Language and display
BE THE BOSS OF AI.ADRIAN DUNKLEY ↗
← This newspaper edition

Gemini drafted a playable game. Try it on a stockroom lesson

Gemini 4 Hydrogen produced a browser game in a creator’s test. A short training exercise lets a supervisor compare it with the tool they use now.

A yellow cylinder sits under a black gate at one end of a layered teal zigzag.
One Turn Before the Gate · AI-generated artwork · AI Boss Ltd

A game that opens in a browser can let a new employee practise a decision before a busy shift. In a test published on 12 November, creator Adam Holter used Google's newly released Gemini 4 Hydrogen to generate a game called Emberhold. The player moves around a settlement and places buildings. In a separate check on 17 November, the saved demo opened in Chromium, the character moved and a cottage appeared after a placement. That was a check of the opening controls.

The creator supplied a detailed brief

Holter used Google Antigravity version 1.5.7, run through typed commands, with Hydrogen at extra-high effort. His brief specified the game's mechanics and asked for a single HTML file, which a browser can open directly. His capture shows two buildings and the second wave of the game in progress. It did not capture the player choosing an upgrade. The report does not establish how much the run cost or how long another person would need to repair and finish the game.

Google launched Hydrogen on 12 November. It says it offers the model on free and paid tiers in AI Studio, its developer workspace. Jamaica is on the AI Studio country list. Google lists Antigravity access for paid AI Pro and Ultra subscribers. To run a named-model trial, select Hydrogen in AI Studio if your account offers it. Record the High setting or whichever option you use. The free Gemini app can choose a model for you, which makes the comparison harder to reproduce. In testing by Artificial Analysis, Hydrogen scored 51 at high effort and 54 at extra-high on its Intelligence Index version 5.0.1. The index combines nine tests across coding, knowledge work, science, and general knowledge and reasoning. The figures are points on that combined scale. They do not estimate staff training time.

Give your two tools the same stock rule

Use this practice brief in Hydrogen and your usual tool: “Make one HTML file for a phone. Show two independent orders, each starting with six boxes in stock: one order for four and one for eight. Give the trainee two choices on each card: release the order or ask a supervisor. Releasing four leaves two boxes. Reject an attempt to release eight and explain the need for a supervisor. Asking a supervisor must leave stock at six. Add a reset button. Use no external files, customer data or network requests.”

Ask each tool for the complete file. Copy only the HTML code into a plain-text editor and save each version as stockroom.html in a separate folder. Use a phone browser in a test setup with no work accounts and the network off. If your phone cannot open a local web page, ask your IT colleague to put both files in an approved test preview. Do each of the four choice-and-order combinations once in each version. Reset to six boxes between attempts. Compare the displayed quantities and explanations with the brief. Record the time through correction and phone review. Keep your usual tool if it produces a version that passes with less work. The trial checks a small prototype and supplies no evidence of what a trainee learns on shift.

Ask colleagues to explain a wrong answer

Use the AI Literacy Quiz with a colleague who will review the lesson. Select Start the Quiz and answer the questions to see the explanation after each choice. The opening question distinguishes a tool that produces new content from software that simply runs online. A practice run using the first option for all 30 questions returned 7 out of 100 with a breakdown by topic. Pick a missed explanation and ask your colleague to put it in their own words before you write feedback for the stockroom exercise. The quiz score describes answers to this set of questions. It does not measure how a colleague will do on shift.

Sources and further reading

Sources checked 17 November 2026.