AI Game Dev Search

Your first platformer, and the three jump fixes players never notice

A first fox platformer from one prompt, then coyote time, jump buffering and variable jump height. Play each before and after, with our measurements and a run without the skill.

Tested with Phaser 3.80.1 · Claude Code · Opus 5.5

This is the real result. Click it and play. Open full screen

Before you start

  • Claude Code installed. Choose your tools shows how.
  • The skills, as one plugin. In Claude Code run /plugin marketplace add aigamedev-ai/platformer-kit, then /plugin install platformer-kit@aigamedev.
  • An empty folder with a copy of Phaser in it (phaser.min.js, version 3.80.1). Open Claude Code in that folder.
What we asked“Add coyote time to the jump, so I can still jump for a moment after running off a ledge. Test it.”

What happened

You
Add coyote time to the jump, so I can still jump for a moment after running off a ledge. Test it.
  • Used the platformer-feel skill
  • Read index.html and the skill's jump notes
  • Checked that Playwright and headless Chrome work
Claude
The game already has coyote plumbing but COYOTE_TIME = 0. Coyote logic is already wired in update(): refilled on the ground, counted down, cleared in startJump, and it doesn't use up the double jump. It just needs a non-zero value.
  • Changed COYOTE_TIME from 0 to 0.10
  • Wrote tests/coyote.test.mjs
  • Ran it: all four cases pass
Claude
I added coyote time and tested it: you now get 0.10 s after running off a ledge where jump still works as a normal ground jump. The test runs the fox off the first ledge and presses jump a set time after it leaves the ground. A press at 0.05 s and at 0.08 s is a ground jump, with the double jump still available. A press at 0.20 s is too late, so the double jump fires instead. With coyote time off, the 0.05 s press only spends the double jump. Separate bug I found but didn't fix: the release handler is only called for a buffered tap and never when you let go of jump, so every press gives the full-height jump. Do you want me to fix that?
You
When I press jump just before landing, the jump doesn't happen when I land, and it uses up my double jump instead. Make the jump still count if I press it a moment early. Test it.
  • Used the platformer-feel skill
  • Read the game, the coyote test and the skill's buffer notes
  • Changed JUMP_BUFFER from 0 to 0.12 and added a landing check
  • Wrote tests/buffer.test.mjs
  • Ran it at 30, 60 and 144 fps
Claude
Pressing jump just before you land now gives you a normal jump when you touch down, and your double jump is left unused. When you press jump in the air, the game works out whether you will touch a platform within the next 0.12 s, using the same gravity and fall speed as the real fall. If you will, it holds the press and jumps the moment you land. If not, the press is your double jump, as before. The last row of the test turns the fix off and shows the old bug, so the test does catch it.
You
I want a quick tap to give a small hop and holding jump to give a full jump. Check how it works now and fix it if needed. Test it.
  • Used the platformer-feel skill
  • Wrote tests/variable-jump.test.mjs first: 9 of 12 checks failed
  • Made letting go while rising cut the upward speed
  • Raised JUMP_CUT from 0.7 to 0.8 because a tap reached only 25 percent
  • All checks pass at 30, 60 and 144 fps
Claude
A quick tap now gives a small hop and holding jump gives the full jump. Before the fix, they didn't: once a jump started, the hero kept rising for the full 0.18 s whether you were still holding or not. A one-frame tap went the full 100 px. While the hero is rising, letting go of jump now ends the rise straight away and multiplies the upward speed by JUMP_CUT, once per jump. The full jump is the same height as before, so every ledge in the level can still be reached.

This is part 1 of a series where we build one complete platformer with Claude Code and the skills from the platformer kit. Every part is a real, unedited session: we paste a prompt into a fresh Claude Code, and the next part starts from what the last one made. The game is Fox Courier: a fox carries a letter across an autumn forest to a lighthouse.

Step 1: a small game that runs

We asked for the smallest thing that is still a platformer.

Make a small browser platformer in this folder. Everything goes in index.html, using Phaser 3.80.1 (phaser.min.js is in this folder). The hero is a fox courier, drawn as an orange square for now. Arrow keys or A/D run, Space or W jumps, and you can double jump. Add a few platforms with coins to collect, spikes that send you back to the start, and a flag at the end that shows "Letter delivered". Play it yourself in a headless browser with Playwright (it is in node_modules) to check that you can run, jump, double jump, collect a coin, die on spikes and reach the flag.

The last sentence matters. Asking Claude to play its own game in a headless browser is what turns “it should work” into “it ran and these checks passed”. It took 10 minutes and cost $0.81.

The game from step 1. Run right, jump the pit, double jump. Keep this one in mind: it is the "before" in everything below.

Step 2: the jump

Players judge a platformer by its jump. Three small helpers make most of the difference, and players never notice them when they work. Claude’s first game already contained all three, so to show what each one does we switched them off in the game above, changed nothing else, and asked a fresh Claude Code to put them back, one at a time, with the platformer-feel skill installed.

Coyote time

Coyote time lets you jump for a short moment after you run off a ledge. Your eyes say you’re still at the edge, so the game agrees with them.

In the demo, the tinted strip past the edge is the ground that coyote time adds: 0.10 s at the fox’s run speed of 180 px/s, or 18 px.

An explainer we built for this page. Press jump a moment after the square leaves the edge, with coyote time on and then off. Slow motion helps.

Add coyote time to the jump, so I can still jump for a moment after running off a ledge. Test it.

The session took 66 seconds and cost $0.33. Coyote time is hard to feel, so we added labels to the games on this page: a yellow strip at the edge of the pit marks the coyote zone, and the top of the screen says what each jump was (“COYOTE JUMP, 60 ms after the ledge”, or “DOUBLE JUMP”). Run off the first ledge (the pit at the end of the start ground) and press jump a moment late. In the first game the late press is a double jump and your second jump is gone. In the second it is a coyote jump and you keep the double jump.

