This website uses cookies

Read our Privacy policy and Terms of use for more information.

The tools got easy. The thinking didn't.

A few weeks ago my daughter's school had a field trip to the CLT Startup House. About 30 kids aged 12 to 15, a couple of teachers, a few parents, and some donuts in the back. Most of the room had never built a piece of software in their life or knew the first thing about building.

A little under 2 hours later they were standing at the TV demoing games they'd made. A two-player retro basketball game called Street Hoops 95 with pixel sounds and a shot meter. An open-world dragon RPG in a neon palette. A boss-fight game where killing a boss drops an item that slows enemies down. A racing game set across a world tour. When I asked each group whether they'd done anything like this before, three separate kids said "first time."

I've run a version of this exercise with product teams at companies a hundred times their size. It's the same loop, and the kids got to a working thing faster, so I want to walk through what I actually taught them and where it broke.. because it broke in the same three places it breaks for the grown-ups.

1. The loop is the whole thing

I gave them one idea to leave with, and it's five steps.

Interview me. Make a plan. Poke holes in the plan. Build it. Check the work.

The first step is the one nobody does. When people freeze in front of AI it's usually because they don't know what to say or how much detail is enough, and the fix is to hand that problem back: “interview me first, ask me whatever you need to know.” Then it asks the questions, and answering questions is much easier than composing a brief from nothing.

The second step is the one people skip. If you say "make me a game" you get somebody else's game. Make it write the plan first, the same way you'd make a person write one.

The third step is the one that saves you. Once it has a plan, tell it to find three to five problems with its own plan (aka an adversarial review where it pokes holes in it’s own reasoning). I explained "adversarial" to a room of kids by asking what an adversary is, and one of them said "an enemy," which is right. You want it hunting for the missing board in the table before it starts cutting wood. It is better at finding its own holes than you are at finding them for it.

Then you build, then you check the work, and the checking shouldn’t be optional. Just telling it “Build me a thing” and trusting it’ll get it right without proper planning, context, and review is why a lot of folks thing AI “isn’t really that good.” Spoiler, it is. If you use it right 😉

Build the ugly version on purpose.

2. "Explain it to me like I'm 12, because I am"

The part I pushed hardest on was reading the plan. It's the step where people turn their brain off and treat whatever came back as true because it's written confidently. They also just get lazy and don’t want to read so they just type “approved” and hope for the best.

So I told them: read it, and if it comes back as mumbo jumbo you don't understand, tell it “explain this to me like I'm 12, because I am.” One of the few things AI is genuinely better at than most people is explaining something at whatever level you ask for. Adults won't ask, because asking feels like admitting you don't know. Twelve-year-olds ask immediately.

That gap is the difference I see between the teams that get value out of this and the teams that don't. The ones who don't are getting output they can't evaluate, so they either accept all of it or reject all of it, and both are the same mistake.

At the end I asked one group what their biggest takeaway was, and a kid who'd never done this before said it taught her how to be more specific, and how to get what's in her head into a form somebody else can understand. That's the skill. It was never the tool.

3. Everything broke, the way it always does

Three things went wrong, and every one of them has a boardroom equivalent.

The logins. A chunk of the room couldn't get into the tool at all, because they were under the age limit and it said so. Their devices are locked down by the school too. So a group would be mid-idea and just stuck, staring at each other, and the fix was somebody nearby with an unlocked browser doing the typing. That's every enterprise rollout I've ever watched.. the access problem shows up before the AI problem, and it doesn't feel like an AI problem so nobody plans for it.

The credits. We ran the whole thing on free tiers on purpose, so I could show them what they'd actually hit, and we hit it. Tokens run out per day and per week, and when they run out mid-build you are done for the day whatever you were in the middle of (unless you spend more money).

Which leads to the thing I'd want you to take from this more than anything else: we built the ugly version on purpose. Low-poly graphics, prototype styling, no polish. I explained a prototype as a styrofoam version of a table you're building so you can check the height and the shape before you go buy the wood. Spending your budget making the first version pretty is how you run out before you've learned whether the thing is any good.

The plan-versus-build toggle. The tools usually have a toggle for whether it's planning or building, and defaulting to build is how you burn a day. Slow down at the front and the back half gets cheap.

4. Building was never the hard part

Near the end a kid asked how you get your game onto the App Store, which was my favorite question of the day. So we talked about business accounts and legal review and why publishing to the web is the easy path and the app stores are not.

That's the conversation I have with founders in this house every single week and with mid-market clients at StealthX, just with bigger numbers. Building got easy. Getting one person who isn't your mom to use the thing did not, and "if you build it they will come" is still a lie (darn you Kevin Costner!).

The other thing I told them, and I mean it for you too: on the free tools, you're the product. Go into settings and turn off training on your data before you upload anything you care about. It takes 15 seconds, and most folks don’t do it.

The access problem comes first.

Wrapping up

If your reaction to a room of kids shipping software in two hours is that you're behind, don't take that away from this. They weren't better at software than you, they had no tools you don't have, and half of them couldn't even log in. What they had was somebody telling them to slow down at the start, ask for exactly what they wanted, read what came back, and check it. That's it, and it's most of the job at every level I've ever worked at. The tools got easy this year. The thinking didn't, and the thinking is still the part you get paid for.

Putting this into action

  • Run one thing through the loop this week before you build anything. Ask it to interview you, make it write the plan, then tell it to find five problems with its own plan. If the plan reads like mumbo jumbo, ask it to explain the plan to you like you're 12 (no shame!).

  • Go turn off training on your data in whatever free AI tool you use most. Settings, two minutes, and then stop thinking about it.

Some AI tips for this week

  • Prototype ugly on purpose. Ask for the roughest working version first and say so explicitly, because otherwise it will spend your budget on polish you're about to throw away. We shipped low-poly graphics all morning and nobody in the room cared.

  • Give it the constraint out loud. "I'm on a free tier, keep this small" changes what comes back. Constraints work on AI for the same reason they work on people, which is that they stop it from solving a bigger problem than you asked about.

I wrote up everything we covered with them, free, for any parent who wants to run this with their own kids on a Saturday: decks.stealthx.co/kids-build-with-ai. Forward it to somebody with a bored teenager.

Onward & upward 🤘

Drew

P.s. the episode above is with Tory Stone Thompson, who runs the school. She's the reason I have opinions about any of this, and her line about the difference between a guardrail and a barrier has been rattling around in my head since recorded.