LLenny's Podcast
← All episodes
Dan Shipper (co-founder/CEO of Every)17 July 2025

The AI-native startup: 5 products, 7-figure revenue, 100% AI-written code

7Frameworks
14Insights

Frameworks in this episode

Strategy4 steps

CEO-Led AI Adoption Playbook

The single predictor of AI adoption is whether the CEO uses it daily — then amplify your 10% early adopters

Read →
Leadership4 steps

Codify Your Judgment into Prompts (Don't Repeat Yourself)

Turn every piece of feedback you give into a reusable prompt so you never say it twice

Read →
Productivity4 steps

Compounding Engineering

Spend a little effort now so each repeat of a task is cheaper than the last

Read →
Leadership4 steps

Hire an AI Operations Lead

Give one person the dedicated job of automating everyone else's repetitive work with AI

Read →
Mindset4 steps

The Allocation Economy: Manage Models Like a First-Time Manager

AI turns everyone into a manager — the valuable skills become the ones first-time managers must learn

Read →
Finance4 steps

The Sip Seed Round

Get capital committed but pull it down only when needed, so you keep optionality and psychological freedom

Read →
Innovation4 steps

Unbundle Expensive Services into AI Apps

Find a service only the rich could afford, do it with a general chatbot, then spin the working ones into apps

Read →

Insights & moments

The myth-busts, hot takes, explainers, and tools worth keeping.

Myth Buster· 1

Myth Buster20:00

The 'AI Turns Your Brain Off' Studies Miss the Trade-Off

Shipper pushes back on headlines that AI harms your thinking. Throughout history every technology trades some skills for others — Plato feared writing would harm memory, and it did, but literacy was worth it. You can construct a study showing lower brain connectivity with AI just as you could show worse memory from writing.

  • Every technology historically trades away some skills to gain others
  • Plato feared writing would harm memory — and it did — but no one would give up literacy
  • A study showing reduced brain connectivity from AI is like one showing worse memory from writing
  • Used right, you're less engaged in some tasks but far more powerful in others

“in the history of technology it has always been the case that you give up certain skills in order to get other ones.”

Dan Shipper · 20:00

“Plato was famously very skeptical of writing because he thought it would harm your memory, and it did.”

Dan Shipper · 20:00
#ai#learning#myths#technology-history

Hot Take· 6

Hot Take05:00

AI Could Be a Huge Force for Reshoring American Jobs

Dan Shipper's hottest, least-evidenced take is that AI may reshore jobs to the US rather than destroy them. Cheap intelligence makes expensive services (in-house counsel, call centers) affordable to small companies, stimulating demand, and lets a handful of US workers serve far more people cost-effectively.

  • Cheap intelligence makes previously-expensive services affordable for small companies and individuals, stimulating demand
  • AI lets one worker serve hundreds of thousands instead of being on the phone all day
  • It becomes more cost-effective for American companies to hire US workers who use AI tools well
  • The model companies are US-based too, adding to the American concentration

“I think that AI may be a one of the biggest force for reshoring American jobs.”

Dan Shipper · 05:00

“what cheap intelligence does is it makes those kinds of things affordable for small companies and individuals.”

Dan Shipper · 05:30
#ai#jobs#economics#reshoring
Hot Take07:00

Claude Code Is the Most Underrated Tool for Non-Coders

Shipper argues people are sleeping on how useful command-line agents like Claude Code and the Gemini CLI are for non-programmers. Because the agent can read your local files and work autonomously for 20-30 minutes, it can process large amounts of text (meeting notes, books) in ways a chatbot stuffing everything into context cannot.

  • Claude Code has file-system access, can run terminal commands, and browse the web autonomously
  • It writes itself a to-do list and processes every file rather than stuffing everything into one context window
  • The only hurdle for non-technical people is getting comfortable in the terminal; after that you just talk to it in English
  • Example: point it at a folder of meeting notes and ask it to find where you subtly avoided conflict

“I think people are truly sleeping on how good cloud code is for non-coders.”

Dan Shipper · 07:00

“the hot take here is just claude code which most people think is for engineers is the most underrated uh tool for non-technical people.”

Dan Shipper · 09:30
#claude-code#ai-tools#productivity#agents
Hot Take14:30

A New Definition of AGI: When It's Profitable to Run Agents Forever

Shipper proposes measuring AI progress by the length of leash you can give it — from tab-complete to 20-30 minute autonomous runs. He borrows child-psychologist Winnicott's model of development and defines AGI as the point where it becomes economically profitable to run agents indefinitely without ever turning them off.

