He Asked for a Jedi. He Got a Marketing System.

My son wanted a Star Wars birthday party at a local indoor play place.

Seemed like a reasonable request.

We homeschool, so filling a 12-kid party room takes a little more thought than sending invitations home with a classroom. The good news is he’s great at making friends in the wild. The kid will talk to anyone.

And I know a few things about capturing leads.

He handles the networking. I handle the follow-up. Apparently we’re a team now.

About 20 minutes after he told me what he wanted, I had an MVP landing page coded. The code went into GitHub, Cloudflare handled the hosting, and a children’s birthday party officially had a tech stack.

The idea was simple. Give him a Jedi invitation card with a QR code. When he makes a friend, there’s something to hand over. Their parents can scan it, find the details and let us know they’re interested.

No trying to remember someone’s number after a conversation at the playground. No invitation disappearing into the bottom of a bag with half a snack.

Over the next few days, AI helped with the party-room booking, designed double-sided 3 × 4.5-inch QR cards and handled correspondence with local printers.

We were discussing rounded corners and card stock.

For a birthday party that was still nearly four months away.

I believe the technical term is a long lead-time campaign.

The Force Is Strong. The Scope Creep Is Stronger.

Of course, the page doesn’t just sit there looking like an invitation.

It goes from “express interest” to “RSVP” to “full” as we move through the process and fill the room.

My son’s birthday party has lifecycle stages.

Behind that is a database, a private admin dashboard and anonymous tracking to see whether people arrived through the printed card.

Naturally, the dashboard is called Mission Stats.

Then there’s the email autoresponder. Someone fills out the form, they get a confirmation. I’m not manually typing the same information every time a parent responds.

And you better believe AI is helping me handle the replies too. That’s the plan. I didn’t build all this to appoint myself full-time birthday support.

I also wanted automated WhatsApp messages as an option for parents. That part isn’t live yet. Meta threw a snag at me, and I haven’t put the effort into sorting it out.

Even the Jedi have to wait for Meta.

Did a 12-kid birthday party require a multichannel communications strategy?

Probably not.

Was I going to build a QR-code funnel and then not know whether the QR code worked?

Also probably not.

Some dads go overboard on decorations. I apparently need a conversion event.

But here’s the part I actually find interesting: I built this from my phone.

Not by spending three days hunched over a tiny code editor. My part was mostly explaining what I wanted, answering questions, reviewing what came back and asking for changes.

There were still things to check. Links needed to work. The emails needed testing. The cards needed to be readable. AI doing the work doesn’t mean I stop paying attention.

But the distance between “that would be cool” and having something usable has become ridiculously short.

And not everything has to be a business.

Sometimes it’s just a thing you want to make for your kid. Something useful, a bit over the top, and fun to build.

A few years ago, I wouldn’t have bothered building a custom app for this. Now I can follow the idea while I’m going about my day, with a few messages and decisions along the way.

That’s a part of AI I don’t think we talk about enough. Not just doing more work. Getting to make things you otherwise wouldn’t have made.

If you’re curious about the page or how it came together, email me. Happy to talk through the build without putting the kids’ details out on the internet.

My son wanted a Jedi party.

Dad built a marketing system.

May the forms be with you.

Why I Named My AI Leeroy

My AI is called Leeroy. Yes, that Leeroy.

If you haven’t seen the Leeroy Jenkins video, watch it first. This will make considerably more sense.

A group of World of Warcraft players are planning their next move. There’s a strategy. There are roles. Someone has even calculated their probability of survival: 32.33%, “repeating, of course.”

I appreciate that level of precision in a plan that’s about to get everyone killed.

Then Leeroy charges in, announces himself at full volume, and the whole thing falls apart.

Naturally, I thought: that’s what I should name the AI I work with.

Hear me out.

AI is moving fast. Most of what I learn comes from actually doing something with it. Trying an idea. Building a version. Finding out what works and what absolutely does not.

You can spend weeks comparing tools and planning the perfect setup. Meanwhile, someone else has built three versions and learned something from each one.

At some point you have to go into the room.

So yes, I want some Leeroy energy. I’d just like better than a 32.33% chance of coming back out.

Repeating the experiment is useful. Repeating the slaughter feels like a process issue.

Because an AI can move very quickly while being completely wrong about what you meant. Give it a vague instruction and enough access and it can have a remarkably productive afternoon doing entirely the wrong thing.

Impressively efficient. Completely wrong.

This is where the infrastructure matters. Not because organising folders is my dream hobby, but because the thing doing the work needs to understand the work.

What are we trying to accomplish? What have we already decided? Which information is current? What can it do on its own, and when does it need to stop and ask?

The context needs to reach the AI before it acts. Otherwise you’re just the guy running behind it yelling “Stick to the plan!”

