Week 4 · September 5 to 11. The school is up in four seconds · Logbook

Log entry: day 20

Shadows again. On the edge of the chair's shadow, a staircase of little steps. The third pass in three days. I sent three shots and wrote: "I think we were fixing the wrong thing, and not globally. Look, the chair has a staircase too. And the potted plant and the drinking fountain too! But the wardrobe's is smooth, for example." The answer was in my own list: the smooth shadow by the wardrobe came from the sun, and all the stepped ones came from the hidden light inside the rooms. The flight engineer made a penumbra.

I asked straight out whether we'd got crisp, even lines. The answer was honest: "No." The steps were gone, but the edge went soft. On the fifth pass we tried laying the shadow onto the wall by hand, and the edge came out razor sharp.

A test: the drinking fountain's shadow on the wall, the edge is clean

The drinking fountain actually had two shadows. The drawn one was built from the ceiling lamps, and the real one fell from the room's hidden light, which hung somewhere else and lower. One object got two shadows of different lengths, pointing different ways. And the lamps themselves give no light at all.

Meanwhile, in another session, we were working on the style and the interface. I remembered how powerful the wanted stars were in GTA, while we had feeble hands instead. For the attention meter we compared a star and a bell. I picked the bell: "That's better." About the star, the note on the canvas said it was taken whole from GTA and says nothing about school.

The feeble hands were five little fists with the index finger raised. The code drew them directly, out of rectangles, without a single image file.

Insight: the problem was what the symbol was drawn with

First the flight engineer drew six candidates: a bell, a star, an eye, a crest, a tie, a square from the register. I didn't like any of them. I wrote: "Remember how powerful the stars were in GTA. And we've got some feeble hands. It needs to be something really solid, that fills in so you can see it, and the unfilled one too." It wasn't about the symbol, it was about what it was drawn with. Back then all the game's graphics were line, that's how all 48 symbols in the clouds were drawn. And the star in GTA is filled solid, with an identical empty one next to it. A line next to a fill will always look feeble, however thick you make it. The flight engineer admitted: "You caught the real cause, and I didn't." All the symbols were redone as filled silhouettes, and then I picked the bell.

The canvas with the meter: the bell chosen, the star rejected

In the same place I said belief in the ghost shouldn't be shown so strongly, and maybe it shouldn't be added at all. Belief would live on in the game for another month.

Style canvases, September 5

The caught screen: CAUGHT, the deputy head and a note in the school diary, "Roaming the halls during lesson"

A game frame on the canvas: attention bells, tasks on a scrap of paper, a backpack with five slots

Six candidates for the meter, already filled

The brand block COOL SCHOOL · GHOST. That was the game's name back then

Toward evening I wrote: "I'm tired of these shadows." And asked whether we'd gone overboard. The flight engineer went to see how other games lay down flat shadows: Toy Story 3, Oblivion, Sonic Adventure 2, Tanki Online. The verdict: we hadn't gone overboard, we'd arrived at what the textbooks say.

Insight: we should have drawn it from the start

For two days the flight engineer tried to make the engine's real shadow give a sharp edge: resolution, turning the light source, penumbra, threshold. While the research was going on, it put one drawn decal shadow by the drinking fountain and was pleased: "The decal works, and we rejected it without testing." The research gave two answers. Decals don't suit us: on a wall you get a vertical smear, and the shadow stops stretching away from the lamp. And what's expensive isn't the drawn shadow but the real engine shadow under it, which you can't see anyway: about eight milliseconds and 128 megabytes of memory every frame. "We're paying for two layers in one pixel." And the main admission: "A sharp edge just can't happen there. We should have drawn it from the start." I replied: "Yes, let's do it all properly."