Before: coyote time off. A late press only gets the air jump. The labels are added for this page.

After: Claude's version. The same press is a full ground jump.

We measured both ourselves. We stepped each game on a fixed 60 fps clock, ran the fox off the ledge, and pressed jump a set time after its last frame on the ground. “Ground jump” means the fox rose and still had its double jump.

Press after the ledgeBeforeAfter
0.05 sDouble jump spentGround jump
0.067 sDouble jump spentDouble jump spent
0.20 sDouble jump spentDouble jump spent

Claude set COYOTE_TIME to 0.10 s, the number the skill recommends. We measured less: the window is only about 0.06 s, because the timer ticks down once before the jump check reads it. Keep that in mind for the comparison further down.

Jump buffering

Jump buffering is the opposite problem. You press jump a little too early, just before you land. The game remembers the press, and if you land soon after, the jump happens the moment you touch down.

An explainer we built for this page. Bounce the square and press again just before it lands, once with buffering on and once with it off.

In our game the early press did something worse than nothing. While the double jump was still unused, an early press spent it in mid-air, and the fox then landed without jumping.

When I press jump just before landing, the jump doesn't happen when I land, and it uses up my double jump instead. Make the jump still count if I press it a moment early. Test it.

The session took 85 seconds and cost $0.37. Jump once, then press jump again just as you come down. Don’t double jump first. The label at the top says “DOUBLE JUMP” when the early press was wasted, and “EARLY PRESS KEPT” when it waited for the ground.

Before: the game after step 1 of this part. The late press fires the double jump, and you land without jumping.

After: the same press waits for the ground, and you jump again as soon as you land.

Our measurement drops the fox onto the ground and presses jump a set time before it lands. A buffered press keeps the double jump.

Press before landingBeforeAfter
0.05 sDouble jump spentJumps on landing
0.08 sDouble jump spentJumps on landing
0.12 sDouble jump spentJumps on landing
0.30 sDouble jump spentDouble jump spent

The 0.30 s row is right: that early, you meant the double jump. Claude did not simply start remembering every press. It checks whether you will land within 0.12 s, and only then holds the press for the ground.

Variable jump height

Variable jump height gives you a small hop when you tap and a full jump when you hold. Letting go while the fox is still rising cuts its upward speed once, so the jump tops out sooner.

An explainer we built for this page. Tap to get under the spikes, hold to reach the ledge. Turn variable jump off and you can't get under the spikes.

I want a quick tap to give a small hop and holding jump to give a full jump. Check how it works now and fix it if needed. Test it.

Both games below show “JUMP HEIGHT” in pixels after every jump, next to the full jump height of about 100. Tap the key as fast as you can, then hold it: you should see about 33 px and then about 100 px. In the game before, both show 100.

Claude wrote the test first. It failed 9 of 12 checks: once a jump started, the fox kept rising for the full 0.18 s whether you held the button or not. A one-frame tap went the whole 100 px. The session took 89 seconds and cost $0.33.

Before: tap and hold both give the full jump.

After: tap for a hop, hold for a full jump. This is the finished game for this part.

Button heldBeforeAfter
Tap (one frame)100 px33 px
Hold 0.1 s100 px59 px
Hold until falling100 px99 px

Claude also raised JUMP_CUT from 0.7 to 0.8. At 0.7 a tap reached only 25 percent of a full jump, the very bottom of the skill’s range of 25 to 50 percent, so it picked a hop of about a third. The full jump stayed at 100 px, so every ledge in the level is still reachable.

What changes with the skill

A fair question is whether the skill matters at all, since Claude Code can fix a jump without help. So we ran the same three prompts, in the same order, from the same starting game, in a second folder with no skills. Same model, one run each.

StepWith the skillWithout
Coyote time66 s, $0.33376 s, $0.55
Jump buffering85 s, $0.3798 s, $0.33
Variable jump89 s, $0.3384 s, $0.23

Both folders ended with a jump that works. We measured the no-skill builds the same way:

CheckWith the skillWithout
Coyote press at 0.05 sGround jumpGround jump
Coyote press at 0.067 sDouble jump spentGround jump
Buffer press at 0.12 sJumps on landingDouble jump spent (its window is 0.1 s)
Tap height33 px (33 percent)26 px (26 percent)
Tests left in the project3 files in tests/None, they were written to a temp folder

Single runs vary, so don’t read much into the times. Two differences are real, because they come from what the skill says. First, the skill sessions leave their tests in the project, which is why the buffer and variable-jump sessions could re-run the earlier tests to check nothing broke. Second, the skill names the good ranges, so the skill session knew a 26 percent tap was too small and moved it.

What to take from this

  • Ask Claude to play its own game. Every check in this series is a script that presses keys and reads numbers.
  • Ask it to check, not only to add. The three bugs were things the code already “had” and got slightly wrong.
  • Keep the tests. A test in the project is a test the next session can re-run.

Get the skill. Next: rooms that teach, and a spring that grows the game.

The skill that made it work

It gives Claude measured numbers for a good jump and makes it test every helper with scripted key presses.

platformer-feel
  1. Keep every movement number as a named constant, in seconds and pixels per second, so you can tune it.
  2. Check the three helpers that good platformers share: coyote time, jump buffering and variable jump height. Read each helper's notes before you touch it.
  3. Test with scripted input, not by reading the code: each helper gets a test that fails without it and passes with it.
  4. Finish by listing each helper with its number and what the test showed, in plain words.

Save it as .claude/skills/platformer-feel/SKILL.md in your project.

Was this useful?

Part 1 of Make a complete platformer with AI

Next in Make a complete platformer with AIRooms that teach, and a spring that grows the game
Open →