Build a Small Game with GPT-6 Astra: References to Playable Scene

Choose a loop you can describe in one sentence

Our starter asks the player to move through a small field, collect three lights and reach the exit. The interface shows how many lights remain. That is enough to test input, state, collision and completion without needing an inventory system or a story campaign.

Write the finish condition before choosing an engine. If your real project needs 3D physics or engine-specific export, say so. If you are only testing a mechanic, a small browser prototype may answer the first question with fewer moving parts.

• OpenAI practitioner guide: building games with Astra

Use references to constrain the scene

Attach references you are allowed to use and explain whether they define mood, materials, proportions or layout. Then ask for an asset list: player, floor, obstacle, collectible and exit. Record each asset's scale and role so individually attractive objects still work together.

A first pass can use simple shapes. Replace them after the controls and win condition are reliable. If you request generated artwork, distinguish concept art from a game-ready asset that meets the engine's geometry, transparency or animation requirements.

Give Astra a playable definition of done

Ask for the runnable project, instructions and a brief report of actual checks. A screenshot is useful evidence of appearance, but it cannot establish that a collision or restart works. If the environment cannot launch the project, the handoff should name that unverified step.

Keep the task small enough that the agent can inspect and repair its own result. Expanding to more levels before the first one works multiplies the places a defect can hide.

Build one small playable scene: move a player, collect three lights, then reach an exit. Include visible instructions, progress and Restart. Start with simple shapes and a coherent palette. Deliver runnable files and setup instructions. Verify movement, one-time collection, the locked exit before completion, the win state and restart. Report what you actually tested. Keep menus, inventory, accounts and additional levels outside this first prototype.

Play the starter and inspect its source

The linked Light Trail starter is a self-contained HTML file. Open it in a browser, use arrow keys or the on-screen controls, collect the three lights and reach the exit. Restart resets the board. Its source is small enough to inspect before adapting it.

This is a teaching artifact with intentionally simple geometry and no external libraries, accounts or network calls. It demonstrates a reviewable game loop; it does not demonstrate a complete game-production pipeline or an Astra performance benchmark.

• Play Light Trail and view its source

Test the edges of the loop

Move into a wall, revisit a collected item and reach the exit too early. These short checks tell you more about state handling than repeatedly playing the happy path. After winning, restart and confirm the count, player position and items all reset together.

For a game with continuous movement, test the frame rate and input behavior on the devices you intend to support. Our grid starter avoids that complexity, so its tests should not be taken as evidence of performance for a 3D engine project.

Add visual assets after the loop survives testing

Choose a consistent style sheet for the next pass: palette, camera angle, materials and object scale. Generate a small set of concept references before commissioning every asset. Recheck readability after replacing simple shapes; a detailed collectible that blends into the floor is a gameplay regression.

Ciyo can help prepare visual references and image assets for the art direction stage. Engine integration, collision geometry and the playable build still require their own implementation and verification.

• Write reference-aware image prompts
• Astra and Blender: a separate 3D workflow