Bro. You left the plan in another chat.

That’s the bit I’m interested in getting right. Give it the relevant information, a clear job and sensible limits. Let it move. Check what actually happened. Improve the next attempt.

Not a made-up confidence score with an impressive number of decimal places. Evidence that we’re making better decisions.

If we’re testing something we can undo, let’s go. Show me a first version. We’ll learn more from that than another discussion about what we could potentially build.

If it’s about to spend money, contact someone or make a change we can’t easily reverse, checking first is sensible. Leeroy can wait ten seconds.

But if every tiny action needs a committee meeting, we’ve built a very expensive way to remain standing outside the room.

I’m still responsible for the brief, the access and the decisions. Naming my AI after Leeroy Jenkins isn’t a liability waiver.

It does make the conversations funnier, though.

I want the willingness to charge in, with better odds of coming back out. Give it the context. Make the plan. Keep the mistakes small enough that we get another attempt.

Then stop standing outside the room.

Preferably with better than a 32.33% chance of survival.

Repeating, of course.

LEEEROOOY JEEEENKINS!

Am I getting dumber or just faster

Sunday is my mountain office day. I love hiking up Grouse, so I pack a laptop, charger, headphones and a change of clothes and take my work up with me. The weight is the point on the way up. The view is the point when I stop.

I've usually got three or four things going. Building and working in VS Code with both Codex and Claude at the same time, talking to Leeroy in the ChatGPT app and having him drive my browser on a task, sorting out accounting, working on the dashboard I use to keep track of my own stuff. And now writing this.

It's crazy how much I have my AIs doing now and how they talk to each other and sort things out that I used to be more involved in.

That's most of how I work now. Give something instructions, let it run, check it, send it back if it's wrong. There's usually a terminal still pondering something while I've moved on to the next thing.

I used to need someone else to do a lot of this. Someone who knew the platform, someone who could build the thing. Now I can get through a lot of it myself in an afternoon, without waiting for someone to be available.

But I was thinking about how much of it I actually understand.

I've never opened the PostHog dashboard. My analytics come through a dashboard I've had built, and that's what I look at. I use PostHog, but I couldn't tell you where anything is in it.

A few years ago I'd have spent hours in there learning where everything was. Now I'm looking for tools with a CLI so I can tell the model what I need and let it do the work.

It's quicker. I'm also not learning the tool the way I would have before.

I don't know if that makes me dumber. It does mean there are things I'm using every week that I'd struggle to operate on my own.

What I still have to do is read what comes back. I'll get a number that doesn't look right, or a summary of a deal that doesn't match the conversation I remember. I'll push back and it'll change its answer. That happens often enough that I don't just accept the first thing it gives me.

With those things, I've got something to check against. I know the deal. I know roughly what the number should be.

It's the stuff I don't know that makes me wonder. If I haven't learned how something works, how much am I going to miss when I check it?

I'm getting more done. I just don't think getting more done answers the question.

Own the infrastructure and test the model

Dan Martell has tested over five hundred AI tools. He publishes the ten that actually matter, rates the models, tells you what to move to. The content's good. It's also different every few months, because the answer keeps changing.

He posted about ChatGPT recently. The comments were full of people saying they'd already switched to Claude. Because of him. He'd moved them there, and now he's over here, and they're stuck with the thing he told them to buy last time.

Nobody in that thread was debating the models. They were comparing which app they'd moved into.

That's the problem. Your context lives in the Claude app. Your projects, your history, the way you've set everything up. It's all sitting inside one company's product. When the next model is better, and it will be, you've got nowhere to take any of it.

The framework I work to is 60/30/10, and that's not mine. It's Jake Van Clief's. He wrote the research paper on Interpretable Context Methodology, which is the actual engineering behind why folder structure works as agent architecture. Most people in this space are selling prompts. Jake's the one who went and did the work on why the structure around the model matters more than the model. If you're building anything with AI and you haven't gone through his stuff, start there.

His breakdown: sixty percent of anything that works is data and how it's organised. Thirty is orchestration, the rules about what happens when. Ten percent is the model call. Everyone's arguing about the ten percent.

So the organisation is the asset, not the tool. Mine is folders and markdown files in git or on my laptop. Named properly, structured so a model can find its way around without being told twice. Data can sit in Supabase, or Firebase, or something you build cheaper. Doesn't matter much. What matters is it's mine and anything can read it.

I was messing around with a Grok bot last week. No plan, just curious. Pointed it at a project I was in the middle of and it was up to speed in seconds. Nothing to migrate, no export, no rebuild. It read the same files everything else reads.

That's not Grok being clever. There was just nothing in the way.

The models will keep passing each other. Someone will post next month that the new one is better and they'll probably be right. The question isn't which one wins. It's whether you're set up to just go and use it.