My youngest son is nine. He’s been imagining a horror game since he was six. Characters drawn on paper, backstories explained over dinner, mechanics that shift a little every time he retells them. This summer we’re going to build it.

The best models aren’t always available, as I found when one got switched off, and they’re expensive when they are. I’d already tested what that means for a project like this. I built a small game with a frontier model, then added a level to it using a cheaper one. The level worked, but the quality drop was visible. Not broken, just worse, in a way I couldn’t point at but kept noticing.

So instead of using the best model to build the game, I used it to build the thing that makes the game: the harness that cheaper models will work inside for the rest of the project.

What the harness looks like

It’s a Godot project where certain parts are done and frozen. Player controller, combat, interaction, the verification pipeline, all checksummed so nothing can quietly edit them. Systems talk to each other through an event bus and nothing else. Every type of content, traps, puzzles, mobs, NPCs, has one working golden example, and the instruction is to copy it, not invent. And nothing counts as done until a verify script prints PASSED, which runs tests designed to catch exactly the mistakes a weaker model makes: traps that don’t hurt, puzzles that never solve, mobs that don’t chase. It even takes screenshots, because floating props and pitch-dark rooms pass every unit test and still ruin the game.

Honestly, I can’t tell you where most of these rules came from. I gave the frontier model the goal, “guardrails clear enough that a smaller model can deploy into this safely”, and it has been testing its way to the answers on its own. There’s a rule in the playbook banning a specific way of wiring up game events because it slowly leaks memory, and I have no idea what failure put it there. It found it, wrote it down, and moved on.

Tight, loose, tight

This approach kind of matches a concept from management literature: tight-loose-tight. Tight on goals, so everyone knows what success is. Loose on execution, so the team owns the how. Tight on results, so outcomes actually get checked.

The harness is the same shape. Tight at the entry: frozen core, hard contracts, one golden example per system. Loose in the middle: any model, or any dad who has never made a game, can build whatever trap they want in the way they want. Tight at the exit: verify has to pass, and you look at the screenshots.

This is similar to how you might onboard a junior developer. You don’t supervise their keystrokes, you scope the work tightly, give them a working example to copy, and review the result. The frameworks we built for delegating to people are turning out to work for delegating to models. The difference is that with models, the seniority of your “team” changes with your subscription, so the scaffolding has to carry more of the weight.

Frontier intelligence is CapEx

One way to think about it is that frontier intelligence is a capital cost, and smaller models are the operating cost. You spend the expensive model once, on the structure, and then the cheap model runs inside that structure indefinitely.

Access to the best model is always temporary in some sense, prices change, models get deprecated, rate limits appear, and the gap between the frontier and the mid-tier is where most of the cost lives. If the smart model’s output is content, you have to keep paying for it, but a harness you pay for once.

Let’s see if it works

None of this is stress-tested yet. No smaller model has built a level inside the harness, no nine-year-old has directed one, and the whole thing rests on a bet that the gates catch the dumb mistakes before a human has to. I’ll find out in a few weeks whether verify.sh is as good a manager as I hope it is.

Cheap execution also means fast execution. Nine-year-olds don’t have the longest attention spans, so when they say “now make the monster chase us,” you have a minute or two before things start to get boring. A frontier model thinking hard about architecture doesn’t fit in that window. A small model dropping a mob into a level that’s built to catch its mistakes might.