Decisions

Jens and his friends had done tower defense plenty of times already. Six minutes after the first research came the memory of a mode where you pick one of twenty heroes and conquer an entire map – that was the better idea.

Jens wanted the Star Slayer setup; Claude then proposed switching from Phaser to PixiJS itself. Joystick, audio, art pipeline and the fixes for the installed iPhone app already existed. The extra work for a custom tile renderer was small.

The server decides hits, damage and loot; the phone only predicts its own movement. It ticks at 20 Hz, and only while someone is connected.

One hero, one wolf, one house and four icons per variant. A has been binding since, and its style prefix appears verbatim in every image request.

The map from back then is the strongest memory. Terrain and places follow the original; enemies, stats and art are new. Greenmark follows, further rebuilds later.

Measured: the same simulation produced different checksums in two JavaScript engines but the same rounded state. Since the server is authoritative anyway, custom maths isn't worth it. Only the world forge has to be bit-identical.

Tested alone on the iPhone 16. Everything ran cleanly, around 60 frames per second. The rig looked wobbly while walking – and Jens' line for the rest of the project: characters and movement are everything.

After five failed painted tests and three intermediate paths, D looked best. At 22:11 that evening Jens reported: no more stutter – only the cape should swing a bit more.

Jens noticed himself that the pipeline still used older models. Claude then compared the models on hero, monster and boss using real costs from billing data. The research budget had a hard cap; until the decision there were no final images.

The two friends should be truly impressed when they first connect. Until then they get no access and no images of the world – updates come spoiler-free.

The first sample was unusable except for one recording. The decision came the same morning.

Claude analysed 15 shifts and found that the code wasn't the bottleneck but the tooling – too many parallel tests on four cores and a performance test that measured the machine instead of the game.

Test scene from 150 down to 17 milliseconds per frame, smoke test from 151 to 58 seconds – with identical images.

Measurements perfectly smooth, saving works, camps fair. But: animations look bad, spells like a giant cloud, and you have to see every shot fly. That became around 20 feedback packages, shipped in four versions within about 15 hours.

For two hours every approach was meant to be tested. Then the trade-off: via the API, 3D only works for bipeds – heroes would look different from wolves and crabs. A consistent painted style matters more. Jens hardly cares about the art cost for it.