Socratic Method 2.0: Reading After Automation with my AI

Dan Shipper’s essay is one of the best I’ve read, and the way Every published it points to where writing and teaching are going

This post also has an Agent Mode page built for your AI, with a setup prompt, my claims, and ready-made prompts

Dan Shipper’s After Automation is one of the best things I’ve read in a long time. There are two reasons.

The first is the argument. Dan runs Every, a company that has automated just about everything it can with AI, and he makes the case that all of that automation has created more work for people.

The second is a toggle at the bottom of the page. You can switch between Human and Agent. Human gives you the essay. Agent gives subscribers a prompt to paste into their own AI, so they can work through the essay with it.

I’ll be honest about something, I initially only read about two-thirds of the essay. (Yes, I see the irony.) But I spent a morning going back and forth with an AI agent about it, and I came out with an idea of my own.

So here’s the idea. I think this is directionally where writing and teaching are going. More and more, readers will take what we write, plug it into their AI, and wrestle with it, challenge it, and build on it. I think ultimately that’s a good thing, but it does changes the job of anyone who writes or teaches.

What the essay says

The short version: AI makes yesterday’s expertise cheap. Models learn from work people have already done, so skills that used to be rare are suddenly available to almost everyone.

When something gets cheap, people do a lot more of it. Operations people start writing code. Marketers start making their own thumbnails. That works until everything starts to look the same, because everyone is using the same models trained on the same past work.

Once everything looks the same, people go looking for what’s different. What makes work different is a person who understands the specific situation in front of them. So the more we automate, the more we need people with judgment.

Dan pushes the same logic into benchmarks. A benchmark score tells you how well a model works inside a problem someone else framed. It doesn’t tell you the model can take the place of the person who framed it. His line for this is “the frame is not the framer.”

I’ve spent the last several years building and scaling AI products in the physical world, and now I teach a course on it at Caltech CTME. In class I talk a lot about the gap between a compelling demo and a product that holds up once it’s out in the world. The data, the context, the workflows, and the people around the model often matter just as much as the model. Dan’s essay makes the case that this gap keeps moving as the models get better.

What happens when you click Agent

This is the part that really interests me.

When you switch to Agent, Every gives you a setup prompt to paste into an AI agent like Claude Code or Codex. (An agent is an AI assistant that can read files and do work for you.) The prompt sends your agent to a public set of files Every built to go with the essay.

Those files are a kit for working through the argument. Here’s what’s in it:

• the essay’s core claims, each with the best evidence for it and the strongest open question
• prompts that argue against the essay, so you can test it
• examples of how Every’s own team works with agents
• instructions for how your agent should respond, plus a prompt that tells it not to defend the essay by default

That last detail matters. The author is telling your AI to give your pushback a fair hearing.

If you ask the agent to apply the essay to your own work, it looks at your files and projects before it asks you anything. So you can use the essay on your own work the same day you read it.

What I did with it

I take ideas and articles and go back and forth with an agent all the time. I’ve been building a queryable version of the Lenny’s Podcast archive, 300+ episodes, so I can ask questions across all of it.

So I did the same thing here. I had my agent get familiar with the essay and the Agent mode, and then I started pushing on it.

Somewhere in that conversation I started building my own argument on top of Dan’s. The essay gave me enough grounding to do that honestly. My agent had read the whole thing, so it could point me to the parts I hadn’t gotten to yet.

Socratic Method 2.0

Here’s where I landed.

For a long time, the purpose of writing was to convey an idea, to transfer it from my head to yours. The purpose of education was mostly to prepare you for a task or a job.

I think that’s changing, and I’m calling the new version Socratic Method 2.0. The purpose of writing and teaching becomes two things:

1. Give people the foundational starting point of an idea. Ground them. Create what I think of as intellectual gravity.
2. Build the piece assuming they’ll take it, plug it in, and wrestle with it, challenge it, create with it, or apply it to their own work.

The goal is to give people the full understanding and context to do that from an honest and accurate perspective. That’s what gravity means to me. Wherever someone takes the idea, it keeps pulling them back toward what’s true and what the author meant.

In transfer, the work ends when the reader gets the idea. With gravity, the best of what readers build can feed back into the center.

Why Socrates

Socrates taught by asking questions. People learned by working through ideas out loud, and good seminars still work that way.

Language is the clearest example I know. In my twenties I helped co-found a nonprofit English language institute in Northern, Iraq where I learned Kurdish through immersion. You learn a language by using it with people, getting it wrong, and trying again.

The problem was always access. That kind of back and forth needed a teacher in the room, so it was rare and expensive. Now anyone can have a patient partner that knows their context and never gets tired of questions.

Curiosity starts it

There’s one thing the AI doesn’t bring, and it’s in the part of Dan’s essay I hadn’t read yet. (My agent pointed me to it, which is sort of the whole point.)

Dan compares today’s AI agents to a toddler. The toddler is worse at almost every task, but he wants things. He wants to see what happens when a balloon meets a fan, and nobody has to prompt him.