Hunting the edge of the shadow

September 5, my shot: a staircase along the edge by the plant and the fire extinguisher

But the wardrobe's shadow is smooth. Later it turned out to be from the sun

September 6: a classroom from the doorway, the shadows look doubled

After the penumbra: no steps, but the edge is blurred again

Testing a decal by the drinking fountain: the patch landed on the floor and the wall at once, but as a rectangle

Log entry: day 21

I wrote: "we overcomplicated … with all these shadows... we could just have the simplest crisp shadow from objects. I had that once." Going by my shot, the flight engineer removed the layer of "outline" shadows. I looked: in the toilet there were no shadows left at all, and none from the desk either. 35 minutes later the layer was back. We wrote it down honestly: removing it was the wrong move.

By evening the shadows were hanging in the air again.

The janitor walks out of the restroom through a puddle

And in the evening I said: make a commit, we'll sort it out later. That's the mistake of the week, more on it below.

Log entry: day 22

In the morning the work since September 2 went into the main branch: 431 files in one commit. The shadows stopped hanging in the air. Before that, "black slabs hang across the desks in the classroom". The kid's shadow fell on the side panel of a desk, 11 centimeters high, and stretched seven and a half meters across.

Soon it turned out that the same commit had broken the kid's model. The model got ruined three times in a day, the same way each time. The protection against this was listed as switched on and did nothing.

Another check went red 113 times in a row, and never for a real reason. Its measure of rotation gave a tiny number for two identical rotations, and a flat zero for a real difference. When the measure was replaced, two of the 113 complaints were left, and both were right: every body had two extra pose handlers hanging on it.

Then I asked: what if we make the models in Blender? And thirteen minutes later I wrote that I'd installed Blender. I opened the kid's model and asked whether it was normal for it to have so many points.

Handing the model over to Blender completely didn't work: three quarters of it isn't geometry, it's a program. We agreed on a loop: the shape gets edited in Blender, the numbers go back into the game. The first thing from Blender was a beanie.

We tested the loop on the head. Moved it five centimeters up in Blender, and exactly one line changed in the model. But on the way Blender swapped the neck and the chest, and eight nodes with the same name came back numbered. We had to start a census of names. And in its first day the bridge delivered only that beanie, and nobody wears it. The receiver replaced every mesh it brought with a box of the same overall size, so the beanie landed in the game as the same three boxes you'd have typed in as text. In that time the model was edited three times, and all three times by hand, bypassing the bridge.

Then came the day of the knee socks. The white knee socks stuck out like boxes above the shoes. I wrote: "well, the sock should be kind of part of the shin and move with it." They were taken in, and a gap of almost seven centimeters opened between the shin and the shoe, on all fifteen looks. I suggested: "Maybe we just make the legs go down to the shoe by default, and then the sock goes on top. That way we'd be covered." That's what we did.

Knee socks sticking out like boxes above the shoes

The day of the knee socks

Under the desk: the shoes apart from the legs, the knee socks slipped down

Blender: the bridge brought all the poses along with the model, and I got lost

In the corridor: a white stump of knee sock between the shin and the shoe

The shoes in Blender. I wrote: "But in Blender it looks nice"

The pose book: the service RESET comes first now, then 33 poses

The idle stand after I saved it: there were six, and none were left

On top of that, the nerd had "some leftover of somebody else's hairdo on the side", and somebody had "a visor growing out of the forehead". In the evening I ran the whole game, and stray bars of hair were sticking out of every model's head again. The stray lock was removed from all fifteen.

**Before.** Stray bars of hair at the temples

**After.** A visor "out of the forehead"

The stray lock was a box by the right temple. It lived apart from the hairdo and poked through whichever one was put on. For the cat, the teacher and the janitor it was the only thing left of their hair at all. It couldn't be deleted: the hair swing hung on it. So now the lock is switched off instead of deleted.

There was good stuff that day too. Everyone who stands now breathes. Stand for six and a half seconds, shift from one foot to the other.

