better way to live.