In AI-Built Tools, the Constraint Is the Product
An award field for games, tools and apps asks you to open the work instead of looking at it. Thirty-five gallery works, 23 visitor comments with makers answering many of the usage questions themselves, and a pattern worth stealing: the memorable ones all refuse something on purpose. A voice tool that needs the word twice, a letter that waits for a satellite, one radio station per town code.
AI-assisted draft. Reviewed and edited by the Phosphene team before publication.
On this page

A video can be judged from a thumbnail and thirty seconds. A tool cannot. You have to open it, use it for something you actually needed, and find out where it breaks. That is why the games, tools and apps field of the International AI Creator Award 2026 is the most informative corner of the whole competition: nothing in it can coast on a striking frame.
The numbers as of 11 November: 36 submissions, fourth largest of the ten fields, 35 works in the public gallery, 23 comments from visitors. The 34 entries not submitted by the organisers came from 27 people. Five of them entered more than once in this field, and six also entered work in character design, manga, music or video, which suggests the tool builders are the same people making images and films with these tools.
I read the field after writing about the open category and the entry distribution, and the pattern that held my attention is not about model quality. Every entry that stuck with me refuses something on purpose.
A voice tool that will not send by accident
Kotocast is dictation software. You speak, the words appear in whatever app you are already using. Speech recognition runs entirely on the machine, so audio never leaves it, and the stated reason is accessibility: people who cannot use their hands or hold a posture for long still need to write.
The design decision worth copying is a small one. Saying "send it" once, mid-sentence, does not send anything. You have to say it twice in a row before the message goes.
That is a refusal built to prevent the single most common failure of voice input. Every dictation tool eventually sends a half-finished message, and the fix is usually a confirmation dialog, which adds a step to the ninety-nine times you meant it. Kotocast moves the cost onto the rare case instead of the common one. Nobody has to learn a setting. Nothing is asked of the user at send time except the habit the interface already taught.
I keep coming back to this because it is the kind of decision a language model would not propose. It is not a feature, it is a rule about what the product will not do, and it comes from someone watching where the tool fails in real use.
Waiting is the mechanic
LOVE FROM SPACE lets you send a letter to someone you care about, and only while a satellite is passing overhead. You hold the phone toward the sky, the app locates Ibuki GW from real orbital data, and the send button unlocks when it is nearly straight up. The transmission is theatre, since no radio is emitted, but the orbit and the physics are real.
Strip the poetry and you have a timing constraint: the product is unusable most of the day, and that is the point. The maker's framing is about recovering the act of waiting and looking up in an era when messages arrive instantly.
As design, it is a strong argument. A messaging app competes on speed and loses nothing when it is fast. This one competes on the moment before sending, and the constraint manufactures that moment. If it removed the satellite dependency it would be an ordinary greeting card with an animation.
One station per town, no matter how small
Kikeru Nihon Chizu, roughly an audible map of Japan, is a web radio for songs about places. Creators across the country make AI-generated local songs, and each town gets a station on a map. The frequency is derived from the national local government code, so Memuro in Hokkaido comes out as FM 01.637 MHz.
The rule does the emotional work. Every municipality gets exactly one station, which means a village of two thousand people sits on the same map as Osaka, and the code that produces the frequency is the same bureaucratic identifier the town already has. Listen long enough and prefecture stamps accumulate.
This is the cheapest form of equity I have seen in an AI project, and it costs nothing to implement. A ranking would have buried the small towns under the cities, because that is what rankings do. Picking an identifier nobody can influence removes the competition entirely, and the songs stop needing to be good in a comparative sense.
Rules small enough to check
Kyojin, a pinching game, is played on a six by six grid. Six pieces each, moves of one or two squares orthogonally, captures by pinching, and a forced-capture rule that removes the option of declining a capture. That last rule is where the game lives, because it turns a free choice into a tactical liability. The maker used AI to verify the ruleset and simulate games while tuning it, and shipped five difficulty levels plus online play with no registration.
Compare that with the usual AI game project, which is impressive for a day and hollow by the weekend. The difference is that this one committed to a rule set small enough to reason about, then used the model to test it rather than to generate content.
Scacelith, from a maker in France, takes the opposite route on presentation and the same route on substance. It is a chess simulator where you play a porcelain robot under tournament rules, the entire screen is drawn procedurally rather than assembled from image assets, and the chess itself runs on Stockfish with speech synthesis bolted on. It was built with Claude over about fifteen days and released open source.
Procedural rendering is an unfashionable choice right now, when anyone can generate a beautiful background in one call. It also means the thing cannot look broken the way a generated asset looks broken, and it stays coherent when the camera moves. Fifteen days for a playable chess game with a distinctive look is a real datapoint about what a single person can now build, and none of it depends on the model inventing a style.
What a veteran keeps
Grassroots Football is eleven-a-side football in a browser, rebuilt with generative AI by someone who has been making football games since the 1990s. He used to work with teams of dozens of people. Now it is a small group and the model.
The part he says he focused on is off-ball movement: the run into space, the player who pulls a defender away, the third-man reception. Not the rendering, not the menu, not the amount of content. The behaviour that makes football read as football when you are watching a tiny figure you do not control.
His line in the entry is the best sentence in the field. Deciding what is interesting and what to throw away is human work.
Anyone who has shipped anything with a model knows this in their hands and struggles to say it. The model will produce every option you ask for, at volume, in the style you asked for. It will not tell you which of those options is the reason someone comes back. That judgment is the job now, and it is not a smaller job than the one it replaced.
The comment box as a spec
Twenty-three comments is a small sample, and I would not build a strategy on it. But the way they are used here is unusual enough to matter. Visitors ask how a feature works, and makers answer in their own words, so the gallery turns into documentation written by the person who made the decisions.
Two exchanges are worth quoting in substance.
A visitor asked LoomStroke, a paint app that mixes raster, vector and AI in one layer stack, whether an AI change can be applied to a single selected layer while the rest of the artwork stays untouched. The maker said yes: save a workflow as an AI Effect layer, drag in the reference layers, and the source layers stay unchanged, with output landing on the AI Effect layer or a new layer and the result kept in gallery history.
That question is the one every image tool gets asked, in some form. Can I change this without losing the rest. Most tools answer it with a mask and a hope. An answer that holds up is architectural, and it is decided before the model choice.
A visitor asked MonogaTree, which turns memories into a one-off picture book when no photographs exist, how it keeps generated illustrations faithful to what the user described. The maker's answer was that the system guides the model with the user's memories and emotions, prioritising emotional truth over photographic accuracy, because the model cannot reproduce the past and the person's words remain essential.
That is the honest version of the promise. Accuracy is not available, so the product aims at recognition instead, and says so.
What to take from a field you have to touch
Every project above is defined by a limit, and the limit is not a compromise forced by the tools. Say the word twice. Wait for the satellite. One station per town. Forced capture. No weapons, no blood, no ads, which is how one football RPG describes itself. Refuse to rewrite the artist's decisions.
None of those limits need a better model, and none of them get worse when a better model ships. That is the test I would apply to any tool idea now: if the interesting decision disappears when the next model lands, it was never the product.
Two cautions. The summaries are written by the organisers, so they describe intent, and the field mixes finished work with trailers and early alphas, including one platform that openly submitted a trailer for an unfinished build. And a gallery entry is not a result. The rules are worth copying; the claims attached to them are not yet proven.
Voting closes 13 November, Anywhere on Earth, one work per field, free. Most of the entries open in a browser with no install, which makes this the rare award where the judging takes longer than the reading.