Game Geometry Representation Choice

Use this skill before writing geometry-heavy game code. Pick the representation that makes the important operation simple, robust, and fast enough, then plan conversion boundaries deliberately.

Primary source: Geometry for Programmers by Oleksandr Kaleniuk (https://www.manning.com/books/geometry-for-programmers), transformed and paraphrased, especially chapters 2 and 10-12.

Core Workflow

1. Name the operation first: render, collide, pick, offset, deform, Boolean,

smooth motion, destruct terrain, generate props, edit geometry, or convert.

1. List the data the game already has and the data the game ultimately needs.

Include textures, normals, animation bindings, gameplay tags, and LODs.

1. Pick the working representation where the core operation is cheapest and

least fragile. Do not default to meshes just because rendering ends there.

1. Define conversion boundaries: input representation, working representation,

output representation, resolution, tolerated error, and attribute transfer.

1. Add invariants and tests for the representation, not just the final visual.
2. Revisit the choice when the operation mix changes. A good rendering

representation may be a poor editing or collision representation.

Representation Guide

• Use triangle meshes for rendering, GPU pipelines, LODs, static level geometry,

and approximating smooth surfaces when attribute data matters.

• Use SDFs or implicit fields for Boolean composition, offsets, hollowing,

organic procedural forms, cheap inside/outside tests, and robust repair steps.

• Use voxels or image masks for destructible terrain, coarse occupancy,

erosion/dilation, denoising, flood fill, gap checks, and reliable Boolean-like operations on a shared grid.

• Use splines or Bezier curves for player paths, camera rails, projectile trails,

roads, rivers, UI curves, and animation controls where continuity matters.

• Use parametric surfaces when procedural objects come from a small number of

parameters, such as surfaces of revolution, terrain patches, tubes, or rails.

• Use vector fields or deformation fields for smooth warps, influence maps,

wind/current effects, procedural variation, and nonrigid deformation.

• Use hybrids when one representation owns interaction and another owns

rendering, such as voxel sculpting with mesh LOD output.

Decision Rules

• If the operation is Boolean or offset-like, prefer SDFs or voxels over triangle

meshes unless exact mesh topology is required.

• If the operation needs exact texture, normal, skinning, or authoring data,

preserve mesh attributes or plan how to regenerate them after conversion.

• If the system needs smooth paths, start with curves and parameterization, not

with sampled points.

• If the game needs thousands of repeated spatial tests, choose a representation

that supports indexing, locality, or precomputed coefficients.

• If memory dominates, prefer boundary data such as meshes or splines over full

voxel volumes.

• If robustness dominates, prefer grid, SDF, or redundant fallback workflows over

a single brittle mesh algorithm.

Output Checklist

Return a short representation brief:

• Chosen working representation:
• Existing input representation:
• Required output representation:
• Operations made simpler:
• Operations made harder:
• Conversion steps:
• Error/resolution budget:
• Attribute preservation plan:
• Degenerate or failure cases:
• Minimal tests or debug views:

Common Mistakes

• Rendering from meshes, then forcing every gameplay operation to run on the

same mesh.

• Applying offsets or Booleans to arbitrary triangle meshes before considering

SDF or voxel workflows.

• Converting representations without a plan for normals, UVs, materials,

gameplay labels, collision layers, or LODs.

• Treating a sign-only implicit function as a true distance field after operations

that destroyed distance meaning.

• Choosing a high-resolution grid without proving the memory and streaming

model.

• Optimizing the algorithm implementation before choosing the representation

that makes the algorithm unnecessary.