How the four-bot build timelapses are made — a making-of

The build timelapse series on KangJu Playground fits a modern villa going up from an empty deck into a vertical Short of about 30 seconds, with four characters working around the building. This is how those videos are made, and what we got wrong on the way.

The parts — design, placement, bots, camera

• Design: a program generates the building from a seed. Storeys, width and palette change with the seed; the shape alternates between a four-storey Sky Villa and a six-storey Stack Villa.
• Placement: blocks are placed on a Minecraft Java server by remote command (RCON setblock). The bots do not place blocks by hand.
• Bots: four characters join the server through a Java bot library (mineflayer). They walk to where blocks are going down and swing their arms, so the scene reads as building.
• Camera: a spectator view rides an armour stand at set positions, and the screen is recorded and assembled.

The bot library is Java-only, so this pipeline is separate from the Bedrock build-shorts series. That is also why the site's patterns come in Bedrock and Java flavours.

Why villas and not castles

At first we rotated castles, towers and villages for variety. Three episodes in, what was missing was consistency, not variety. Even among our own videos, a 2,711-block villa read far better than a 4,101-block castle. Stone castle walls hide whatever is inside; a glass villa shows its furniture. Contrast of materials and a visible interior make the picture — not block count.

Mistake 1 — the house was off screen

Shorts are vertical (9:16). Minecraft's field of view is vertical, fixed at 70°, and only the horizontal angle follows the aspect ratio. Worked out, that is about 102° across in a landscape frame but only about 43° in a vertical one.

The camera-distance code still had 102° in it. For six episodes the camera was placed for a wide frame and exported as a narrow one. Measured afterwards, 28% of the house, and for one shape over a third, was outside the frame. Our preview tool drew without perspective or aspect ratio, so it never showed; now a tool renders the first frame exactly as the capture will.

It changed which shapes we build, too. Wide villas become a thin strip in a vertical frame, so only the two tall shapes stayed. The six-storey Stack Villa fills 71% of the frame's height.

Mistake 2 — the bots were not really working

Someone said the bots looked like they were wandering rather than building, so we measured. When a block went down, the nearest bot was on average 9.6 blocks away, and only 18% of blocks were placed within arm's reach (5 blocks).

We first concluded it was a speed limit. Then the code showed that bots did not walk at all on ticks with no block to place. With a longer build more than half the ticks were like that, so bots moved at half their set speed. Walking every tick, a wider set of candidate blocks near each bot, and a rule that nothing goes above an empty cell raised within-reach placement to 55–82%.

Bots also stood glued to one spot: while a long wall went up, each bot kept choosing the same standing position. Mixing a counter into that choice cut time standing still from 18% to 2.2%.

Mistake 3 — smoothing it made it choppy

The game renders only about 21 frames per second while building. To shorten videos we ran the build at 2× speed and heard that the first two-thirds were too choppy to watch. 2× had more distinct frames per second, but uneven gaps between them — judder. Trusting one number was the mistake.

So building is filmed at constant speed, and length comes from building faster in game, not from speeding up video. The orbiting camera went too: cameras now hold still at fixed positions and hard-cut between them, which removes rotation stutter entirely.

Why the video and pattern block counts differ

The video includes landscaping — trees, flowers, hedges. The pattern keeps only the building you would follow, so Stack Villa #09 is 2,935 blocks in the video and 2,066 in the pattern. And when the design code has changed, an old seed produces a different house, so those episodes get no pattern.

What we took away

• Measure before giving up. An "it's mathematically impossible" conclusion was hiding a bug.
• One number does not beat your eyes. Smoothness was frame spacing, not frame count.
• Previews must match the real output. A preview without the aspect ratio fooled us for six episodes.