POV: Your kid wanted a new game, you created one in GPT-6

Ask an AI to "make me a game" and you usually get one of two disappointments: a wall of code you can't run, or a pretty landing page about a game that doesn't exist. The problem isn't the model. It's that you handed over a one-line brief and it filled in ninety decisions on your behalf.

This prompt fixes that by inverting the order. Instead of you specifying a game, the AI interviews you, one question at a time, adapting each question to your last answer. At the end it writes a short brief, you approve it once, and only then does it build. It also gets explicit instructions about the things AI game builds usually get wrong: fake transparency on cut-out characters, feet sliding on a static sprite, a level-select screen described as an "open world", and robot narration announced as a real voice.

Copy it, paste it into a capable assistant, and answer honestly.

How to use it

• Paste the whole prompt as your first message. Don't add your game idea yet, the first question is the idea.
• Answer one question at a time. Short answers are fine.
• "Choose for me" is always a valid answer. Use it whenever you don't have a preference.
• Upload assets when it asks: a photo for the main character, a location reference, music you're allowed to use.
• Approve the brief once. After that, let it build without re-litigating every detail.

Part 1 · The interview

The prompt tells the assistant not to dump the questionnaire on you. It asks one question, waits, then adapts. It skips anything you've already answered and anything that doesn't apply. Here are the ten topics it walks through, in order, with the questions it actually asks.

AGame idea

Then it asks whether you have a reference game, screenshot, video or link, and specifically which parts you like about it: the controls, the camera angle, the world, the challenges. References are used for inspiration, not for cloning, your game keeps its own identity.

BIntended player

If it's for children, it digs one level further: reading ability, familiarity with games, and whether they'll play alone or with an adult. Those three answers quietly change everything downstream, instructions, controls, difficulty, audio, and how long a session should run.

CPurpose

It also settles the language for on-screen text and spoken guidance here, before any of it gets written.

DMain character and avatar

This is the section most people get burned on, so the prompt is strict about it. If you supply a real person's photo or a specific avatar, it asks which features must stay recognisable: hairstyle, skin tone, face, outfit, accessories. Then it asks what kind of character you want, and explains the difference in plain terms:

• Your existing avatar, used directly
• An animated 2D character
• A character in a 3D illustration style, which is still a flat image that happens to look 3D
• A genuine 3D model, which has real geometry and can turn and animate from any angle

EWorld and location

You can upload location photos, maps, or backgrounds. Then it pins down the shape of the world: a guided path, separate levels, freely explorable areas, or a combination.

FVisual style

Whatever you pick applies to everything: character, scenery, objects and interface all have to look like they came from the same world.

GDevices and controls

Controls only get their own question when the choice actually matters. For mobile it defaults to simple swipes, taps and large touch targets, and it must not require keyboard input or hover states, which silently don't exist on a phone.

HChallenges and progression

From your answer it establishes six concrete things: how easy level one should be, how difficulty climbs, what happens after a mistake, what counts as success, what the reward is, and roughly how long a session should last. For young children it biases toward encouragement, clear goals, short rounds and gentle retries, and introduces one new mechanic at a time.

IMusic and voice

You can upload music, narration or sound effects you're authorised to use. Music should suit the audience, loop smoothly, and duck in volume while someone is speaking, with simple controls the player can reach.

JOther assets and delivery

Part 2 · The brief you approve once

After the interview it writes a short brief, not an essay, covering the working title, intended player and age, the main activity and learning goals, the character and its identity requirements, visual style and world, devices and controls, challenges and difficulty curve, music and voice, delivery format, and any capability limits it already knows about.

You confirm it once. The prompt explicitly forbids restarting the interview or asking for approval again on routine implementation details, which is what turns a fun collaboration into a tiring one.

Part 3 · Build small, then expand

The first build isn't the whole game. It's the smallest version where the core loop genuinely works:

• You can control the character
• The character can interact with the world
• At least one challenge works end to end
• Success and retry behave correctly
• The controls work on the device you named

Part 4 · Make the movement believable

This is the section that separates a game that feels good from one that feels broken, and it's worth reading even if you never use the prompt. Animation has to be tied to actual movement:

• The walk or run cycle is driven by distance travelled and stride length, not by a timer
• Scenery and obstacles move at the same world speed as the character
• No feet sliding while the character runs in place
• Locomotion stops when movement stops
• Sensible acceleration, deceleration and turning
• Consistent jump, landing and collision timing
• Motion is independent of screen refresh rate, so it plays the same on a 60Hz laptop and a 120Hz phone

Part 5 · A loop worth repeating

Engagement here isn't a feature list, it's a five-beat cycle the prompt asks it to build around: understand a small goal → take an action → get immediate feedback → earn a reward or learn from a retry → move to the next small challenge.

For educational games there's one more constraint that gets ignored constantly: learning has to be part of what the player does, not a quiz bolted onto a game. Questions and answers stay factually accurate, examples suit the age group, and nothing requires reading ability the intended player doesn't have. Difficulty climbs through meaningful changes, an extra choice, a new obstacle, a slightly longer sequence, not just by making everything faster.

Part 6 · Sound that doesn't fight you

• Audio starts correctly after a user interaction, which is a hard requirement on mobile browsers
• Music sits at a sensible volume and spoken clues stay audible over it
• Sounds don't pile up or constantly interrupt each other
• Pause, mute and resume actually work
• Narration is short, clear and age-appropriate

If an audio service isn't available, it finishes everything that doesn't depend on it and names the remaining gap plainly instead of quietly shipping silence.

Part 7 · Test the real thing

The prompt contains an explicit authorisation for browser, visual and interaction testing, which matters, many assistants won't drive a browser without it. It's asked to check both a phone-sized and a desktop-sized viewport and verify that:

• The character still looks like the avatar you approved
• The body and clothing aren't accidentally see-through from bad background removal
• The walk and run animation visibly changes and matches movement
• Touch controls and gestures work
• Challenges can be completed and mistakes trigger the intended retry
• Difficulty increases the way you agreed
• Music, voice, mute and pause work
• Text and controls fit the screen, assets load, and it's still playable after several rounds

Part 8 · Changes that don't break what worked

When you report a problem later, the prompt asks it to inspect the actual implementation, explain the cause briefly, fix the underlying issue rather than the symptom, preserve the approved character and the features that already worked, test the change, and update the same game instead of spawning a new one.

And the one that saves the most heartbreak: don't solve an animation problem by replacing your character with a different-looking person.

The full prompt

Copy everything below and send it as your first message.

Why this works

Nothing in this prompt is a clever trick. It works because it does three unglamorous things: it forces the specification to come out of your head one question at a time instead of being invented, it gets one clear approval before any building starts, and it names in advance the specific shortcuts that make AI-built games feel fake, so the model can't take them without breaking an explicit instruction.

That pattern generalises well beyond games. Any time you're asking an AI to build something you can't fully specify up front, an interview plus a brief plus a small working version beats a long request every time.

Found this useful? Browse more articles or follow @m.redefinedbyai for more prompts and setups as they land.