GPT 6 game prototyping: build the first playable loop

GPT 6 game prototyping around one playable loop

This GPT 6 game prototyping example asks GPT-6 Astra at medium reasoning effort in Codex to build Catch the Stars: a small browser game with original shapes and no external dependencies. The September 7, 2026 experiment preserves the first source snapshot and the check record. Load the demo when you want to play; the example remains within this page rather than creating a separate indexable game route.

The input is a rule contract, not a request for an expansive game. Arrow keys and visible touch buttons move a basket. Catching a star adds one point, the player starts with three lives, the third miss ends the game, and ten points wins. The game must stop changing score and lives after either terminal state.

What the artifact must make visible

Score, lives and status are real text alongside the named canvas. That gives the player information outside the drawn scene and makes the values directly inspectable. The downloadable source exposes the game rules and control logic; the first-output snapshot gives the comparison a fixed starting point.

Restart is the most demanding part of this small contract. It must clear score, restore three lives, reset objects and timers, release held input and leave only one animation loop. A new-looking screen is insufficient if an older loop still changes the state in the background.

Read browser evidence alongside state tests

Deterministic checks can place a star at a known position and step the engine to examine a catch, miss or terminal transition. Those checks are useful because waiting for random events can hide a defect. They are test-only mechanisms and should remain separate from ordinary gameplay.

Real browser controls answer another question: whether a player can actually move, use the visible buttons and restart the running interface. Desktop and mobile viewport checks also expose controls that leave the screen or a canvas that overflows its container. Screenshots of start, scoring, loss and win help explain the states, but do not prove the transitions between them.

The first browser assertion expected a miss before the falling star had crossed the bottom boundary: it expected two lives but observed three. Extending the test wait by 200 milliseconds corrected the timing assumption. The game implementation remained byte-identical to its first snapshot; the failure and harness diff are retained.

The final record reports nine passing engine tests and 14 browser checks across 1440 × 900 and 390 × 844. Keyboard and mobile touch move the basket; release and touch cancellation clear input. Both terminal states freeze correctly, and repeated restarts clear state without duplicating the animation loop. Restart returns to Ready; Start begins the next round. A separate integration run on this article passed 16 checks, including loading the game on demand, both control methods, scoring, both terminal states and repeated restarts. The original prototype captures and the article integration records are preserved separately.

Official context and a useful next slice

OpenAI’s Playco account, rechecked September 7, reports three themed prototypes from one grey-box foundation and 50% fewer manual fixes than its previous model. One cyberpunk version still needed a performance fix. Those are Playco-reported outcomes, not a speed or quality benchmark for Catch the Stars.

After inspecting the loop, change one rule at a time and repeat the terminal-state and restart checks. A short play session can reveal unclear feedback, but it does not establish long-term enjoyment or target-device performance. The frontend QA example demonstrates a related failure-and-repair record. Use the practical guide to preserve your input and revisions, or compare all four examples before choosing a first task.