A whole 3D game, with one person directing and an AI writing the code.

The problem

Everyone has seen an AI write a small demo. The open question for a team is whether a person can direct an AI through something large, with many parts that have to keep working together, and come out with software that is finished, tested and live rather than a pile of half-working pieces.

A game is a hard way to answer it. It has to look right, feel right and run on a phone, and every change can quietly break something three levels away.

Why it is hard

• Size. About 23,400 lines of the game's own code in 114 files, 74 3D models and 17 pages a player can open. No one keeps all of that in their head, human or AI.
• Coupling. A save file, a points book, an outfit and a hat all have to survive moving between pages, a failed level, a return visit and a new game.
• Feel. Tests can prove a sum. They cannot prove a camera feels right, so the human has to play it and say what is wrong, and the fixes have to land without breaking the rest.
• Respect. It is set on the country of real peoples, and shows real places. That needs care, review and plain statements of what is and is not real.

What I did

I wrote the story, designed the levels, picked every place and character, and set the rules the AI worked under. The AI wrote the code in VS Code and ran the tests. The rules that mattered:

• Every rule of play lives in a plain module with its own tests, apart from the drawing, so it can be checked without a browser.
• Nothing is reported as working until it has been run. Claims are labelled as tested or not.
• When I find a bug by playing, it is fixed and a test is written that would have caught it.
• The AI says plainly what it did not check, and never fills a gap with a guess.

The proof

• 962 unit checks in 27 suites cover the rules: the trip and its keys, scoring, return visits getting harder, the currency, hats and outfits, the leaderboard's limits, and the feedback form's defences.
• 87 browser checks drive the real game in Chrome on desktop and phone sizes: every page loads clean, a whole level is completed and then failed, the shop sells and saves a hat, the rhythm game floats the boat on the beat and not off it, and the live site passes the same suite.
• Bugs the tests caught before players did: a level that crashed only on phones, scores that could come out as fractions a leaderboard would refuse, a swim where holding forward pinned the player against an invisible wall, and pages that lost their place on the hosted site because it serves addresses without ".html".
• Bugs I caught by playing, each now covered by a test: a map that, closed with its X, left the camera circling with no player in sight.

Try it

It runs in any modern browser, on a phone or a computer, with nothing to install: playbutterflyfx.com/chomp. It is free, has no ads, and saves your game on your own device.

What it does not claim

• It is a portfolio piece, not a commercial product.
• The AI did not design it. The ideas, the story and the decisions are mine; the AI implemented and tested them under my direction.
• The cultural content is going to Aboriginal and Torres Strait Islander people for review, and will change where they say it should.
• How it runs on every phone is not yet measured across devices.