Making a game
What actually happens between “I want a racing game” and a playable one, and the handful of decisions on your side that decide whether it is good.
The shape of a build
Whether you are using the built-in agent or your own AI over MCP, a build follows the same six stages. Knowing them tells you where you can intervene.
-
Brief
A genre-aware questionnaire, or your one-line description if you skip it. This is free.
-
Research
The AI reads the platform's knowledge packs — the marker convention, mechanics, balance rules, screen requirements — plus a gameplay blueprint if your genre has one. Also free, and the single biggest factor in whether the mechanics are any good.
-
Character design
Concept art for the main cast, shown to you before any code exists. You pick one. Costs a few images.
-
Code
A small
game.htmlshell first, then one file per character, scene and system. Free. -
Assets
Sprites, backgrounds, UI, sound effects, music, voice, 3D. This is where nearly all the credits go.
-
Test
A real headless browser runs the game and reports actual JavaScript errors. Anything broken is fixed before it reaches you.
Stages 3 and 5 are the only ones that cost money. Stages 1, 2, 4 and 6 are free on the platform — so asking the AI to think harder, restructure the code or re-test costs you nothing.
Writing a brief that works
The questionnaire covers most of this for you. When you are writing free-form — especially over MCP — these are the details that change the outcome:
| Say this | Because otherwise |
|---|---|
| The core loop in one clause — “dodge cars and cross the road”, “match three to clear the board” | You get a genre label rather than a game, and the mechanics are invented rather than borrowed from something that works. |
| The art style — pixel art, flat cartoon, anime, hand-painted, neon | Every generated image is a fresh guess and the game ends up stylistically incoherent. |
| How you lose — lives, a timer, a tower that falls | Games with no failure state are not tense, and the difficulty sliders end up controlling nothing that matters. |
| How difficulty grows — faster, denser, fewer moves | The whole game plays at wave one's settings forever. |
| Portrait or landscape, and whether it is for phones | The layout is designed for a desktop window and feels wrong on a phone. |
| Your budget, if you have one | The AI cannot tell whether you want a jam prototype or a full production. |
Do not specify numbers. Speed, gravity, spawn rate and score values are exactly what the Inspector is for — you will feel the right value in ten seconds of dragging, and no amount of describing it in a prompt gets there faster.
Starting from a template
For well-known genres, the platform ships gameplay blueprints reverse-engineered from real shipped games: the actual mechanics and the actual tuning constants, not a guess at how the genre works. Pick one from the Use a template button on the Make page, or just name the genre and let the AI pull the matching blueprint itself.
Blueprints exist for genres including:
- bubble shooter (Puzzle Bobble style), match-3, hexagonal rotation puzzles;
- Pac-Man style maze chases, Bomberman style grid deathmatch;
- Crossy Road style lane hopping, endless runners, antigravity racers;
- parking-jam and block-sliding puzzles;
- roguelike battlers, auto-battlers, card battlers, tower defence;
- dating-sim and horror visual novels, point-and-click adventures;
- beat-em-ups, run-and-guns, adversarial Tetris, cooking co-op, fishing sims.
A blueprint fixes the mechanics and the numbers. Theme, art style, characters and story stay entirely yours.
Budgets
Turn on Budget mode (in Agent Lab) and the AI prices the scope, tells you what your budget buys and what it does not, and asks before it spends anything. These are the rough shapes, in credits — at 1,000 credits to US$1.
| Budget | What that is |
|---|---|
| ~3,000 | One screen, one mechanic. Around 25 images, a few sound effects, one music track. |
| ~10,000 | Several enemy types and scenes. Around 80 images, a full sound set, three or four tracks, a cover. |
| ~40,000 | Multi-stage. 300+ images, or fewer images plus animation and 3D, music per game state, story CGs. |
| ~100,000 | Many levels with their own art, animated bosses, voiced lines, a full soundtrack. |
A budget is treated as about this much, not a tripwire. Past roughly 85% the AI stops starting new things but finishes what it began — overshooting slightly beats shipping half a sound set. If it goes meaningfully over, it is expected to say so and why.
Where the money goes if you leave it open
Given spare budget, the platform directs the AI down a fixed ladder rather than letting it stop early. In order: real animation for the player and key enemies, distinct hit and death effects, enemy variants so repeated objects stop looking copy-pasted, a background per level or state, voice and per-state music, story CGs, a proper UI kit, and finally re-generating its own three weakest images.
What every finished game needs
Two requirements are enforced by the platform's guidance rather than left to taste, because they are the most common ways an AI-made game reads as unfinished:
Screens
A title screen with the game's name, one obvious Start button, a controls hint and the high score; a working pause; and a dedicated result screen with the score, the high score, a “new record” callout and a one-tap Play again. A game that boots straight into gameplay and ends on bare “Game Over” text is the single most common defect — if you get one, ask for the screens and you will get them.
A difficulty curve
Difficulty lives in a waves table in the game's config: one row per stage, each stricter
than the last on at least one number, with a final row a good player can still fail. Because it is
in the config, it is also in your Inspector — you can retune the whole curve without touching
code.
Iterating afterwards
The decision that saves the most time and money is knowing who should make a change.
| Change | Who | How |
|---|---|---|
| Speed, gravity, spawn rate, score values, difficulty curve | You | Inspector slider, then Save. Free, instant. |
| Colours, palette | You | Colour pickers in the Inspector. |
| Level layout | You | The grid editor, if the AI exposed the level as a grid. |
| Moving, resizing or deleting one object | You | The Hierarchy panel. |
| Replacing one image or sound | You | The assets dock — upload your own, or generate a replacement. |
| New mechanic, new enemy type, new scene | AI | Ask. It writes a new file rather than rewriting the game. |
| “It is not fun” | Both | Tune first — it is usually a number. If tuning cannot reach it, the loop itself needs changing, and that is an AI job. |
Point at the problem, do not diagnose it. In the Editor you can screenshot the running game, draw on the screenshot and send it to the agent. “This gap is impossible to jump” with a circle around the gap works far better than a paragraph describing it.