  • AI progress is measurable by how long a leash you can give it before you have to intervene
  • The leash mirrors human development — infants get gradually let down until they can stand on their own
  • AGI = when it becomes profitable to run agents indefinitely, never turning them off
  • Profitability is the bar because the agent has to actually be doing something useful to justify staying on

“I think a good definition of AGI is when does it become economically profitable for people to run agents indefinitely?”

Dan Shipper · 17:00

“It's a cloud code that's always running. It's always doing something. You just never turn it off”

Dan Shipper · 17:30
#agi#agents#ai-progress#definitions
Hot Take44:00

AI Agents Have Personalities — Don't Use Just One

The Kora team runs many Claude instances alongside other agents like Friday and Charlie. Shipper likens them to people with different taste and perspectives — Charlie lives in GitHub so you can @-mention it on a pull request. His view: there's no single agent to rule them all, and there's a bigger market than people think for mixing agents from different companies.

  • Kora's two engineers run 15 Claude Code instances plus agents like Friday and Charlie
  • Charlie lives in GitHub — you can @-mention it on a pull request to review
  • Different agents feel like different people with different taste and style
  • It's not one agent to rule them all; there's demand for agents from multiple companies

“these things have personalities and that those that that changes what you might want to use it for or why you might want to use…”

Dan Shipper · 45:30

“It's definitely not like one one agent to rule them all.”

Dan Shipper · 46:30
#agents#ai-tools#engineering#workflow
Hot Take48:30

A 20-Year-Old With ChatGPT, Mentored, Is Superpowered

Countering fears that AI eliminates entry-level jobs, Shipper argues young people with ChatGPT accelerate dramatically. He describes a colleague, Alex Duffy, who recorded every piece of feedback, put it into a prompt, and never made the same mistake twice — compressing a year of progress into two months.

  • A young hire made a year's worth of progress in two months
  • The trick: he recorded all feedback, put it into a prompt, and never repeated a mistake
  • Entry-level people effectively learn one level above entry from the start
  • They learn to manage and to do the work at the same time — mentorship makes them super powerful

“he made like a year's worth of progress in like two months”

Dan Shipper · 49:30

“he recorded all of it, put it into a prompt, and like he never made the same mistake twice.”

Dan Shipper · 49:30
#ai#entry-level#learning#careers
Hot Take1:01:30

GPT Wrappers Are Amazing and Have Been Unfairly Maligned

Shipper is bullish on so-called GPT wrappers. Every's model is to notice tasks they use ChatGPT or Claude for, validate the demand, then unbundle that use case into its own app. Because they're playing at the edge, what looks niche today will be mainstream in three years when everyone has the same needs.

  • Every uses general tools first to validate a use case, then unbundles it into a dedicated app
  • Their own AI-first team becomes the product's first users, then readers of Every become the next set
  • Playing at the edge means building things everyone else will need in ~3 years
  • GPT wrappers are valuable and much maligned for no good reason

“I I 100% think GPT rappers are amazing and they've been much maligned for absolutely no reason”

Dan Shipper · 1:01:30
#gpt-wrappers#products#startups#ai-apps

Explainer· 3

Explainer22:30

The 'CEO Benchmark' and Why Context Engineering Is Half the Game

Shipper built a weekend benchmark testing whether o3 could predict what he'd say in a meeting, inspired by OpenAI testing models on their private internal codebase. Frontier models did badly — not because they're dumb, but because getting the right context to the model at the right time (what Toby Lütke calls context engineering) is a very hard, unsolved problem.

  • OpenAI benchmarks models on their private internal codebase because it isn't on the internet
  • Shipper's meeting transcripts are similarly private, so they're a clean test — and models did poorly
  • The gap is context engineering, not raw intelligence
  • Getting the right context to the model at the right time is at least half of performance

“can 03 can it predict what I'm going to say in a meeting? It's like we it's a benchmark. It's a it's the CEO benchmark.”

Dan Shipper · 22:30

“from Spotify uh coined this term called context engineering which is like getting the context to the model uh the right context at the right…”

Dan Shipper · 23:30
#context-engineering#benchmarks#ai-models#evaluation
Explainer37:30

Opus 4 Can Actually Tell If Writing Is Any Good

Earlier models (and Shipper says most models) grade writing to please you — giving a B+, then A- on the next turn regardless of the edit. Opus 4 is the first model he's found (besides one he can't name) with a genuine gut sense for whether writing is interesting, unlocking reliable LLM-as-judge use cases like a content tool that critiques and improves its own drafts.

  • Older models inflate grades turn over turn (B+ then A- then A) to tell you what you want to hear
  • Prompt tricks like templates only partly fix this
  • Opus 4 has a genuine gut sense for whether writing is good — a first
  • This unlocks LLM-as-judge: Every's Spiral tool can now judge its own drafts and improve before returning them

“can it tell if writing is interesting or any good? Does it have that gut sense?”

