Blocked at step one

This spring I spent a month teaching people who do not write code to set up their own AI agent. Friends, business owners, my father, and eventually rooms of strangers. The thing itself is simple to describe: an AI that can act on your computer and that knows enough about you to be useful. We call it a Jarvis. If you have seen Iron Man, you already understand.

I want to write down what I learned, and most of what I learned came from watching things break. The sessions that went smoothly taught me almost nothing.

the guide was clear to me

Before the first public session, I handed the workshop guide to my roommate. He is sharp and completely non-technical, which made him the perfect reviewer. He read it and said: wow, this is very complicated. If I did this myself, it would take a lot of time.

I had read that guide many times and found it clear. It was clear to me. That sentence is the whole problem, and it took me weeks to hear it properly.

The guide's authors were engineers, a data scientist and a data engineer, both generous, both good teachers. On a planning call I finally said out loud that none of us were in a position to identify what needed simplifying. We could no longer see the material from outside. When you know what a terminal is, you cannot remember what the screen looked like before you knew.

The same week gave me a number. I assumed maybe half of any audience would already pay for an AI subscription, because almost everyone I know does. A friend on the planning call estimated 10 percent. He was right and I was living in a bubble, and a workshop designed for my bubble would have died in the first twenty minutes.

The hardest and most useful feedback came from a friend in his sixties, the least technical person to ever test the guide. He got stuck at nearly every step. He did not know what cmd meant. The document he was supposed to copy commands from turned out to be read-only, so even copy-paste failed him. And then he gave me the verdict I needed, which I will translate from French: if you show Windows users a command prompt, you have lost 80 percent of the room, including young people. Even with a hyper-clear instruction sheet, three quarters of people will quit by the second step.

He was describing himself and he was right about everyone. I rebuilt the workshop around his two sentences.

teaching my father

At the end of March, news broke that Russia might cut internet access. My parents live there. My dad had been curious about what I was building, and suddenly the window to teach him had a possible expiry date. He said let's do it now. I canceled my evening plans.

He had never opened a terminal. He uses a computer for work, carefully, the way you drive a rental car. Everything on the screen looked like code to him, including things that were just folders.

We did it over Zoom with remote screen control, which I now consider essential for teaching anyone remotely. I could click the button on his screen, tell him where I clicked, and hand control back. Without that, the session would have been impossible. We renamed his project folder from Cyrillic to Latin characters so file paths would not break, a detail I would never have thought to put in a guide, and exactly the kind of detail that ends a beginner's evening.

In an hour and a half he drafted an email, generated a report, and built a small website. Then the website refused to open locally, because Python was missing from his machine.

That error was the best thing that happened all night. I showed him how to paste the error message into the chat, ask for an explanation as if he were twelve, and let the AI walk toward the fix step by step, running the commands together. He watched the machine explain its own failure and then repair it. On a call later that night I called it workshop squared: my dad was getting workshopped by me and by the AI at the same time.

He was cautious about the subscription, so we started him on the free tier, around 50 requests a month. I told him the tool could plausibly earn him more than it costs. He said, in effect, let me get dizzy first. Fair.

the inflection point

Teaching my father crystallized the one idea I now build every session around.

In the beginning, you explain every single click. Total patience. You repeat yourself without irritation, because to a beginner nothing is obvious and everything is risk. You keep doing this until a specific moment: the moment they understand that their questions should go to the AI and no longer to you.

I call it the inflection point. Before it, you are the interface. After it, they paste their own errors, ask their own follow-ups, and troubleshoot with the machine directly. Once they grasp that, you step back, and the growth from there is theirs. It compounds without you.

This also explains why detailed written instructions fail beginners. A step-by-step document is a promise that nothing will go wrong. Something always goes wrong: a missing dependency, a full disk, a blocked browser on a work laptop. The document cannot answer, and the person is stranded on step four of forty. So I stopped trying to write perfect instructions and started teaching the recovery skill instead. The error message is the curriculum.

what a room of thirty does to a plan

The first public test was a workshop for friends in early April. I expected a handful of people. Over twenty came, close to full attendance. We kept it to basics: working in an editor, voice input, one file of personal context, and a demo of my own system so they could see the ceiling. One friend took those basics and started building a website for a surfboard business he is incubating. He jokingly called me the Jesus of AI, which I am obligated to reject as theology while accepting as feedback: when one thing clicks, even basic teaching delivers real value.

Three weeks later I ran the same idea for a public room of about thirty, and the format cracked in ways the small room never showed me.

Follow-along collapses past a certain size. With ten people, "raise your hand when it boots" resyncs the room. With thirty people of wildly mixed skill, the facilitators sprint between rows while the fast attendees finish and the slow ones sink. One attendee was blocked at step one, downloading the installer. Another ran out of disk space mid-install and visibly deflated. Meanwhile someone in the ninth step sat waiting, bored. Nobody was well served by the same pacing.

The paid tools bit us too. The voice-input tool hit a paywall for part of the room. The same tool produced the strongest moment of the night: a woman spoke into her laptop, watched her words become a clean draft, and asked out loud, wait, you just said that with your voice? I now engineer for that exact reaction. Voice is the fastest door into all of this, because talking to your computer feels impossible right up until the second you do it.

One more moment stays with me. A machine learning expert, someone far more technically credentialed than me, did an exercise where the system helped him work through a real blocker in his life. He turned around and said: wow, this thing is like my psychologist. He had been considering paying for a closed agentic platform. After the workshop he owned a system where the files were his and the model was swappable. That is the sovereignty argument landing in a person rather than in a slide.

And I fumbled the logistics. My opening demo was me speaking into my phone and my agent emailing the setup document to everyone in the room. Great demo. It only reached the people we had checked in, and some attendees brought friends or walked past the check-in table. The fix went straight into the playbook: send to every RSVP, always.

what I changed

After the room of thirty, my team pushed to send installation homework before events: install these tools at home, arrive ready. I refused. At home, alone, with no one beside them, the install takes three or more hours and people give up before they ever see the good part. The failure has to happen in a room where someone can sit next to you. Setup pain is part of the workshop, on purpose, and I say so at the start: you will hit errors, that is intentional, learning to debug with AI is the skill.

Everything else changed within five days. The follow-along became phases with sync checkpoints: everyone finishes a phase, the room resyncs, someone explains what was just installed and why, the next phase opens. The big room became a talk plus demo, and the actual teaching moved to small sessions, ten to fifteen people at most, because helping someone through genuinely technical material takes a human tutor within arm's reach. We even sketched a help button attendees could click on screen instead of raising a hand, because it turns out adults are shy about being stuck.

There is a limit to what a workshop can do, and naming it was part of the lesson. I can take someone from zero to a working system and the understanding of why it matters. The rest is theirs: filling it with their own context, their own files, their own life. I can hand someone the suit. The wearing of it is on them.

What did teaching non-programmers actually teach me? That my expertise had quietly deleted the beginning of the path behind me. I had stopped seeing that "open a folder" is an instruction, that markdown is jargon, that a C drive can be full, that a paywall is a wall. Every session put one of those deletions back in front of my face.

You cannot see your own assumptions from inside your own competence. Put the work in front of someone who has never opened a terminal, and count what breaks. That count is the curriculum.