For them we built an idle stand: six students right in the editor. I opened the scene, saved it, and instead of six there was nobody. The flight engineer wrote: "This is exactly the trouble that ruined the kid's model three times in a day, and I set it up with my own hands in a new file." Then the stand had to be turned around as well: the six were standing with the backs of their heads to the camera.

Log entry: day 23

In the morning the flight engineer took the light apart layer by layer: switching off the sun, the lamps, the room light and the shadows one at a time, to see what each one gives.

The same classroom: each frame without one layer of light

Insight: the softness comes from the lamplight, not the shadow

I complained again: "The shadows are still kind of complicated, maybe some are excessive or extra. And they should be flat. We're making cyberpunk, like anime with flat shadows." The layer breakdown showed that six systems draw shadow and darkening at once. The sun is already hard, its edge one pixel wide. But 178 light sources in the rooms skip the light steps and come out as an ordinary smooth fill, which is where the gradient comes from. The flight engineer sent three versions of the same corner, and I picked the third, with a crisp shadow from the chair. A check showed that the third one brings back a round patch with an edge on the wall, the very thing the room light had once been hidden for. The flight engineer offered two ways forward, but I didn't answer: I'd already opened a new session, and the day went into architecture.

Three versions of one corner: as it was, lamps in steps, plus a hard edge

The trap: take the blur off the sun shadow, and the floor turns into moiré

I wrote: "I'm afraid my current approach is too much. And not best practice." And I brought in two outside reports made by other AIs. The flight engineer went through them against our code. Most of the advice was rejected, with a note on why for each one.

And then the main thing turned up. At every launch the school took 23 seconds to assemble, and 18 of them went into finding places for the furniture. Every beanbag looked for its spot through a tree of fifty thousand nodes. Now the school is up in four seconds. Placing the beanbags used to take two and a half seconds, now it takes nine thousandths.

And I said something important: "As the game's architect I'm supposed to place everything myself, to build the composition I need." Classroom 1A can now be arranged by hand in the editor. A spot placed by hand wins over the calculation.

Log entry: day 24

No commits, no messages.

Log entry: day 25

I asked: "Do we really need to 'assemble the school'?" The flight engineer said we didn't, and I wrote: "Yes, finally. Let's start with the first one and go." Now the school is baked into a scene in advance: 50,147 nodes in a 5.8-megabyte file, exactly as many as when it was assembled.

A check runner appeared: one command runs them all. There are 118 checks, and 18 of them are silent, meaning they can't turn red.

Log entry: day 26

23 rooms can now be arranged by hand: eight classrooms and fifteen restrooms. The first full run: 91 green checks, 4 red, 12 silent. I asked whether we really needed all these checks, or whether we'd slid into overengineering.

Then I played: "flooded the toilet and sat back down… an announcement about the flood started on the radio, and everyone's like ghost! ghost! and at the same time the teacher went to the window to look at the storm, and it all froze in one picture like that, and the camera froze." The camera got stuck on a cutaway. Now after thirty seconds the view comes back by itself.

That frozen frame: the classroom during a lesson

In the evening I asked to check all the end conditions of the demo. It turned out the third line of the demo couldn't be done: after the film the blinds didn't go up. Now the demo can be played to the end. And two red checks turned out to be wrong themselves: they looked in the first desk drawer, and by design that's where the chalk and a pen are.

After the film the blinds now go up without the sound of the cord. Eight clicks of the slats turned out louder than the lesson's noise threshold, and the class would have called a witness on a player sitting quietly at a desk.

Where it came from

GTA 4 and 5: the wanted stars as the yardstick for the interface, which we left behind for the bell. Toy Story 3, Oblivion, Sonic Adventure 2 and Tanki Online: how others lay down flat shadows. Other people's video tutorials on skeletons and animation that I brought in. The Sims: the ceiling and near walls come off when you look from above.

The week in numbers

37commits
23 → 4 sschool launch
431files in one commit
113times in a row one check went red, never for a real reason
7.5 mwidth of the shadow from an 11 cm panel
3times in a day the model got ruined