← All postsSawdustOffcuts from bigger efforts, shared anyway.
Neon 3D proves that AI can one-shot code a game with high trust
Rob MonteroExperiment · AI
I wanted a game I could play on a plane. Something small, pretty, that installs on a phone or a laptop and does not care whether there is Wi-Fi. The result is Neon Stack 3D: a falling-block puzzle in a glowing 3D well, with a chiptune soundtrack that does not use a single audio file.
It is also the newest entry in a habit I keep on purpose: every few projects, I change the harness, the model, or both, and see what happens. This time it was OpenCode running a free model, DeepSeek V4.1 Flash Free. I did not know what that combination could do. Now I do, and the answer surprised me.
Try it below. Click the game first so it gets the keyboard, then press Enter or Space to start.
How to play
- Move with the left and right arrows. One press is one step, on purpose: no runaway key repeat.
- Rotate clockwise with Up or X, counter-clockwise with Z or Ctrl.
- Drop: hold Down for a soft drop, press Space for a hard drop.
- Hold a piece for later with C or Shift. You get one hold per piece.
- Camera: drag to orbit, scroll to zoom, R resets the view.
- Sound: M toggles the music, N the sound effects. Escape pauses, A opens the About panel.
On a phone, an on-screen pad appears with the same moves.
Modern rules, not 1984 rules
The pieces come from a 7-bag, so every seven pieces you get one of each and there are no long droughts. Rotation follows the Super Rotation System, with wall kicks that let a piece slide out of a tight spot. A grounded piece waits a moment before it locks, and moving it resets that wait, but only 15 times, so you cannot stall forever. There is a ghost piece showing where it will land, T-spin detection with its own scoring, and bonuses for back-to-back hard clears and combos. Four lines at once is called a QUAD here.
Built to work on a plane
The version at neon.robs.ws is an installable web app. A service worker caches the whole game on the first visit, so after that it boots and plays with the network off. You can check this yourself in thirty seconds: open it, go to DevTools, Network, Offline, and reload.
- A first visit downloads about 201 KiB gzipped across 14 files.
- The full offline cache is 507 KiB gzipped across 110 files. Around 306 KiB of that is Three.js add-ons the game does not use yet. That is deliberate headroom, and the obvious first thing to trim.
- On the test machine it held the display’s 120 Hz refresh rate, with a mean frame time of 8.29 ms at 2208 × 1722 pixels. That is one machine and one run, not a promise about every device.
Every sound is synthesized at runtime with Web Audio: a pulse-wave lead, a triangle bass, noise for the hats and snare, and a sine kick, in an original loop in A minor.
Who built it
The game code was written by an AI coding agent: the model DeepSeek V4.1 Flash Free, running inside the OpenCode harness, in a single working session on 28 October 2026. No person wrote or edited the game code. I did the parts that are not code: I decided to make it, asked whether the name it started with was a trademark problem, chose the domain, and decided it was good enough to ship.
That trademark question is a nice example of the boring parts mattering. The game had been built around a famous falling-block game’s name. When I asked, the agent explained the issue and renamed everything: title, icons, manifest, README, start screen, scoring labels, and even internal names like the cache and the debug hooks. Then it bumped the cache version so copies already installed would pick up the new name.
Why I keep switching harnesses
A model on its own is raw power. The harness is what turns it into work: the tools it can reach, the loop it runs in, what it is allowed to check, and when it has to stop and ask a person. Change the harness and the same kind of model behaves like a different collaborator. That is why I keep trying new ones instead of settling on a favorite.
- Inclusive Mayan Sudoku was built in Kiro, Amazon’s agentic IDE, where every request becomes a spec with requirements, a design and a task list.
- Lotería de los Pueblos started as a Claude Design canvas and was built by a Claude Code orchestrator that never wrote the app itself: it handed 41 subagents their tasks and had a second set of agents attack each result.
- nombra.la, my domain finder, was built with Claude Code, and at runtime it mixes on-device AI with free models that rotate, so a daily job re-picks the best ones.
- Neon Stack 3D is OpenCode with a free model, and a real browser in the loop.
Each one taught me something the others could not. Kiro taught me to write the spec first. The Lotería orchestrator taught me that an agent whose only job is to find bugs is worth its weight. This one taught me how far a free model goes when the harness lets it check its own work. I would not have learned any of that by sticking to one tool.
The part I actually find interesting
An AI writing a falling-block game is not news. There are thousands of these games out there for a model to have learned from. What caught my attention is that the harness gave the model a real browser. It could load its own game, run JavaScript inside the page, read the console, and check the live site with curl. So “done” did not mean “the code looks right”. It meant twelve checks that it actually ran, including:
- every file the service worker promises to cache exists (110 of 110),
- thirty pieces played through moves, rotations, holds and hard drops with no errors,
- a real line clear moves the line counter from 0 to 1 and the score from 0 to 134,
- the game still boots after the server is switched off,
- the live site serves the right headers, and the new version was live about 45 seconds after the push.
And it found two real bugs after the game already looked finished. In the first, a movement helper called without its vertical argument quietly turned the piece’s position into NaN, and every later write to the board went to a cell that does not exist. The crash showed up in the locking code, but the agent traced it back to the missing argument instead of patching the symptom, and it checked that no real player input could trigger it. In the second, the locking code guarded one edge of the board but not the other. Both fixes were a couple of lines. A static code review would most likely have shipped both.
There was also a bug that was not a bug: a test reported a crash that turned out to be a mistake in the test itself. The agent said so and re-ran the test properly instead of “fixing” working code. Being cheap to check is what makes it cheap to admit you were wrong.
What this does and does not prove
This is one project, one session, one model. No benchmark was run, and I am not claiming this model is better than any other. The genre is very well known, the browser tools did a lot of the quality work, and a person steered. What it does show is the value of the loop: when an agent can run what it wrote and read the result, it catches the kind of mistake that reading the code would miss. Harness it well, and even a free model can ship something real.
Stay curious
If you only ever use one model in one tool, you only ever learn what that pairing is good at. Pick a small project, something you could finish in an evening, and build it with a harness or a model you have never tried. Give it real tools, make it prove that its work runs, and keep the decisions that matter for yourself. Then try another one. The tools are changing every month, and the curiosity to keep trying them is the skill that does not go out of date.
One small honest footnote: the brief the agent wrote for this post said the cache was on version 4. By the time I checked, it was on version 5, because of a later commit that made Ctrl work as a rotate key, as the README promised. The live site wins.
A note on this embed
The game above is the single-file share build, shown in a sandboxed iframe on this page. It loads Three.js from a CDN, so it needs a connection, and the sandbox means it cannot save anything: your music and sound settings reset when you reload. For the installable, offline version, go to neon.robs.ws. The code is on GitHub.
Neon Stack 3D is an independent project and is not affiliated with Tetris Holding.