---
title: "Skills adds 3 game skills for mesh scans, raindrops and 3D rosters"
id: "story-92"
canonical_url: "https://emergent.fm/stories/92"
markdown_url: "https://emergent.fm/stories/92/article.md"
published_at: "2026-11-21T14:17:29.361Z"
modified_at: "2026-11-21T14:21:49.156Z"
topic: "agent_skills"
trust_label: "Reported"
publisher: "Emergent.FM"
---

# Skills adds 3 game skills for mesh scans, raindrops and 3D rosters

> The raindrops skill packs two WebGL 2 shaders to bend a live scene through water beads and runners.

## Key points

- The Skills repository commit adds three game development workflows, growing the library from 206 to 209 skills across ten categories.
- The build-lens-raindrops skill bundles two WebGL 2 shaders, drop-field.frag and refract.frag, to refract a live 3D scene through water beads and runners.
- The build-3d-character-selector guide notes that its master models are unrigged, requiring a skeleton and weight painting if a character must walk, gesture or drive.

A commit to the Skills repository adds three browser-game workflows under its game development category, expanding that section from 22 to 25 skills. According to the updated README, the additions cover build-point-cloud-scan, build-lens-raindrops, and build-3d-character-selector, targeting playable Three.js and WebGL projects.

## How the two-pass raindrop shaders work

According to the build-lens-raindrops guide, the effect takes a rendered 3D scene color target with clamp-to-edge sampling and runs two WebGL 2 fragment shaders. Builders first draw drop-field.frag into a linearly filtered RGBA8 target at half the scene resolution to store droplet slopes and height, then composite the scene with refract.frag. The guide notes that keeping the HUD in a later pass or DOM layer preserves text readability, while speed ramps can drive wetness.

## Building mesh-driven point-cloud cages

The build-point-cloud-scan skill ships an ES module, scan-cage.mjs, whose buildScanCage function takes packed XYZ vertex positions, optional triangle indices, and a positive cell size, according to the source code. It clusters vertices into cells, averages their positions, and connects cells along source triangle edges, returning typed arrays of point positions and edge indices. In the driving game example cited in the guide, a character handoff sweeps bottom-to-top over roughly 1.2 seconds, while a district pulse spans 215 world units over 6.4 seconds.

## From full-body references to a 3D roster

For build-3d-character-selector, the accompanying production guide documents starting from approved identity portraits or existing GLB models. In the driving game's workflow, Seedream 5.5 generated full-body references that fed Tripo H3.1 through Higgsfield, before builders decimated the masters in Blender into browser copies. The skill's stage workflow aligns each model to one canonical rest transform, adds rotation and keyboard controls, and uses request tokens to prevent stale loads during rapid roster switching.

## Stated limits for generated characters

The character production guide states two limits for generated casts. Because the driving game's master models were unrigged and relied on a small vertex-shader deformation for breathing and gaze, builders must add a skeleton and paint weights on a working copy for walking, gesturing, or driving. It also notes that a single reference image leaves the back design to the generator.

## Sources

- [File tree](<https://github.com/MengTo/Skills/commit/3ec2dc0857d36e45d032d10063f62f2c773329a3>) (github.com/mengto, release)

Canonical page: <https://emergent.fm/stories/92>
