This book promises four things: that you can push back on a model, build a small tool for yourself, hand ongoing work to an agentAgent An AI that carries out a task on its own, step by step, instead of only answering in a chat. It can open files, run programs and check its own work before it reports back. with rules, and form your own picture of what is coming. I did not learn them in that order, and I did not learn them quickly. This is the route I actually took, with the step where you can start.
1. Pushing back
I started with ChatGPT like everyone else, and I did not see it coming. "Had you asked me in the summer of 2022 if I thought AI was right around the corner [...] the only answer would have been 'no'."
My first serious try was to make it write like me. I trained a custom GPT on twenty years of my blog posts and fed it 42 of my newsletter intros. "and yet 10/10 suggested outputs were crap. Uninspired, flat and obvious." So I kept writing the intros myself, and I still do.
The first habit that lasted was a question at the end of every answer: "what else should I know?" The second was noticing that the model agrees with me. "Language models are inherently sycophantic [...] so they struggle to push back when we need them to." So I learned to push back first: "no i dont think this holds water. too complex and not specific enough." And when it answered from a guess: "Don't guess. Look it up."
The last step came with agents. A second agent grades the first one's work without seeing it. When they disagree, nothing gets published.
Where you start: end your next three chats with "what else should I know?", and once a day tell the model it is wrong when you think it is. See what it does. (Thinking)
2. Building a small tool
My first app was a tarot reader, at the end of 2023. "A week, ten days of work went into building this thing. Absolutely from scratch in, like, the worst possible way, which is, like, pasting things between files." I asked ChatGPT for "the whole new file" every time, because I could not edit code.
In 2024 I moved to Claude and then to Cursor, an editor with the model inside it. "I went through 500 requests developing on Cursor last week." In 2025 I shipped an iOS app: "I haven't manually written a single line of code in this app. [...] You just need time & patience." That year a command from a chat wiped my changes, and learned that "Unsupervised coding agentsCoding agent An AI agent that works inside the files of a software project. You describe a change in plain words, and it writes the code, runs it and fixes what breaks. will turn anything to spaghetti with great confidence."
What changed things was splitting the work. ChatGPT helped me think, then wrote the brief: "give me natural language instructions of what to build and where, as you dont know my architecture or app intimately." The agent built from the brief. By early 2026 I was on Claude CodeClaude Code Anthropic's coding agent. It runs on your computer, reads and edits the files of a project and runs commands for you., paying "about €100 maybe a month in AI apps", and an app went from mockup to the App Store in a weekend.
It still breaks. A display board went black while an agent flashed it. "how the hell can i program if this can happen?" The rules I keep now came from moments like that: never guess, back up first, test it myself on the device.
Where you start: the brief, not the code. Talk the idea through in the chat you already use, ask it to write instructions for a builder, and paste them into one. (Coding)
3. Handing over ongoing work
This took me the longest, and most of my first tries failed.
In 2023 I tried to have GPT and ZapierZapier A service that connects apps and runs steps between them automatically. summarize articles for the newsletter. It "ultimately wasn't worth the hassle." In early 2026 I bought a Mac mini for OpenClaw, an agent that ran on its own: a morning briefing, tasks from every meeting and every email. "It feels less like flipping a switch and more like onboarding a new coworker. One who's eager but needs constant guidance." After about ten weeks it was "much to rickety". Some of the little apps it led to are the ones I still use.
The first job that ran without me and did something I cared about was saving my code. In July 2026 I asked an agent to commitCommit A saved, named version of your code. Each commit goes into the project's history, so you can always go back to it. and push every project every half hour, 42 of them. Then I asked about it every day for eleven days. "this isnt working, is it?" It wasn't: for three days it skipped every run and still looked healthy. Later it flooded my password store on two Macs. "I KEEP GETTING THE BLOODY MODALS!!!!!!!!!! ON TWO DEVICES NOW." In October I switched it off: "first of all remove the auto git". What it left me was a rule I still use: leave every project in a working state, because something may save it as it is.
The one that worked was smaller and had an exit. For a festival I helped with, a job pulled the programme from the organisers' sheet every 30 minutes, because I did not want to "1) be in charge 2) run the risk of antyhing breaking on calendar". It had a gate that stopped any change that would break people's calendars, and it told me in plain words what it changed. The gate stopped it on the first evening, and I saw it from the message.
Only after that came the factory: at six every morning an agent picks issues, works them, and leaves me a report and its questions. The first night it sat for three hours waiting for answers. The rule that came out of it: ask me only when the choice is hard to undo, otherwise take the safe default and tell me in the morning.
Where you start: pick one small job that only reads or copies, never deletes. Make it tell you in words what it did. Check it every day for two weeks, and decide in advance how you will switch it off. (Delegating, How I run my agents)
4. A picture of what is coming
I have worked in foresight since 2010, and the centaur idea is older than ChatGPT for me. What changed is how I keep my picture current. First a weekly newsletter, since April 2023: "We (still) can't automate good thinking." Then asking several models the same question: "I'm using language models not at all as an oracle, but more like crowdsourcing." Then building: "Keeping up with AI is a building problem, not a consumption one."
My picture is not fixed. In January 2026 I said "I used to be more optimistic", about agents that reach into everything on a phone. And this year I had our own old forecasts graded against what happened. Fewer than half came true.
Where you start: write down three things you expect in your field in a year, with a date. Ask two models what they expect, and keep both answers. Check them all on that date. (Navigating, Deciding)