Make a Platformer Game With AI: Jump Feel and Levels
Make a Platformer Game With AI: Jump Feel and Levels
Describe the character, the jump, the platforms and what ends the level, and Zugo builds a 2D browser platformer you can play with the arrow keys: gravity, collision with ground, hazards that reset you, and a goal at the end. A build costs 6 credits and each edit 3, and nearly all the editing goes into the jump.
Everything else in a platformer is furniture. Players forgive ugly tiles and forgettable enemies, and they do not forgive a jump that lands a pixel short of where they aimed. That is where the credits should go.
What does the first build actually contain?
A character that runs left and right, a jump with gravity pulling it back down, solid platforms, a camera that follows, hazards that send you back to the last checkpoint, and something at the end of the level that counts as winning. That is a playable game on the first pass.
One prompt also carries a life counter, collectible coins with a score, a simple patrolling enemy, a double jump, and a level that scrolls sideways for a while rather than fitting on one screen. All of those are rules with clear edges.
What one prompt cannot deliver is a level that is interesting to play twice, because interesting level design is a sequence of decisions rather than a description. You get a corridor with obstacles, and you turn it into a level through edits.
The broader question of what generation reaches in games at all is answered in can AI build a game.
What should the prompt name?
Seven items. The first four make it work, the last three make it playable.
| Element | Weak version | Strong version |
|---|---|---|
| Character | "a guy" | "a small character that runs left and right at a steady speed" |
| Jump | "he can jump" | "a jump about three tiles high, holding the key jumps higher" |
| Ground | "platforms" | "solid platforms and floating ledges, the character lands on top of them" |
| Goal | "reach the end" | "a flag at the right edge of the level ends it with a win screen" |
| Hazards | (unspecified) | "spikes and pits send the player back to the last checkpoint" |
| Camera | (unspecified) | "the camera follows the player horizontally with a slight lead" |
| Controls | (unspecified) | "arrow keys plus space, and on-screen buttons on touch devices" |
The camera line is worth spelling out. A camera locked exactly to the character makes players motion sick on a fast level, while a camera that leads slightly in the direction of travel shows what is ahead and feels professional at no extra cost.
A prompt assembled from the table: "A 2D side-scrolling browser platformer. A small character runs left and right at a steady speed and jumps about three tiles high, jumping higher when the key is held. Solid platforms and floating ledges, spikes and pits that send the player back to the last checkpoint, three lives. The camera follows horizontally with a slight lead. A flag at the right edge ends the level with a win screen. Arrow keys and space, with on-screen buttons on touch devices."
Which edits fix the jump?
Four, in this order. Each one is a single edit at 3 credits and each one is felt immediately.
Ask for coyote time. Give the player a brief window after walking off a ledge where the jump still works. Nobody notices when it is there, everybody feels cheated when it is missing, because the game reads as ignoring inputs.
Ask for a jump buffer. If the player presses jump slightly before landing, hold that press and jump on touchdown. Between coyote time and a jump buffer, most of the "the controls are unresponsive" complaints disappear.
Ask for variable jump height. Tapping should give a short hop, holding should give a full jump. Without it, every jump is the same height and platform spacing becomes an exercise in memorization.
Ask for faster falling than rising. A character that comes down slightly quicker than it went up feels weighty and precise. A symmetric arc floats. Phrase it as "increase gravity while falling so the descent is faster than the rise."
Why does the level feel empty?
Because a generated level is usually a flat line with obstacles placed at even intervals, and evenness is what makes it forgettable. The fix is rhythm, and it costs one edit at a time.
| Symptom | Cause | What to ask for |
|---|---|---|
| Nothing is ever risky | Gaps too small | "widen two of the gaps so they need a full-height jump" |
| The whole level looks the same | One platform height | "vary platform heights so the player moves up and down" |
| Enemies are ignorable | No threat | "an enemy that patrols a ledge and resets the player on contact" |
| No reason to explore | Single path | "a higher ledge with three coins that needs a precise jump" |
| The end arrives suddenly | No pacing | "an easy opening stretch, a harder middle, a short calm run to the flag" |
Add these one at a time and play after each. Platformer difficulty is not additive: a wider gap plus a patrolling enemy plus a moving platform can turn a fair section into a wall, and you only see it by playing.
The practical loop for this kind of incremental change is described in how to edit after generation.
Does it work on a phone?
Only if you ask. A keyboard platformer opens on a phone and cannot be played, which is the most common reason a shared link gets closed in three seconds.
Ask for on-screen controls explicitly: a left and right pair on the bottom left, a jump button on the bottom right, both large enough for a thumb and both semi transparent so they do not hide the level. That is one edit.
Then ask for the game canvas to scale to the screen instead of assuming a desktop window. A platformer designed for a wide view and shown on a narrow phone screen either crops the level or shrinks the character to an ant, and both are fixed by asking for the view to fit the device.
Test it yourself on an actual phone before sharing. Touch controls that look fine in a desktop preview can sit exactly where your thumb rests, and no description catches that.
What does a platformer cost in credits?
A build is 6 credits and an edit is 3. Hi-Fi mode doubles both to 12 and 6.
Realistic spending: one build, four edits for jump feel, two or three for level rhythm, one for touch controls, one for a title screen. Ten or so edits on top of the build gets you something you would share. Pro at $25 a month gives 200 credits, which is 66 edits if you spend nothing else, and Business is $99 a month for 800 credits.
Free gives 5 credits once, slightly less than a single build, so it shows the tool rather than finishing a game. Zugo also ships 25 templates with 5 of them games, and starting from a working game and reshaping it usually costs fewer edits than describing one from a blank page. A simple build takes about a minute.
Where is the ceiling?
Four limits worth knowing before you start.
- 2D and browser. No 3D, no console build, no installer. That is the shape of the product.
- A long game is authored, not generated. One level generates well. Eight levels with a difficulty curve, a new mechanic every second level, and a boss at the end are a sequence of decisions you make edit by edit.
- Saved progress across devices needs a backend. Progress in the browser works immediately. A player who switches phones and keeps their save means connecting Supabase, which is its own task.
- Art and sound carry identity. A generated character is a placeholder. Your own sprites and a jump sound are what make it yours.
Each build is launched in a sandbox before it reaches you, and one that fails to open is reported as a failed build. That removes the blank-screen risk and says nothing about whether the jump feels good, which only playing answers. Zugo does not replace a development team on a complex product, and very specific behavior is refined through edits rather than captured in one prompt.
What should you do first?
Build one level, play it for two minutes, and spend your first four edits entirely on the jump. A platformer with a great jump and a plain level is fun. A platformer with a clever level and a mushy jump is not, and no amount of extra content rescues it.
Then add touch controls, then rhythm, then art. That order wastes the fewest credits, because every later change is judged against a character that already moves correctly.
A version of this genre where the obstacle is thinking rather than timing is covered in make a puzzle platformer with AI. Published games live at an address like your-game.zugo.run, open on phone and desktop, and can sit on a custom domain. Describe your character and your jump at zugo.dev and see how the first landing feels.