How to use GPT-6 Astra: A brief, reference images, and a feedback loop

Thomas Ricouard used GPT-6 Astra to build a space game with more than 10,000 planets. You can approach a planet, land, and leave your ship.

He also developed an interactive house where you can open drawers, switch lights, and operate an espresso machine. Even the coffee fills the cup.

Projects like these make you wonder what to put in the prompt. The detailed walkthroughs reveal more than a clever request: visual direction, production tools, and repeated inspection shaped the results.

Here's a practical way to apply those ingredients: make a project brief in ChatGPT, generate reference images, then set up Astra in Codex to build and check its own work.

This is a recommended workflow drawn from the examples, rather than the exact sequence every creator followed. Let's make it concrete with a small browser project: a snowy mountain observatory you can explore.

1. Create the project brief in ChatGPT

Start a conversation in the ChatGPT app. Explain your idea, then ask it to help resolve the important decisions before writing the brief.

Help me write a project brief for a snowy mountain observatory that runs in a browser. A visitor should be able to walk inside, use the telescope, and look at the stars. I want it to feel quiet and mysterious. Ask about the choices that materially affect the project, then produce a concise brief with scope and completion checks.

The brief should say what you're making, who it's for, what someone can do, and what is outside scope. Keep it small enough to finish. Our observatory might include one building and one telescope, with no accounts or multiplayer.

Put “Finished when” in the same document. For this example:

• A visitor can enter and reach the telescope without getting stuck.
• They can use the telescope and return to the room.
• The route remains visible in the intended lighting.
• The experience works in the agreed browser and on the intended device.

These checks will become Astra's stopping criteria. Save the agreed brief as a document in the Codex project so it can refer to the same requirements throughout the build.

The brief defines the experience. Now give it a look.

2. Generate reference images in ChatGPT

Use the brief to ask ChatGPT for a few visual directions. Compare them, choose one, and explain what needs changing.

Generate three concept images for this observatory. Keep the building and viewpoint similar so I can compare the visual directions.

Once you choose a direction, refine it. Perhaps the building works, but the interior is too bright or the telescope looks too futuristic. Make those decisions while they're still images.

Then generate the important views in that direction: the approach, the interior, and the view through the telescope. Save the selected images alongside the brief and explain what each establishes—lighting, materials, layout, or atmosphere. If you already have suitable references, use those instead.

In the Void Explorer walkthrough, Ricouard refined concept images that were first too realistic, then too simple, before saving visual targets for the game. The images gave later revisions something specific to aim at.

For your project, they serve the same purpose. Astra can compare an actual view with an agreed reference, instead of guessing what you meant by “mysterious.” Images guide appearance; the brief still defines how things should work.

3. Set up the feedback loop in Codex

Open the project in Codex with Astra selected. Give it access to the brief, the reference images, and the tools needed to create and inspect the result.

For a browser project, the setup needs three things:

• Project access: Astra can read and edit the files.
• A running preview: The project can open locally in a browser.
• Inspection tools: Astra can operate that browser, capture screenshots, and inspect relevant errors.

These tools form the feedback loop's supporting setup, sometimes called a harness. A prompt alone doesn't connect them. Ask Astra to check what's available, start the project's local preview, and demonstrate that it can open the page and operate an interaction. If access is missing, resolve that before relying on automatic checks.

Once that works, Astra can run the loop:

Build → run → inspect → compare → correct.

For the observatory, it should try entering the building, using the telescope, and returning to the room. It should also capture the important views and compare them with the saved images.

A screenshot might reveal lighting that misses the target. Trying to reach the telescope might reveal furniture blocking the route. Each observation gives the next correction a specific purpose.

Ask Astra to identify and fix those problems before returning to you. In Ricouard's spacecraft example, Astra noticed a plain bow and an overpowering background planet, then revised both. That is the kind of inspection you want it to take responsibility for.

Hand over the brief and start the loop

With the brief, images, and working inspection tools in place, use a prompt like this:

Build the project from the brief and saved reference images. Run it in the local preview, inspect the important views, and test the “Finished when” checks. Identify the most important mismatch, fix it, and repeat the relevant checks. Preserve what works. Make routine decisions within scope; ask before changing the direction or cost. Stop when the agreed checks pass and the result is ready for me to try. If progress stalls or a check is unavailable, explain what remains unresolved and what you tried.

Then try the result yourself. Astra can check interactions and compare images, but you still decide whether the experience feels right. Give specific feedback and let it run another focused cycle if needed.

The recipe is simple: the brief defines done, the images define the visual target, and the feedback loop gives Astra a way to close the gap.