Rooms that teach, and a spring that grows the game
Two prompts built twelve rooms: six that teach run, jump and double jump with no text, then six around a new spring pad. Play them, see every room measured, and compare with a run without the skill.
This is the real result. Click it and play. Open full screen
What we asked“Add a second area of about six rooms that brings in one new thing: a spring pad that bounces the player high. Teach it the way the first area teaches the jump, mix it with the double jump, and end the area with a room that tests both.”
What happened
- Used the platformer-levels skill
- Read the game, the old tests and the skill's measuring notes
- Ran the existing tests first: all pass
- Wrote the six rooms as data, each with a one-line story
- Measured every room with the skill's script: it plays each room 200 times with simulated players
- Added single-jump routes to prove rooms 4 to 6 really need the double jump
- Widened the pits in rooms 5 and 6, raised room 6's block from 130 px to 145 px
- Found and fixed a bug: a coin touched from above counted as ground and refilled the double jump in mid air
- Rewrote the parts of the old tests that assumed the old layout
The game from part 1 has one short level. A good level teaches without a single word of text: each room asks the player for one new thing, in the same order every time. Show it safely, use it, combine it, test it.
The most useful tool for this is knowing what one jump can reach. The demo below grades a platform against the fox’s real jump. A ledge a few percent past the reach is the worst: players keep trying, and keep missing by a few pixels. One that is clearly out of reach teaches them to try the double jump.
An explainer we built for this page, with the fox's jump numbers. Drag the platforms around and watch how each jump is graded.
Step 3: six rooms that teach
Turn this into a short first level of about six rooms that teaches a new player to run, jump and double jump, then ends at the flag. Make sure every room can actually be finished.
It took 19 minutes and cost $2.14. Claude loaded the platformer-levels skill, wrote each room as data with a one-line story, and measured it with the skill’s script, which plays every room 200 times with simulated players who press a little early or late.
Claude's level with the skill. Walk off the right edge of a room to go on. Spikes and pits restart the room you're in. Starts at The tall wall; the rest of the game is ahead of you if you want it.
| # | Room | What it teaches | New players clear it |
|---|---|---|---|
| 1 | First steps | Run: flat ground, coins leading right | 100% |
| 2 | Hop the crates | Jump: a 40 px crate, then a 56 px one; nothing can hurt you | 100% |
| 3 | Mind the gap | Jump with risk: a 100 px pit, then one spike | 97% |
| 4 | The tall wall | Double jump: a 140 px wall, when one jump reaches about 104 px | 99% |
| 5 | Long leap | Double jump with risk: a 270 px pit, when one jump covers about 150 px | 91% |
| 6 | Delivery | Everything together | 74% |
Rooms 4 to 6 can’t be passed with a single jump, and the script checked that: without the double jump you stop at the wall or fall in the pit.
The same prompt, without the skill
To see what the skill adds, we gave the same prompt to a fresh session in a second folder, starting from the same game with no skills installed.
The level from the run without the skill. Walk off the right edge to go on. Starts at the double-jump room; the rest of the game is ahead of you if you want it.
| With the skill | Without | |
|---|---|---|
| Time and cost | 19 min, $2.14 | 11 min, $0.94 |
| Rooms | 6, one story line each | 6, one screen each |
| Text on screen | None | A sign in each room: “Space to jump”, “Hold jump to go higher” |
| Proof every room can be finished | A bot clears each room, and fails without the double jump | The same, and also a bot that takes off 15 px early and late: 39 checks pass |
| Numbers a designer can use | A pass rate for each room, for new, average and skilled players | None |
| Bugs found on the way | The coin that refilled the double jump | None reported |
Both levels can be finished and both teach in a sensible order, and the no-skill session did something we like: it checked that each route survives a 15 px early or late take-off. Two things differ. The skill’s level teaches without words, because the skill says to show it safely and use it before explaining anything; the no-skill level puts a sign in every room. And only the skill session can say how hard each room is: 100, 100, 97, 99, 91 and 74 percent. That table is what lets you tune a level without finding six friends to play it.
The skill session was also slower and more than twice the price. It measured every room 200 times and tuned it. If you only need a level that can be finished, the no-skill result is enough.
Step 4: one new thing, and an area for it
One new mechanic is enough for a whole area. The prompt asks for the same lesson order again:
Add a second area of about six rooms that brings in one new thing: a spring pad that bounces the player high. Teach it the way the first area teaches the jump, mix it with the double jump, and end the area with a room that tests both.
It took 26 minutes and $4.05. Claude first worked out what the spring can do. It launches you about 293 px, the double jump reaches about 200 px, and a spring plus a double jump at the top reaches about 396 px. Then it built the rooms around height:
| # | Room | Beat | What it asks | Measured |
|---|---|---|---|---|
| 7 | Spring camp | rest | Walk in. A pad on a ledge out of reach previews the new thing | 100% |
| 8 | Boing | introduce | Walk into a pad in front of a 240 px shelf a double jump can’t reach; nothing can hurt you | 100% |
| 9 | Bramble steps | develop | Run off a shelf onto a pad in the thorns; jumping at the edge, as world 1 taught, overshoots it | 77% |
| 10 | Second wind | mix | The spring stops about 70 px short of a 368 px cliff, so jump again at the top | 83% |
| 11 | Under the eaves | calm | A pad under a roof only bumps your head; an open one lifts you onto a hill | 100% |
| 12 | Express delivery | test | Spring, double jump and thorns together | 56% |
Each room has a note in its data saying what the player believes going in, what the room shows, the obvious wrong move and how it visibly fails. With the pads removed in a copy of the game, every room in this area fails, so the spring is needed everywhere. Most first tries fail (17 to 25%), but 84 to 98% of simulated players clear a room within three tries.
The whole game after this part. Rooms 7 to 12 start after the first six: walk off the right edge of a room to go on. The fox is still an orange square. Starts at Spring camp; the rest of the game is ahead of you if you want it.
What to take from this
- Ask for the lesson order, not for “levels”. “Teach it the way the first area teaches the jump” is what made the second area read as one idea. Without that sentence you get twelve unrelated rooms.
- Numbers keep the rooms honest. “A 368 px cliff, 70 px beyond the spring” is checkable. “A tall cliff” is not.
- Take the pass rates with a grain of salt. The simulated players only measure timing.
The skill that made it work
It makes Claude plan the game in areas, give each area one new element, and check every room against the jump numbers the game really has.
- Keep rooms as data (a list of tiles, hazards and pickups), separate from the game code.
- Write each room's one-line story first, then teach in order: show it safely, use it, combine it, test it.
- Check every room with the game's own jump numbers, then play it with a scripted route.
- Measure each room with simulated players, and report it the way a level designer would.
Save it as .claude/skills/platformer-levels/SKILL.md in your project.
Was this useful?
Thanks, we read every message.
Part 2 of Make a complete platformer with AI

