My AI stack
I do not write the code. I direct it, and I have been here since before it had a name.
In 2016 I wrote a law thesis about unsettled questions in cyber warfare. In February 2022 I got early access to MidJourney v1 and understood what was coming. In November 2022 I signed up for ChatGPT on the day it opened, and I have used AI every day since. In January 2025 I stopped writing code and started directing it. This is what that actually looks like: the tools, the method, and what I think happens next.
··· days
Using AI, since launch day
···
Age of the AI coding era
The arc
Ten years of paying attention to this.
Two of these were turns rather than steps. Both times the thing in front of me was crude, most people around me were talking about its flaws, and I could see the slope it was on.
2016
A thesis about a technology the law had not settled
I wrote my bachelor thesis on cyber attacks and the use of force: the legal challenges of jus ad bellum in cyber warfare. International law was not absent. The first Tallinn Manual had appeared in 2013, an interpretation of how existing law applied rather than a law of its own, and what was unsettled was how established principles translated to a medium that could carry espionage, disruption, sabotage and an armed attack alike. What interested me was the gap between what technology could already do and what anyone had thought to regulate. That gap is still where I like to stand.
2017
Language models before they were called that
QuillBot, which predates GPT-1. Crude, narrow, and enough to make me pay attention.
2021
The first useful wrappers
Jasper, launched that January, and Rytr beside it, both built on the GPT-3 family OpenAI had opened through an API the year before. Still primitive by any standard we would use today, but the direction was unmistakable. Nobody I knew was aware this existed, so I read obscure forums to learn anything at all.
Feb 2022a turn, not a step
MidJourney v1, and the moment it clicked
I got early access. I had seen the DALL-E demos and was impressed, but the moment I tried it myself I was sold. I could have an idea, describe it, and watch it arrive in minutes. Today it takes seconds. That was when I understood we were at the beginning of the beginning, and even then I had no idea how fast it would move.
Nov 2022
ChatGPT, day one
I signed up the day it opened to the public. I have used it, and everything that came after it, every day since.
Q4 2024
Two things happened at once
o1 arrived: the first model that thinks before it answers. And so did the editors, Bolt, Replit, Cursor, Windsurf. Separately either would have been interesting. Together they were a new job.
Jan 2025a turn, not a step
I started building, and did not stop
At that point I barely knew what a terminal was. The tools were about as good as MidJourney v1 had been, and everyone I talked to focused on the flaws. I did not care, because I had seen that shape before. I decided early that I would not spend my time learning to write code, not because it is not valuable but because AI would improve faster than I could learn. So I learned the other half instead: git, deployment, databases, what a build actually does, enough of the language that the machine and I understood each other. For months, pnpm run dev was the phrase I typed most.
Late 2025
When Claude Code changed my own work
This is when Claude Code and Opus 4 became how I work, not when either was released. It is the point where this stopped being assisted coding and became directed coding, and it is hard to describe to anyone who was not doing it before and after.
Now
First thing in the morning, last thing at night
I build every day. It is the first thing I do when I wake up, the first thing when I get home from work, and the last thing before bed. I am a builder at heart, and that is why this technology speaks to me like no other: it raises the floor, so anyone with an idea and the will to chase it can reach the sky.
How I actually work
A terminal, two agents, and a harness I wrote myself.
I decided early that I would not spend my time learning to write code, because AI was going to improve faster than I could learn. So I learned the other half: what a build does, how a repository is shaped, what to check before something ships, and how to give an agent enough context that it does the right thing without me watching every line.
The harness
A request crosses in. Nothing leaves as a diff until the hooks and the checks have each had a turn.
What I do not do
I do not write code by hand and I have not for a long time. No computer science degree, no years on an engineering team. On code quality I cannot be your last line of defence, and I say so before anyone has to ask.
What I do
I run agents from a terminal. Claude Code and Codex, most days, side by side, on repositories I own. I decide what gets built, in what order, against which constraints, and I read every diff that lands.
What I own
The harness around the models: my own skills, hooks, and workflows. That is the part that survives a model change, and it is the part that is actually mine.
