Two AI models, one prompt: a chai-stall game
Fable 5.2 and Opus 5.5 each wrote a complete browser game from the same one-shot prompt. A test bot played both. What each one built, and what it missed.
Can an AI model write a whole game that people want to play, from one message? I gave two models the same prompt and got two different chai-stall games. Both are live: Fable 5.2’s and Opus 5.5’s. Play one on your phone before you read on.
The experiment
The prompt is 316 words. It asks for “a complete, polished browser game called Kernel & Chai: Tapri Rush” in one self-contained HTML file, with no external assets or libraries. It lists the game in detail: customers with speech-bubble orders, a kettle you hold to boil (too long and it spills), kulhad or glass, snacks, a patience bar, rush hour, rain, a picky regular called Sharma ji, a sunset summary and an upgrade shop. It also asks for mouse and touch, a pause button, a saved high score and share tags.
Each model got it once, as one message, at High effort. No follow-ups. From my run log, Fable 5.2 took 18 minutes and Opus 5.5 about 30. Each returned one file: 124,105 bytes from Fable 5.2 and 100,833 from Opus 5.5. Every picture is drawn with canvas and CSS, and every sound is made in code with Web Audio.
To host them, I changed only the share tags, added a credit line and deleted one leftover test line from each. The game code is as the models wrote it.
How I tested them
I didn’t judge from screenshots. A test bot played both games in Chromium, with real mouse and touch events, as a player would. It reads the game’s state to decide what to tap, and changes nothing. It played 8 in-game days of each game from an empty save, then days 1 and 2 again on a 390 × 844 phone screen with touch.
Across all the runs, about 40 minutes of play, there were 0 console errors and 0 network requests in either game. No crash, no stuck customer, no negative money. The review was written before I told the reviewer which model made which game. One limit: the reviewer was an AI agent and the player a bot. No human has playtested them yet, and I didn’t measure frame rate on a real phone.
What each one built
Both games do every item in the prompt. They differ in what they added on top.
Fable 5.2 made the deeper game. Sugar is 0 to 3 spoons, not on or off. From day 3, orders mix two spices, and from day 4 customers order two things at once. Rent rises every day, from ₹50 on day 1 to ₹1,280 on day 8, so each day has a target. It has seven upgrades, a street with a cow, a kite and a sleeping dog, and Continue picks up a saved run at the next day. Its share tags were complete.
Opus 5.5 made the cleaner phone game. Its layout is designed for a 400 px wide screen, so text and buttons stay large on a phone. The hearts, the day clock and the order bubbles read at a glance, pours queue while one is running, and the how-to sits inside the menu. Sugar is on or off, and a wrong order costs one of five hearts. It is the easier game to pick up.
The review scored fun, look, polish and completeness out of 10: 8, 9, 7, 9 for Fable 5.2 and 7, 8, 8, 8 for Opus 5.5. That is 33 and 31 of 40, one reviewer’s view of one run.
The rough edges
Fable 5.2’s game is small on a phone. At 390 px wide, order text is about 8 px and the pause button is 24 × 24 px, under the 44 px usually advised for a tap target. Its first help card is taller than a phone screen, so the OK button is cut off unless you scroll. And the rent outruns the money: on day 8 the bot, playing near perfectly, earned ₹887 against ₹1,280 rent, so day 9 can’t be paid.
Opus 5.5’s game has the opposite problem. It has no rent or other way to spend money, so by day 8 the bot had ₹5,528 and nothing left to buy. Its share tags were left half done: the image and URL sat in a comment with a placeholder domain, so a link to it showed no picture. It also left a debug hook that exposes the game’s internals (I removed it from the hosted copy).
Both lock pinch-zoom, and neither turns off screen shake for people who ask their device for reduced motion.
What this means if you build with AI
The prompt was specific, and both models met it line by line. What they missed were the things it didn’t say: how big text should be on a phone, whether the economy balances over eight days, reduced motion. A one-shot result is a strong first draft. It still needs a check on a real phone, a few days of real play and someone who asks what the prompt forgot.
This is one prompt, one run per model and one reviewer. A second run could come out differently. So I’m not naming a winner. Play both and see which one you keep coming back to.
Sources
- Tapri Rush, Opus 5.5's build (playable) (opens in a new tab) kernelchai.com