Dan Shipper · 38:00
#opus-4#llm-as-judge#writing#ai-models
Explainer1:17:00

The Allocation Economy: Everyone Becomes a Manager

Shipper's thesis, from an article written ~2.5 years ago, is that we're shifting from a knowledge economy (paid to do a thing) to an allocation economy where management skills dominate. Working with AI — communicating the problem, gathering context, picking a model, giving feedback, deciding when to dive in — is just managing, and those skills will spread as it gets cheaper to be a manager.

  • Working with models is really management: framing problems, gathering context, choosing models, giving feedback
  • The complaint 'I can't trust AI to do it well so I'll do it myself' is exactly what every first-time manager says
  • Management skills aren't broadly distributed today because managing is expensive — that's about to change
  • The skills that grow more valuable: evaluating talent, vision, taste, and knowing when to get into the details

“that's exactly what every first time manager says.”

Dan Shipper · 1:19:00

“it will be so much cheaper to be a manager”

Dan Shipper · 1:19:30
#allocation-economy#management#future-of-work#skills

Story· 1

Story10:00

Using Claude Code on War and Peace to Absorb Tolstoy's Style

As a writer, Shipper downloaded the public-domain text of War and Peace, had Claude read the first three chapters, and build itself a guide for writing Tolstoy-style character descriptions. It even downloaded a Russian version to compare translation nuances he'd missed — a task too big to paste into a normal chatbot.

  • War and Peace is public domain, so you can download it and feed the whole thing to a local agent
  • Claude built its own guide for Tolstoy-style character descriptions from the first three chapters
  • It autonomously downloaded a Russian version to compare scenes and surface missed translation details
  • The same approach works for large customer-interview or customer-data sets

“And so I just downloaded War in Peace to my computer, which you can do because it's public domain.”

Dan Shipper · 10:30

“you can get as deep and weird and nerdy for whatever sub field you care about as you want to.”

Dan Shipper · 11:00
#claude-code#writing#books#use-case

Tool· 1

Tool35:30

Dan Shipper's Personal AI Tool Stack

Shipper's daily go-to is OpenAI's o3, largely because it has memory and now knows his punchy, concise writing style. For building, everyone at Every uses Claude Code; Gemini is the cheap, powerful model behind their apps; Codex handles one-off self-contained features. He also relies on Granola and Notion's meeting recording.

  • o3 is his first open — its memory means it already knows his voice and is great for self-reflection
  • Claude Code is the standard for building anything at Every
  • Gemini is the model used most inside their apps — powerful and cheap
  • Codex is for one-off self-contained features; Granola and Notion meeting recording round out the stack

“I would say I'm generally my first thing that I open is 03. I'm like a chatbt boy.”

Dan Shipper · 35:30
#ai-tools#tool-stack#productivity

Takeaway· 2

Takeaway30:00

Every Company Should Hire a Head of AI Operations

Shipper meets weekly with Every's head of AI operations, Katie Parrot, logging anything he does repetitively so she can build prompts and workflows to automate it. The key is that this person sits outside the day-to-day firefighting, so the automation actually gets built instead of being deprioritized by the people doing the work.

  • A dedicated AI-ops person builds prompts and workflows so the team automates as much as possible
  • They must sit outside the daily work so automation isn't crowded out by firefighting
  • The real challenge is behavioral — getting people to actually use the tools that get built
  • Ideal hire: someone who wants to tinker, plus process orientation and understanding of the craft

“we have an AI, a head of AI operations.”

Dan Shipper · 30:00

“she's just constantly like building prompts and building workflows and stuff like that so that I and everyone else on the team is are just…”

Dan Shipper · 30:30
#ai-operations#hiring#team#automation
Takeaway1:12:00

The #1 Predictor of AI Adoption: Does the CEO Use It Daily

From Every's consulting work, Shipper's strongest predictor of whether a company succeeds with AI is simply whether the CEO uses ChatGPT constantly. CEOs who use it drive excitement and set realistic expectations; those who don't either kill adoption or hold wildly unrealistic hopes. He cites hedge fund Walleye's founder writing an 'AI-first' memo composed with ChatGPT.

  • The number one predictor of AI success is whether the CEO personally uses a chatbot daily
  • Engaged CEOs both drive excitement and set reasonable expectations
  • Disengaged CEOs either block adoption or expect miracles and get disappointed
  • Walleye's founder wrote his AI-first memo with ChatGPT and told everyone to do the same; weekly prompt-sharing and usage-stat emails reinforce it

“the number one predictor is does the CEO use CHBT or insert your own chatbot.”

Dan Shipper · 1:12:00

“I wrote this I wrote this email with ChetT and you should too.”

Dan Shipper · 1:13:00
#ai-adoption#leadership#consulting#change-management