The stack
Four tools I run, and one lab I read.
A wall of logos proves nothing. Four are tools I have used long enough that losing one would change how I work. The fifth I do not use at all, and it is here anyway.
Where this is headed
I have said all of this out loud.
I am a job and business consultant at Falck and I sit on our internal AI team in labour market services, so I argue about this most weeks with people who have to act on the answer. Before that I built StudieAkademiet and sold it in 2022. I have used this technology every day since ChatGPT opened. Some of it I have already changed my mind about once.
What changes for one person is not that the work gets faster. It is which part of the work is scarce.
When everyone has a tutor with more expertise than most textbooks, a year of AI-assisted learning can cover more ground than a five-year degree did. The knowledge does not stick the same way, and I am not convinced it has to. Learning a new domain quickly now looks more valuable than knowing one deeply, because the model already knows most domains deeply. What it does not know is when to apply what.
Good ideas come from iteration, lived-in understanding of constraints, and experience with the subject. A language model, almost by definition, returns the average of what has been said. Asking it for the idea walks you straight into the most crowded space there is. It is a fine partner in ideation and a terrible engine for it.
For what to cook tonight, ask away. For anything that needs to be novel, if the model does the thinking, the output is the opposite of novel.
Your prompts, notes, procedures, examples and saved context become part of your professional skill. If that system lives only in a company account, you leave without a large part of what made you good.
AI can make weak work look professional. So the signal moves to visible reasoning: which alternatives were rejected, which risks were named, which judgments can be defended, and what actually happened afterwards.
A beautifully written essay used to be the proof. Now the proof is whether the student can defend the argument and improve it.
There is real value in treating your system as if it were alive, but only while you know it is not, or you end up somewhere unhealthy. Speaking to a model in anger is technically free and practically expensive: it spends its attention managing you instead of answering you. This is anecdotal. Try it yourself.
Six ways this fools everyone
I fall for all six of these.
Knowing the name of a bias does not switch it off. I have anchored on a benchmark, believed a wrong answer because it sounded sure of itself, and stopped watching an agent that had been right all week. The dates are the part I did not expect. The newest of these studies is from 1999 and the oldest from 1951, and all of them describe people being fooled by something far simpler than a model. I keep wondering what we will name the ones that are only possible now.
A new model drops
They had people spin a rigged wheel, then guess what share of UN countries were African. Wheel lands on 10, median guess 25. Wheel lands on 65, guesses jump to 45. A number everybody knew was random still moved the answer.
The press release leads with one benchmark, and that figure quietly drags every judgment you make afterwards. The only number that matters is whether the model made your work better.
Why hallucinations fool us
Students were given identical information. Some were told it came from a credible source, some from an untrustworthy one. Same words. The credible version landed as more persuasive.
A model hands you a fabrication in exactly the confident, well-dressed language it uses for the truth. The voice sounds like an authority. That has nothing to do with whether it is right.
Letting an agent run
People paired with an imperfect automated assistant sometimes followed its wrong advice, and missed problems it failed to flag. It helped with some of the work and quietly changed what the humans bothered to notice.
Once a system runs smoothly for a while, we check it less. The smarter it looks, the easier it is to stop watching.
The way you write the prompt
People were shown 2, 4, 6 and asked to find the rule. Most only tested guesses that confirmed their first hunch. Six of 29 got it without a wrong guess first.
Ask why remote work is destroying productivity and the model will happily build you that case. A leading prompt comes back as a polished, well-argued report.
AI and jobs
They asked whether English words more often start with K or have K as the third letter. Most said start with, because those come to mind faster. K is more common in third place.
A slick video shows ten minutes of expert work done in ten seconds, and a whole profession feels like it is about to vanish. One unforgettable demo is not a typical Tuesday at work.
The hype cycle
People judged which of a few lines matched in length, an almost insultingly easy task. When actors confidently gave the same wrong answer, real participants caved about a third of the time.
When everyone is talking about the same product, popularity starts to pass for quality. Attention breeds attention until you cannot tell consensus from evidence.