Designing the courses

There are twenty-three parkour courses on this site and I did not place a single jump by eye. Not because I am rigorous, but because the first few I did place by eye were unfair in ways I could not see from inside the editor, and building the tooling turned out to be less work than re-testing the levels every time I touched the movement.

The jump is asymmetric, which breaks the obvious formula

The character falls three times faster than it rises. That is the single biggest lever on how a platformer feels — it is why the jump has a snap to it rather than a float — and it has an awkward consequence for level design: the textbook airtime formula assumes one gravity, and with a fall scale of three it is wrong by roughly 40%.

Forty percent is not a rounding error. It is the difference between a gap you clear comfortably and one you cannot clear at all. So the arc is computed from the movement config rather than written into a comment, and every course is built against the computed value.

The version of this that nearly shipped was a sentence in a source file saying roughly “with this jump velocity the character rises about one and a half units, so every rise is at most one cell and every gap at most two”. That was true when it was written. It is a constant folded by hand, and the failure mode is horrible: change the jump, the comment stays, the courses stay, and some of them quietly become impossible. Nothing fails. Nobody finds out until a player cannot make the fourth jump.

Twenty percent is the difference between hard and unfair

No course is allowed to ask for a jump at the full computed reach. They are built against 80% of it.

The reason is that a jump designed at 100% is one you make by leaving the platform on the last available frame at exactly full speed. That is a frame-perfect input, and asking for a frame-perfect input is fine once in a boss fight and ruinous for a move the game asks for a hundred times in a run. The remaining 20% is not slack. It is the entire difference between a course that is hard and one that is unfair.

My Expert courses spend that margin. They are supposed to feel like they are on the limit, and they are — of what the design rules allow. The safety underneath it is not spent by anything, because it is also absorbing an optimism in the model: the arc assumes you keep full horizontal speed through the air, and air control in this game is 0.15, so somebody who takes off slow stays slow.

The unit is empty cells, and getting that wrong is how you break a course

This sounds like pedantry and it is the thing I got wrong most often.

A gap is measured in empty cells between platforms, not centre to centre. Two pads three cells apart have two cells of air between them. The character takes off at the leading edge of one pad and lands on the leading edge of the next, so two cells of air is what has to be flown, not three.

Confuse the two and every jump in the course is one cell harder than you designed it. That is enough, at the top of the ladder, to turn a deliberate jump into a frame-perfect one — by accident, invisibly, across a whole level.

Every course is proved finishable, including from every checkpoint

A course gets turned into a graph: every surface you can stand on is a node, and there is an edge wherever the arc says one can be reached from another. A test walks that graph and refuses to ship a course where the goal cannot be reached from the spawn.

That much is obvious. The check I would actually recommend to anyone building a platformer is the second one: the goal has to still be reachable from every checkpoint.

A checkpoint you cannot continue from is not a mercy, it is a soft lock. The run is alive, the clock is running, and the course is over — and unlike an unfinishable course, it only happens to the players who already made a mistake, which is the worst possible group to do it to. I have shipped exactly one of those and only because I found it before anyone else did.

The medal times are measurements, not targets

The same graph gets solved for its cheapest route: walking edges cost distance over run speed, jump edges cost their airtime. The result is a hard physical floor — it assumes instant acceleration, perfect pathing, and never looking at where it is going.

Nobody will ever match that time. It is not a target, it is a measurement of the course. The Author time is that floor multiplied by a single constant for how much slower a very good human is, and the other three medals are multiples of the Author time.

I want to be straight about which part of this is a guess. The constant is. It is one number representing an assumption about human play that I have no data for yet, and it is written down in the source as the least defensible thing in the file. What the derivation buys me is not accuracy — it is that a medal can never be impossible, because it is anchored to a route the geometry admits, and that it cannot silently rot, because moving one pad moves the floor and the test fails with the new value.

When there is real play data, that constant is the one thing that has to change, in one place, for every course at once.