Sand Garden.

I asked to see the mathematics in two dimensions before taking the pattern further. The AI-assisted build had produced paths that looked plausible, but the tool and its marks had been worked out separately.

One moving rake became the source of every groove. All four teeth share a position and orientation, and the same path drives the 2D study and the 3D scene. Checking the teeth through a turn made the correction visible before more material work hid it.

That rule governs the tool. It does not force someone’s drawing to become a perfect circuit. Auto uses composed paths; a person can stop a stroke wherever they want.

The hand now describes a path while the rake follows it at a controlled pace. The waiting curve is smoothed and remains editable. It is the route the rake will take, rather than a second line painted on the sand.

A test with the same gesture at three input speeds exposed the old timing problem and checked the replacement. The rake’s peak speed stayed at its configured limit across all three.

The surrounding controls follow the same approach: handles on the line, tools beside the board, and six composed gardens to begin with. Someone can also leave Auto running without drawing anything.

Each pass replaces the swept strip beneath the rake before adding its new grooves and ridges. Shallow height changes give the marks some displaced sand without turning the project into a full sand simulation.

This also shaped the interaction. There is no undo stack: another pass can change a mark, and the smoother can work it away. Photo mode and replay preserve a moment or its process without changing that rule.

Auto works on the same surface. It keeps making new compositions over existing sand instead of clearing the board between patterns. Switching to a manual tool keeps what is there.

The upload path still sent large buffers for small changes. Packing only the changed area cut the recorded short-contact update from 172 ms to 1 ms in a Mac WebKit test. Output comparisons checked that the faster path left the same sand behind. Retesting on my iPhone confirmed that raking worked well there.

Lowering the resolution was less useful. It passed a desktop fixture, still failed on the Pixel, and made the sand visibly worse. Full-resolution textures and antialiasing returned; the lighter mode instead reduces terrain detail, filtering, shadows, and particles.

The Pixel failure remains unresolved. The app now offers a restart that preserves the garden if graphics stop. TV testing also caught unnecessary rendering above the screen’s native density, which could waste work even with the rake lifted.

Raking worked well on my iPhone after the upload change. Testing on the TV exposed a different problem: it could be slow even when the rake was lifted. Those observations helped separate the cost of changing the sand from the cost of drawing the scene.

No one else has used it or given feedback yet. My own testing does not tell me whether a visitor understands the first stroke or wants to come back.

That is the next useful test: watch someone begin on their own phone, see where they hesitate, and find out whether they choose to rake or simply watch. Longer display sessions need testing too. The unresolved Pixel failure remains part of that work.

The original idea was to give visitors somewhere to spend a little time. There is now a working garden to put in front of them.