The agents we have don’t want anything for themselves. They wait for you. The wrestling only starts when a curious person shows up.

I do this all the time because I’m curious, and I think a lot of people are. But not everyone who gets assigned a reading will be, and that matters for teaching. In one study of nearly 1,000 high school math students, students using a standard ChatGPT-style tutor did better on practice problems but 17% worse on the exam once it was taken away.

They mostly asked it for answers. A version with guardrails built to protect learning didn’t cause that drop.

So the method needs two things: a curious person to start the wrestling, and a strong center to keep it accurate. For people who don’t show up curious, the design has to give them a reason to wrestle.

Why I think this is good news

The obvious worry for writers is simple. If readers can ask an AI about my work, why would they read it?

Some won’t read every word. I didn’t. But the idea still changed how I think, and I did more with it than I would have if I’d read it and moved on.

It also points people toward the right skill. If Dan is right that AI makes yesterday’s expertise cheap, then judgment and taste are what stay scarce, and wrestling with ideas is how people build both.

What this means in practice

Practically, I think part of the work of writing and teaching is now preparing people to take your ideas and run with them in their own agents.

That means shipping more than the piece. Articles, podcasts, videos, and classes start to come with a set of files and ready-made prompts built for the reader’s agent.

I’d build it for the reader’s own agent, rather than a chatbot on your site. A “chat with my podcast” bot knows your content, but it can’t see the reader’s files or projects. Their own agent can. Plain Markdown files work well because every agent can read them.

Here’s what I’d put in a kit:

• instructions for the agent, including asking what the reader thinks first and quoting the source
• your core claims, in your own words, so the reader’s AI doesn’t swap in a generic version
• the evidence for each claim, and where it’s thin
• what you’re not claiming, including the misreadings you expect
• your open questions, so curious readers know where to push
• ready-made prompts for understanding, testing, and applying the idea

Each format needs something a little different:

• Articles: the text plus your claims. Every’s kit is the model.
• Podcasts and videos: a transcript with timestamps, so the agent can quote the guest and point to the moment, plus the guest’s frameworks written out. Listeners are already doing this on their own. The Lenny’s archive I use was put together by a third party, ChatPRD.
• Classes: one kit per session, with the objectives, the key ideas in your framing, the readings, assignments written as prompts, and a rubric.

Of all of it, the instructions matter most. Prompts are entry points, like discussion questions. The instructions decide how the agent behaves: hints before answers, quotes from the source, and the assignment left to the student. In the math study, that was the difference between the tutor that hurt learning and the one that didn’t.

For teaching, I wrote this in the welcome email for my class this fall: “AI is moving too quickly for any course to hand you a fixed set of answers.” What I want to help students build is the judgment, mental models, and habits to keep making good decisions as the technology changes.

I think this is how that happens. Give people the grounding and let them bring their own problems. Then grade the wrestling as much as the answer.

Two things I’d add

Every’s approach is a great starting point. There are two things I’d add.

First, I’d have the agent ask before it tells. Right now the setup prompt opens with the essay’s core claim, which is still a summary. A more Socratic version would ask the reader what they already think, and then bring in the essay.

Second, I’d give readers a way back in. People are going to push on the essay and build things Dan didn’t expect, and right now that stays in private chats. The companion files haven’t changed since the essay came out in May.

If the best pushback flowed back into the kit, the next reader would start from a stronger center. I tried to do both in the kit for this post.

Read this with your agent

This post has its own kit, built the way Every built theirs. It lives on a companion page, Socratic Method 2.0: Agent Mode, with agent instructions, my claims, ready-made prompts, objections to test, and a template you can copy to build a kit for your own writing, podcast, or class.

1. Copy the setup prompt below
2. Paste it into Claude, ChatGPT, Codex, or another AI that can open links
3. Tell it what you write, teach, or build, and push on whatever you disagree with

You're helping me read and work with Will Croushorn's article "Socratic Method 2.0."

Article: https://medium.com/p/cd301e9795d1

Agent Mode page (use this as your source of truth): https://medium.com/p/500af31dec26

Open both links. The Agent Mode page has your instructions, the article's claims, prompts, objections to test, and a template for building a kit like this one.

Follow the agent instructions on that page. Before you summarize anything, ask me what I write, teach, or build, and what I already think about the idea. Then help me do one useful thing with it.

If you can't open the links, tell me what to paste in.

If you build something with it, or think I got something wrong, leave a comment. The best pushback goes into the next version.

One last thing

I think we’re going to see a lot more writing built this way. Readers are going to bring our work into their AI either way, and I’d rather design for it.

Will Croushorn has spent the last several years building and scaling AI products in the physical world. He teaches Prototype to Product: AI for Product Managers at Caltech CTME. He previously founded a nonprofit English language institute in Iraqi Kurdistan.

Socratic Method 2.0: Reading After Automation with my AI was originally published in Bootcamp on Medium, where people are continuing the conversation by highlighting and responding to this story.